It seems like there's news about new models, capabilities, and costs every few days ever since 2023. Model efficiencies and cost per token have started to really hit the main stage as companies invest more in frontier labs' capabilities, while others try to operationalize workflows at lower costs using open-weight models.

That's useful, but not the whole story.

Harnesses are starting to get attention, but they still aren't getting enough of it. Recent experiments have shown that harnesses not only govern what models can actually do, but can significantly impact total costs.

Simply put: the model's harness is the system that gives the model permissions, capabilities and can turn a chat assistant into an autonomous worker.

A model inside a workbench that can inspect files, use tools, make changes, and verify results is another thing entirely.

Subscribe now

Databricks Just Put Evidence Behind It

Ali Ghodsi, CEO of Databricks, recently described an internal evaluation across Databricks’ own tasks, codebase, and infrastructure, produced by more than 3,000 software engineers and spanning three hyperscale clouds. His key finding: for the same model,

“the choice of harness can significantly save costs (~2x).”

That is a significant, executive-level point. Token price is not the same as task cost. A cheaper model can consume more context, require more retries, or fail more often. A more expensive model can sometimes finish faster. And the harness can change the cost curve without changing the model at all.

The Difference Is Not Subtle

I experienced this moving from ChatGPT to Codex for my personal coding projects. ChatGPT is excellent for many knowledge questions. It is fast, conversational, and economically sensible when the task is simply to ask, think, draft, or understand.

But coding in a chat harness often leaves the human as the integration layer. The model writes code, but the user copies it into files, runs it, finds errors, returns to chat, and repeats the loop manually. This is very painful!

Codex changes the pattern. It can inspect the project, edit actual files, run checks, reason across steps, and iterate. The model is not just talking about the task. It is operating inside the full flow and acting more like a digital partner on your goals.

Subscribe now

The Same Pattern Shows Up Outside Coding

As Treasurer for a nonprofit, I produce monthly and quarterly financial analyses and reports. I had tried to get ChatGPT CustomGPTs and Gemini Gems to learn my style, follow my template, and automate the reporting process from raw data. They helped, but the process remained manual and fragile.

With ChatGPT Work, I provided prior examples, the template, and the raw data and it produced a strong report in one pass. That was not just a model difference. It was a harness difference. Chat is for knowledge finding and research whereas Codex moves through technical workflows using your computer or the cloud. Work helps produce real business deliverables (also using your local computer and files).

The right question is no longer only, “Which model should we buy?” It is, “Which harness should govern this work?”

Model Capability Still Matters

The lesson is not that model choice no longer matters. It is that model choice increasingly becomes something the harness should manage. Anthropic described Claude Fable 5 as its most capable generally available model, especially strong on longer and more complex tasks. Anthropic also distinguished Fable 5 from Mythos 5, a limited-release variant, and noted that Fable includes safeguards that may route some requests away from the top model in fewer than 5% of sessions on average.

That's the power of Fable 5's harness: it acts as an enabler and a governance safeguard.

At the same time, Moonshot’s Kimi K3 shows how fast cost-performance is moving. Moonshot describes Kimi K3 as a 2.8T-parameter open model with native vision and a 1-million-token context window, built for long-horizon coding, knowledge work, and reasoning. Recently it was evaluated next to Fable 5 as a potential open-weight challenger to Anthropic’s most powerful, generally available model. Moonshot says it still trails the strongest proprietary models overall, but it's showing frontier-level performance across its evaluation suite.

Subscribe now

What Executives Should Consider

The winning AI stack will likely be multi-model and multi-harness. In many companies, a proprietary or deeply customized harness may be the way to personalize AI around actual work. A custom harness could more accurately leverage company context, choose when to use a conversational AI, determine the tools and right model for different coding tasks, and quickly apply company assets to a document workflow. Harnesses could determine which governed enterprise agent may be required for a challenging or specific task to ensure consistency and business rule adherence. The harness can also determine when to route to frontier models versus cheaper or open-weight models.

The harness is what makes those decisions operational at scale.

It governs context, permissions, memory, tools, retries, verification, routing, and cost. It can turn model capability into leverage while removing the human guesswork and error, especially when employees are left to guess which tool, model, or workflow fits each task.

Moving Beyond Model Selection

The model matters. But the harness determines whether AI is an answer machine, an assistant, or an operating capability.

Executives who learn this distinction early will make better AI decisions. They will buy more intelligently, govern more effectively, and measure the economics of AI by the work it completes, not the tokens it consumes.

Subscribe now

A couple of weeks ago I had an epiphany: many AI strategies I’ve heard lately are basically the same as pounding a square peg into a round hole.

Subscribe now

It’s not anybody’s fault. For the past few years, we’ve heard experts say we should use AI to automate repeatable tasks. We’ve heard that agents will come online quickly and cheaply to do what humans do, only faster. We’ve heard that businesses should be looking for every routine process they can hand over to AI.

The problem is that people have not always thought deeply enough about the amount of work it takes to force an agent to do truly repeatable tasks correctly, accurately, and consistently every time.

I use an agent at work to build a morning report of email and meeting activity, and I’ll be darned if that report is not a little different every day despite how detailed the skills and instructions are. That is fine for my email management activity. It will not hold water for regulated companies that need transactions handled the same way every time.

Subscribe now

What is adding fuel to this fire is our new friend: tokenomics.

Tokenomics is the analysis of how much money your agent will save compared with the token costs required to run it. Companies are learning that they can build agents to do a lot of repeatable work, but at scale, the cost to run those agents can outweigh the human or traditional software alternatives they are used to.

That is not just a theory. A July 2026 KPMG report, covered by TechRadar, found that only 35% of organizations have full visibility into AI operating costs and that nearly half have delayed, paused, or shrunk AI strategies over cost concerns. Rob Fisher, KPMG’s Global Head of Advisory, put it plainly:

“AI is now as much a financial management priority as it is a technology one.”

The epiphany I had is that we, as leaders, need to start looking at our business problems through the lens of deterministic or probabilistic.

If a problem is truly repeatable and needs to be done the same way every time according to structured rules, then it is a deterministic problem. You should use AI to help build the tool, workflow, or system that can be customized to your business and scaled across your enterprise. The ROI comes from building that system faster with AI coding tools, implementing features more quickly as you scale, and still getting the human efficiency and redistribution benefits you were hoping for.

This is where AI helps you create the machine. It does not need to be the machine forever.

Meanwhile, when you find a probabilistic problem that requires research, judgment, decision-making, synthesis, and response generation, go with an agent. You can create the bones of that agent quickly with skills files and instructions. You can work with it side by side with humans. And as you roll it out to pilot teammates who reinforce what is right or wrong, you strengthen its context, memory, and usefulness over time.

Subscribe now

Anthropic makes a similar distinction in its December 2024 engineering article, “Building effective agents”. They describe workflows as systems where tools follow predefined paths, while agents dynamically direct their own process. Their advice is a useful gut check for leaders:

“Workflows offer predictability and consistency for well-defined tasks, whereas agents are better when flexibility and model-driven decision-making are needed at scale.”

That is the square peg and round hole problem in one sentence.

Let’s look at customer service for a great example.

Everyone has called into a customer service center. We call these Level 1 centers. The person on the other end usually has limited scripts, limited permissions, and access to the most common actions. Most of the work these people do is deterministic, with a light touch of traditional classification.

They ask mostly canned questions. They look up what to do in that circumstance. They try the one or two options they have to complete the transaction. If they cannot help you, they gather your information, open a ticket, and move it up to Level 2.

You should not automatically build an AI agent to do this work.

You can save a lot of time by using AI development tools and old-fashioned data science to build a scalable system that does these things and accelerates response times. You may not eliminate all of the people on the Level 1 team, but you can triage many of the items they would normally work on and free human operators to focus on items that do not follow the happy path.

At Level 2, the job is inherently more probabilistic.

The human agent has to research what happened, look for similar issues, determine how those issues were handled in the past, reference materials about related problems, come up with a recommendation, and then try solutions until one hopefully works.

This is what a probabilistic problem looks like. It involves research, analysis, judgment, and decision-making.

An AI agent that does most of those steps overnight and is ready with recommended actions for each ticket in the queue can make the Level 2 agent capable of handling many more cases than before. As the Level 2 agent confirms the AI agent’s recommendations, they train the system and help build confidence scores. In time, business leaders may decide that recommendations above a certain confidence score can be handled automatically, giving human agents more time for the novel and ambiguous problems that need real attention.

Subscribe now

This is also why the ROI conversation has to get more precise.

A 2026 arXiv paper by Longju Bai, Zhemin Huang, Xingyao Wang, Jiao Sun, Rada Mihalcea, Erik Brynjolfsson, Alex Pentland, and Jiaxin Pei, “How Do AI Agents Spend Your Money?”, found that agentic coding tasks can consume dramatically more tokens than simpler code reasoning or chat tasks. The authors call agentic tasks “uniquely expensive,” and their findings reinforce the bigger point: leaders need to understand cost behavior before scaling agentic work.

You have heard me mention “repeatable activities” a few times.

Repeatable tasks are like the penny machine in a souvenir shop. You put the penny in, turn the crank, and out pops a smashed penny with an image on it. You insert another penny, crank again, and get another souvenir.

That is repeatable.

Tasks related to coding, bug fixes, incident response, and anything else that sounds like the Level 2 example above are not repeatable in that same way. They are also not routine just because they happen often. They require thought, analysis, action, testing, additional resolution, and validation before they go into production.

Every bug is a new mini software project. It may not take long, but it still involves ambiguity.

Make sure you are using AI coding tools to accelerate this work. But I would be careful about fully outsourcing these kinds of tasks to AI agents yet.

There is real promise here, but only if leaders stop treating AI like one big hammer. The MIT NANDA report on enterprise AI adoption, covered by Tom’s Hardware in August 2025, found that only about 5% of enterprise generative AI pilots achieved rapid revenue acceleration. The report’s lead author, Aditya Challapally, explained the successful group this way:

“They pick one pain point, execute well, and partner smartly.”

The question shouldn’t be, “Where can we use agents?”

The better question is: “What kind of problem are we solving?”

Subscribe now

Five things AI leaders should do today

1. Classify your work before choosing the tool. Separate deterministic problems from probabilistic ones before you talk about agents, copilots, automation, or vendors.

2. Use AI to build deterministic systems faster. If the process needs the same result every time, use AI coding tools, data science, and workflow automation to build the system instead of running every transaction through an agent.

3. Reserve agents for ambiguity. Agents are best when the work requires research, judgment, synthesis, recommendations, and adaptation.

4. Measure tokenomics before scaling. Do not wait until the bill arrives. Model expected usage, compare it with human and software alternatives, and build cost visibility into the strategy from the beginning.

5. Keep humans in the learning loop. The best agent deployments will not be “set it and forget it.” They will be side-by-side systems where humans validate, correct, reinforce, and gradually expand the agent’s authority.

The leaders who win with AI will not be the ones who put an agent on everything.

They will be the ones who understand the work deeply enough to know when they need a deterministic system, when they need a probabilistic partner, and when they are just trying to pound a square peg into a round hole.

That is the kind of AI strategy I am most interested in: practical, measurable, and tied to the way work actually happens. If you are thinking through these questions inside your own organization, stick around. I’ll keep writing about the gap between AI hype and AI that actually earns its keep.

%%first_note_embed%%

Notes is a new space on Substack for us to share links, short posts, quotes, photos, and more. I plan to use it for things that don’t fit in the newsletter, like work-in-progress or quick questions.

Go to Notes

How to join

Head to substack.com/notes or find the “Notes” tab in the Substack app. As a subscriber to AI d_AI_ta POINTS, you’ll automatically see my notes. Feel free to like, reply, or share them around!

You can also share notes of your own. I hope this becomes a space where every reader of dAIta POINTS can share thoughts, ideas, and interesting quotes from the things we're reading on Substack and beyond.

Share AI d_AI_ta POINTS

Why Substack Notes

Fellow writers and readers are spending time in Notes for a variety of reasons.

Ted Gioia sees Notes as a forum for dialogue.

The idea behind Notes is simple. Millions of people now participate in Substack as writers and readers—but much of this is built on long articles and essays. We now have a forum for dialoguing and sharing shorter posts.

Chris Ryan is drawn to Substack as an alternative for legacy social media.

One of my main reasons for joining Substack was to wean myself off exploitative social media platforms. I’m hoping this is going to help move us along on that journey!

Sherman Alexie wants to create a kind and creative new space in Notes.

I’m going to focus on being positive, with my own thoughts and photos and by linking to songs, stories, poems from around the web and from other Substacks.

If you encounter any issues, you can always refer to the Notes FAQ for assistance. Looking forward to seeing you there!

Go to Notes

📌 THE POINT IS: AI is lowering the cost of creating highly specialized software. That changes the traditional tradeoff between standardized vendor platforms and expensive custom development.

Technology leaders should standardize commodity capabilities while retaining ownership of workflows that create competitive advantage.

The answer is not unmanaged app sprawl. It’s hyper-customization built on governed data, enterprise controls, automatic cataloging, and application lifecycle management.

For most of my career, “app sprawl” has been something technology leaders were supposed to eliminate.

Too many tools meant duplicated costs, fragmented data, inconsistent controls, and security exposure. The responsible strategy was consolidation and standardization.

That strategy made sense when software was expensive to build, difficult to maintain, and slow to change.

But what if the economics of software creation are changing faster than our operating assumptions?

Subscribe now

New world economics

I have been thinking about this while building my own software through AI-assisted development, or vibe coding. While using the workflow I created for cataloging items, I kept noticing small improvements I wanted: a contextual field, a faster flag, a different view of related records, or a shortcut for a repeated action.

These were not roadmap items. They were micro-features.

Traditionally, most would be too small to justify a development cycle. I would have absorbed the friction through extra clicks, a spreadsheet, a sticky note or a mental workaround.

Instead, I could add a feature while the idea was fresh, test it locally, and move it into production in a tight loop.

The software began to mold itself around how I worked, rather than requiring me to mold my work around the software.

That experience has made me reconsider whether all app sprawl is inherently bad.

When small problems become worth solving

When software is expensive, only large problems deserve software. When software becomes cheaper and faster to create, small frictions become addressable.

Ethan Mollick described a related shift in “Claude Dispatch and the Power of Interfaces”. He suggests that the future may not be “one interface to rule them all,” but AI generating the right interface for a particular moment, including “a custom app to solve a problem.”

That is the promise of hyper-customization

Enterprise applications are designed for averaged workflows. A vendor cannot design every screen around one employee, one team’s exceptions, or one company’s operating process.

Generalized software creates scale, but it also creates quiet mismatches:

• The workflow almost fits.

• The data model captures most of what matters.

• The exception lives in email, a spreadsheet, or someone’s memory.

• AI-assisted software can begin to occupy that gap.

• Are we standardizing our secret sauce?

At the Gartner Research Board Cycle 2 2026 meeting in Chicago, my takeaway from the discussions was that this question is larger than personal productivity.

Large enterprises frequently implement vendor applications for entire processes or deploy multi-module ERP platforms such as SAP. These provide mature controls, reliable transactions, regulatory support, integrated data, and an upgrade path.

But these decisions often carry an assumption that receives too little scrutiny:

The business process should conform to the software.

That may be reasonable for commodity capabilities. Few companies differentiate themselves through the mechanics of maintaining a general ledger or administering standard benefits. Standardization there can reduce cost and risk.

Subscribe now

The calculation changes when the process is part of the company’s advantage

How a company prices, fulfills, underwrites, merchandises, manages risk, or serves customers may be part of its secret sauce. Forcing those capabilities into a vendor’s generalized model can produce consistency while removing some of what makes the company distinctive.

Historically, executives accepted that tradeoff because:

• Custom software was too expensive.

• Customization made upgrades painful.

• The vendor was presumed to have encoded an optimized process.

Those constraints were real. They should not become permanent beliefs after the economics change.

Buy the commodity. Build the advantage.

AI-assisted development does not mean replacing an ERP with thousands of casually generated applications. That would exchange rigidity for chaos.

The useful question is:

Where should the boundary between packaged and custom software now sit?

My working principle is:

Buy the commodity. Build the advantage. Connect both through governed data and explicit contracts.

Core platforms can continue to manage durable records, commodity transactions, and enterprise controls. A governed data platform can preserve shared definitions, history, quality, and access policies.

Around those foundations, specialized applications can express how the company wants to operate. Some will extend vendor platforms; others may replace modules where customization produces enough value.

New but not necessarily novel

This is not entirely outside the vendor playbook. SAP’s clean-core guidance recommends standardizing the core while prioritizing extensions for business-differentiating processes. Gartner’s work on postmodern ERP and composable enterprise architecture similarly challenges the idea that every capability must live inside one monolithic suite.

Mollick provides an early signal about the changing cost curve. In “Management as AI Superpower”, he describes executive MBA students, most without coding backgrounds, creating working prototypes in four days.

A prototype is not a production system; security, resilience, integration, and regulatory requirements remain. But the cost of building, testing, discarding, and revising software is falling.

That changes which problems are economical to solve.

Subscribe now

The governed hyper-customization stack

The architecture I am imagining has three layers:

1. Stable enterprise core: ERP, CRM, systems of record, commodity transactions, identity, and financial or regulatory controls.

2. Data and control plane: Governed data products, APIs, security policies, observability, lineage, audit logs, and application lifecycle management.

3. Adaptive application layer: Personalized interfaces, workflow tools, specialized business modules, and AI agents operating within defined permissions.

The adaptive layer can change quickly because the foundations beneath it remain stable.

The small app becomes the interface, not the prison for the data. If an application is retired, its durable data remains available to whatever replaces it.

The harness matters more than the app count

This is not a defense of unmanaged shadow IT. The relevant measure is no longer simply how many applications exist. Leaders should ask whether every application:

• Has a named owner and business purpose.

• Inherits enterprise identity and permissions.

• Uses approved data connections.

• Passes security and deployment checks.

• Produces logs and usage information.

• Has traceable versions and a rollback path.

• Can be retired without losing important data.

Mollick makes a similar point in “Claude Code and What Comes Next”. He highlights what AI coding agents can accomplish “with the right harness,” while warning that they can delete files, run unintended code, or expose sensitive information.

The capability and the control problem arrive together

Every employee-built application should therefore be cataloged automatically. The catalog should capture its owner, purpose, users, data sources, permissions, versions, security status, and last-used date. Microsoft’s Power Platform governance model already points in this direction through centralized inventory, usage monitoring, and administrative controls.

Applications should also have an ending. An automated process can flag inactive tools, notify owners, preserve durable data, and recommend archival or retirement.

App sprawl becomes dangerous when software has no lifecycle.

Subscribe now

What technology executives can do today

This does not require an immediate ERP replacement. It requires a controlled way to learn.

1. Choose one differentiated workflow. Find a process where your company’s way of working matters and current software creates obvious friction.

2. Map the hidden workaround layer. Document the spreadsheets, emails, duplicate entry, manual reconciliations, and knowledge carried in employees’ heads.

3. Run a governed experiment. Give a small team an approved AI-assisted development environment with enterprise identity, governed data, source control, testing, and logging.

4. Measure operating value. Track cycle time, manual steps, decision quality, errors, adoption, and iteration speed rather than lines of code or applications created.

5. Catalog from the beginning. Capture the owner, purpose, users, data, permissions, versions, and last-used date during deployment.

6. Define retirement before launch. Decide when the application will be reviewed, where its data will remain, and how it can be shut down safely.

7. Update the build-versus-buy framework. Ask: Is this capability a commodity we should standardize, or an advantage we should own?

A different kind of sprawl

Some app sprawl is still waste, duplicated effort, security risk, or poor technology leadership.

But some may be creativity made operational.

It may represent employees solving problems while they still understand them, companies learning faster than major platforms can adapt, and distinctive capabilities scaling beyond software designed for the average customer.

The future of enterprise software may not be one perfect platform used the same way by everyone: it may be a governed ecosystem of stable core systems, shared data products, and many adaptive tools close to the work. Those tools can be created quickly, observed centrally, connected securely, and retired when their usefulness ends.

That is not the app sprawl we were warned about.

That is hyper-customization with a control plane.

Subscribe now

Share

App sprawl is the new BI sprawl

Subscribe now

Vibe coding is not just a developer productivity story: it’s the beginning of app sprawl.

For the last 20 years, companies have dealt with BI sprawl: too many dashboards, inconsistent definitions, and competing versions of the truth. That was messy, but most dashboards were read-only.

AI-built apps are different. They can move data, trigger workflows, encode business logic, expose permissions, and create security risk. And now the ability to build them is moving from software teams to almost anyone who can describe a problem clearly.

That is exciting! But, it’s also an executive governance problem.

Testing the waters hands-on

Over the weekend, I tested this myself. I used Codex to build a web app for cataloging my dad’s collection of 160+ vintage vinyl records. I expected to have a rough prototype after a few afternoons. Instead, I had a usable app quickly, iterated on features as I entered the records, and finished the cataloging workflow far faster than expected.

That experience changed how I think about citizen development in an AI world. The bottleneck was not writing code. The bottleneck was knowing what I wanted, testing whether it worked, and deciding what risks mattered.

The skill shift

At a recent Executive Roundtable event in Charlotte, I was asked to be on a panel of speakers about data, AI, and the future. I was asked what college students should be studying today and I replied, “good product manager skills.” The ability to describe the problem in terms of outcomes and value; identify killer features and how to prioritize them; and being able to do the user testing that will ferret out any gotchas or misses in functionality are going to be the most valuable skills. Simon Sinek in his book, “Start with WHY,” speaks about a world where the “what” is a detail, but the “WHY” is what gets people excited about the work you’re doing. As good product managers, that’s where we should be upskilling and preparing ourselves.

Aleks Bass, Chief Product and Technology Officer at Typeform, told TechTarget: “GenAI has lowered the fluency bar…” from coding to reasoning about the problem. What we can do best is give it meaningful problems, then, that drive value and outcomes for ourselves, our associates and our customers.

Subscribe now

Enterprise risks are increasing at exponential speed

This does come with some risk though for organizations. A cyber architect colleague and I were recently chatting and the reality is that while dashboards are mostly low-risk, apps that aren’t inventoried or governed, and that proliferate but don’t necessarily ever get shutdown, exponentially expand your company’s attack surface without anyone knowing. Even when many people create similar dashboards with slight differences, the primary risk is usually making the wrong decision from inconsistent information. App sprawl will give us executable business logic, data movement, permissions, and security exposure. That’s a much bigger deal.

Without skilled operators making sure those apps are coded with “industrial grade” protections, your customers’ data will be shared, copied, used in insecure databases, and be exposed to constant cyber threats. This will (yes: will) create vast opportunities for offensive AI systems used by threat actors to exploit. It almost seems inevitable, so CEOs and technology leaders need to be thinking about the trade-off between self-service and speed vs. their company being hacked, attacked, ransomed, or worse.

Speed needs guardrails

All is not lost: the better move is to learn from the BI example. With the right guardrails, company agreements, and cultural expectations, leaders can steer their teams to doing the right thing while still taking advantage of speed. One example is ensuring that you have enterprise-grade AI protection infrastructure in place. The next is making sure that your company uses a harness that has your guardrails built in. This is what we mean when we say governance as code: your build environment needs to automatically guard against vulnerabilities. With coding tools as sophisticated as they are, companies will need AI agents driving more of the development and maintenance lifecycle. The winners will be the companies that allow those agents to automatically inspect code, fix vulnerabilities, run security checks, update related issues, execute regression tests, and prepare releases for human review in record-breaking time. The human being will still own the decision, but AI agents will accelerate the cycle.

Subscribe now

5 Things To Do Today

1. Create an AI-built app inventory now.
   Ask teams what they have already built with AI, low-code, spreadsheets, scripts, or workflow tools.

2. Classify apps by business risk.
   Personal productivity tools, team workflow tools, and business-critical/customer/data-sensitive apps should not have the same review process.

3. Give citizen builders an approved sandbox.
   Make the safe path easier than the rogue path.

4. Require an owner, data classification, and review date.
   Every small app should have a named human, known data sources, and a review point. Even small apps need to have a lifecycle including retirement.

5. Teach product management as a core AI skill.
   Train people to write requirements, test outputs, document decisions, and know when to call IT, data, cyber, legal, or compliance.

If you are leading data, AI, technology, or operations work, this is the new operating layer to watch: not just what AI can build, but how companies govern what gets built.

Vibe coding is here, and it is not just for software developers. It is for anyone who can see a problem, describe a better outcome, test what gets built, and keep improving it. That should excite leaders, but it should also sober them up. App sprawl is coming, and it will move faster than BI sprawl ever did. The answer is not to slow everyone down. The answer is to teach people how to build responsibly and give them the guardrails to do it safely.

Subscribe now

Share

Most companies still think about AI memory like it is a bigger chat history. That is the wrong mental model.

A long context window can help an agent remember what happened earlier in a conversation. That is useful, but it is not enterprise memory. It does not automatically know which lessons should apply again, which were one-time exceptions, which require approval, and which are outdated.

If AI agents are going to become digital workers, they need a way to learn without turning every correction, preference, or exception into an uncontrolled operating rule.

Subscribe now

“Without memory, AI agents forget user preferences, repeat questions, and contradict previously established facts.”

– Mem0, Building Production-Ready AI Agents with Scalable Long-Term Memory

Knowledge tells the agent what is true. Memory tells it what changed.

In the first article in this series, I wrote about Knowledge Graphs as the deterministic control plane for agentic AI. They help agents produce answers that are correct, consistent, and explainable.

In the second article, I wrote about Context Graphs as the layer that captures institutional judgment, especially when the answer is not cleanly written in a process document.

The Memory Graph is the third piece. It captures what the organization has learned over time: human corrections, customer outcomes, policy changes, approved exceptions, escalation patterns, and decisions that were valid yesterday but are no longer valid today.

That last point matters. Memory without time is dangerous.

“Traditional knowledge graphs treat facts as static, but real-world information evolves constantly.”

– OpenAI Cookbook, Temporal Agents with Knowledge Graphs

If an agent remembers that a customer exception was allowed last quarter, but does not know the policy changed this quarter, it is not smarter. It is confidently outdated.

Subscribe now

A Memory Graph is not just a database

A Memory Graph is not one tool. It’s an architecture pattern.

In practice, it may include a graph database to represent entities and relationships, a vector store for semantic retrieval, an event log for auditability, a rules engine for deterministic policy enforcement, and a governance layer for approval, retention, access, and invalidation.

A rules engine tells the agent what it must do. A Memory Graph tells the agent what the organization has learned. Those are related, but they are not the same thing.

• “Customers in this segment usually ask for an escalation” is memory.

• “Customers in this segment must receive this disclosure” is a rule.

Blend those together carelessly and your agents will start treating patterns like policy. That is how risk gets scaled.

Personal memory and enterprise memory need different governance

Personal memory improves user experience. Enterprise memory improves organizational performance. Those two memories cannot live under the same governance model.

A personal memory might be, “Matt prefers concise answers.” Helpful. Low risk. Easy to delete. An enterprise memory might be, “This claim scenario requires escalation when these three conditions are present.” That needs lineage, approval, access controls, retention rules, and a clear invalidation process.

LangChain’s memory framework separates memory into facts, experiences, and learned behaviors. Those distinctions matter because different kinds of memory require different controls.

The most valuable enterprise memory will not come from generic chat logs. It will come from human correction. When a customer service agent gives the wrong answer, or when a manager explains why an exception was allowed, that learning should not die inside the transaction.

Subscribe now

5 things you can do now to prepare for Memory Graphs

1. Choose the memory architecture before you choose the agent.

   Do not let every team bolt memory onto agents in a different way. Define the enterprise pattern first: graph storage, vector retrieval, metadata, audit logs, and business-rule integration.

2. Separate memory from rules.

   Rules define what must happen. Memory captures what has been learned. If your agents cannot tell the difference between a pattern and a policy, they are not ready for autonomy.

3. Create separate lanes for enterprise memory and personal preference.

   Personal memory should be scoped to the user, transparent, and easy to delete. Enterprise memory needs lineage, approval, access, retention rules, and invalidation.

4. Start with human corrections.

   Pick one workflow where humans already correct AI outputs. Capture the correction, reason, source, outcome, and whether the lesson applies once, sometimes, or always.

5. Design forgetting into the system from day one.

   Memory without expiration becomes operational debt. Policies change. Regulations change. Strategy changes. A Memory Graph must retire, supersede, or challenge old memories before they become scaled risk.

This is where your digital workforce starts learning safely

Knowledge Graphs help agents know what is true. Context Graphs help agents understand how judgment gets applied. Memory Graphs complete the architecture by helping agents learn from what actually happened.

But enterprise memory cannot be a junk drawer of chat history, user preferences, and stale assumptions. It has to be governed infrastructure.

When designed well, the Memory Graph turns human feedback into institutional learning. Policy changes travel faster. Agent behavior gets more consistent. The digital workforce gets a safer way to improve over time.

That is the real unlock: agents that learn and scale under control.

📌 THE POINT IS: Big, old companies are not too late for agentic AI. They may actually have the most to gain, if technology leaders are willing to invest right. The agentic race will be won by technology leaders in large enterprises who fund the infrastructure that lets AI scale without breaking trust. Do that well, and the payoff is bigger than efficiency alone: it’s the chance to redeploy talent, improve service, and build the next version of the company.

Why big companies feel behind

You’re the CTO of a large company. Every day you wake up to news about AI agents taking over workflows, working all night on projects, and shipping code at never-heard-of speeds. Then you bury your head in your hands wondering, “Why can’t I have this life!?” Most of the incredible news that you hear about these days that tout large, agentic wins are companies that were built on digital rails (e.g. newer, usually smaller ones).

No wonder you’re feeling frustrated and behind…their realities are very different than yours.

Subscribe now

I’ve lived the mess

For most of my career, I’ve worked at large, old enterprises. GE and Bank of America were > 100 years old when I worked at each and now Nationwide is celebrating its 100th year of protecting people, businesses and futures with extraordinary care. As large-sized corporations, they all have diverse technologies that have either been bought, built, or acquired through purchases of other companies. The technology environments are old, vast, complicated, and most of the time held together with digital bubblegum and duct tape.

This is not the environment in which you let AI agents (let alone humans!) run free. There are decades of crazy code written well before CI/CD automation and testing were ever a thing embedded in some of the most critical systems that keep the place going! In a large enterprise, bad automation does not just break a toy process. It can break customer trust, compliance posture, or core operations.

Don’t freeze. Build the stack.

I’m here to tell you that while all of this is true, there is no reason to fear or freeze in your tracks when considering starting with AI agents. I’ve talked a bit about Knowledge Graphs and Context Graphs as early steps to getting your data environment ready for AI. There are a number of other “infrastructure” components such as AI gateways, orchestration tools, and model selection and evaluation wizards that you’re just going to have to buckle down and implement. Deloitte found that while 42% of companies felt strategically ready for AI, they felt less prepared on infrastructure, data, risk, and talent.

AI isn’t going to be an overnight project for you like it is for the Ma-and-Pa shops out there that aren’t carrying the burden of providing electricity to homes, facilitating everyday people’s wallets, or protecting your family’s home. I hate to say it, but the large, old companies out there in many cases are providing the most critical elements that we all need to live safely. There’s just more responsibility on your hands and that means you need to spend more money to get there.

Subscribe now

The ROI story that most leaders get wrong

But it’s the ROI upside that you have to keep your eyes on. The largest failing that technology and business leaders have today is that they think that ROI should come in the form of decreased costs and a wider margin on the bottom line. That’s a good secondary goal to have, but the real goal should be growing your business, increasing your top line through offering new products and services, and finding ways to reimagine your company for its next 100 years. Having AI process call center transactions will improve your margin, sure, but if you deploy the humans who were spending days answering phones and train them to talk to those same clients to uncover potential new business offerings or better ways to offer value to your clients, you’ll see an even larger overall impact.

Let’s take the very recent example of American Airlines. The airline is launching electronic boarding gates at DFW to streamline boarding and free up gate agents to help customers with more complex issues instead of focusing on the repetitive task of monitoring boarding-pass scans. That is the kind of “boring” use case that scales in a large enterprise: automate the simple, repeatable work so your people can spend more time on higher-value service and operational needs. Over time, that kind of change can give operations leaders more flexibility in how they deploy staff during disruptions, delays, and high-touch customer moments.

Subscribe now

5 things you can start with today

So how do you get started if you’re staring at a valley of mainframes, COBOL code, and tears? Here are your first five steps:

1. Pick one ugly, high-value workflow and map it end to end. Don’t start with a moonshot. Start with a process that is painful, repetitive, and economically meaningful...like American Airlines!

2. Identify the systems of record, data sources, and decision points that workflow touches. If you do not know where the context lives, your agent will not know what “right” looks like either.

3. Put guardrails in before autonomy. Define who approves what, where humans stay in the loop, what actions require escalation, and how logging and audit trails will work.

4. Build the context layer on purpose. That may mean a Knowledge Graph, Context Graph, memory layer, or all three. Whatever you call it, your agents need business meaning, not just tokens.

5. Track value beyond labor savings. Measure cycle time, quality, customer impact, conversion, risk reduction, and the new revenue opportunities your people can pursue once AI takes the grunt work off their plates.

Act your size, but not your age!

If you’re a large, old company, this is the part where you stop envying the startups and start acting like the institution you are. Your edge is not speed alone. Your edge is trust, scale, customer reach, and the ability to reinvent critical work in ways that matter. The winners in this next era will not be the companies that rushed agents into production with the fewest guardrails. They will be the ones that invested early in the infrastructure, context, and governance required to scale AI without blowing up the business. That’s how old companies will continue to get older…and better!

Share

📌THE POINT IS: The ROI on your AI investment is directly tied to the level of autonomy you can safely grant your agents. By combining the deterministic rules of the Knowledge Graph with the institutional judgment captured in the Context Graph, you build the foundation for a digital workforce that can execute high-volume, high-stakes transactions with explainable, human-like expertise, finally unlocking the path to scale.

Subscribe now

The AI Autonomy Crisis

You have spent millions on AI pilots and data infrastructure, but most of your digital workers are still stuck in “human-in-the-loop” monitoring, afraid to make a decision without supervision. Why? Because despite the hype, the vast majority of AI projects fail to scale across the enterprise. The fundamental issue is trust. When agents are granted autonomy, they must be reliable, and they are only as safe and scalable as the data foundation beneath them.

The Knowledge Graph is Not Enough

In my recent article on the criticality of Knowledge Graphs for enterprise AI, we discussed why a foundational investment in these graphs is a critical first ingredient in making sure that agents reply with correct, consistent, and explainable results. But just getting the right information isn’t going to empower your agents to make good judgement calls. Somehow, your agent needs to be able to look up solid information and policies, but also know when humans “bend the rules” in the name of customer service and satisfaction without breaking your company.

“Agents can read data and take action, but they still don’t know why decisions get made.” - Jaya Gupta, Founder, Foundation Capital

Subscribe now

Context: The Key to Human Judgment

Context Graphs are a new concept to many people. Jaya Gupta defines them well:

“A context graph is institutional memory for how an organization actually makes decisions: not how the process doc says it should, but how it works in practice.”

However, Malvika Jethmalani from “Reworked” puts it a little more simply:

“Context graphs protect judgment.”

Context Graphs help your AI agents think more like your human workforce. You have several experienced customer service representatives today. They know the ins and outs of your policies, but they also know what kind of brand is important to your company and thus how to flex policies when needed. When a customer is on the phone making a request, your experienced human agents will override a policy using “tribal knowledge” to satisfy the situation.

Your AI agents need to know how and when to do that if they’re going to be successful!

Unlocking Superhuman Digital Workers

When your AI agents know where to get information and “how-to” instructions (KGs) and know how to apply that context to elevate customer service and your company’s brand, you’re on the precipice of unlocking superhuman value for your most high-volume transactions. Your human-in-the-loop cycle will quickly move from needing approval for every transaction to only getting involved with specialty cases and new situations that the AI needs help learning about.

This is where the ROI is for your investment.

AI agents must be able to offload human toil and increase throughput to grow your top and bottom lines at the same time. Think of selling agents, service agents, claims agents. If you snap together the Knowledge Graph and Context Graph, all of those dreams for your business start to come into focus.

Subscribe now

But there’s ONE more piece to the puzzle: The Memory Graph

In a future article, we’ll talk about the last piece of the emerging puzzle that fully transforms your agent experience. Memory Graphs are the key to automatic learning that should take place whenever a human gives feedback to agents about a decision they made. For example: an agent may make a correct decision, but later a human indicates that it’s no longer valid due to a change in the company’s direction. The agent will remember that decision, update itself to flag future conditions, and make the new decision that’s being enforced.

Now imagine that change gets rolled out to all of your digital workers at once without having to send an email, hope people read it, hope people weren’t on vacation when there was training, etc. That’s the power of Memory Graphs in your digital workforce.

“The companies that win will invest in a living record of how their organization thinks, decides, and evolves. That’s a compounding asset and a trillion-dollar opportunity.” - Klarity

The Architecture Mandate: 5 Steps Executives Can Take Today

The choice to build autonomous AI agents is happening everyday. You must ensure that you capture your organization’s most valuable, high-stakes asset: human judgment. This investment is the only way to transform investment into ROI.

Here are five immediate, executive-level steps to begin building Context Graph readiness:

1. Document the Decisions That Matter Most: Identify the 50 critical processes that drive your highest value and risk and start recording the rationale, options considered, and the owner—not just the final outcome.

2. Make Exceptions Traceable: Institutionalize the practice of recording why a policy or Standard Operating Procedure (SOP) was bypassed.

3. Instrument Your Workflows: Mandate that cross-system communication and informal handoffs be captured as data. Auto-scribe meetings, and ensure approvals occur within systems, not just hallway conversations.

4. Create a Shared Vocabulary for Judgment: Ensure cross-functional alignment on key concepts like “onboarding complete” or “priority one incident.” Without a common taxonomy, neither humans nor AI agents can reliably interpret the context needed to apply judgment.

5. Pilot a High-Value, Exception-Heavy Workflow: Select a single domain (like mid-market deal cycles or compliance reviews) where “it depends” is the honest answer. Capture its steps, measure the time and outcomes, and use this data to inform your first Context Graph model.

Start scaling your investment today!

Your company’s ability to reach a trillion-dollar valuation depends on how well you implement AI architectural decisions like these. The Context Graph transforms the intangible, high-value asset of human judgment into a machine-readable format, making it the non-negotiable prerequisite to increasing autonomy in your agents. This is how you scale AI investments. The time to start capturing this critical context is now.

Subscribe now

Share

📌 THE POINT IS: The rise of AI has created a crossroads for capitalism, making the historical model of compensation tied to "toil and hours" obsolete. Business leaders must urgently redefine human value around Wisdom, Creativity, and Judgment, or face an "Implosion Cycle" where mass job displacement collapses the consumer base, ultimately bringing the entire economy down. The solution lies in "Creative Dividends" humans can bring to your company.

Subscribe now

The Fork in the Proverbial Road

“First solve intelligence, then use intelligence to solve everything else.” — Demis Hassabis on the philosophy of DeepMind

There’s been a lot of chatter in the AI news about trillion-dollar companies, companies that will have one employee and huge margins, and companies that will pop up like weeds every time someone has an idea that they can “vibe code” into a product. But the challenge remains that many companies, especially larger ones, still expect that they’ll be able to dramatically downsize their workforce. Recent news has announced smatterings of companies like Amazon, Block, UPS, even Duolingo who have let go of contractors or employees (sometimes in the thousands) related to AI advancements.

Some of us have started to pick up our heads and ask a few questions:

1. Who will ensure that companies which have been around for over 100 years will last another 100 years?

2. Who will be the customers of these trillion-dollar companies when the population of income-earning consumers dwindles with their employment rosters?

Even the gig economy is showing signs of cracking as companies like Waymo start to advance their fleets of self-driving cars and leaders like Elon Musk start seriously pivoting manufacturing efforts to expand that technology.

We need to think about cutting a new path in the woods for our system of economics, from governments to company heads and investors, before the path most taken leads us off a cliff.

“The historical model of compensation—reward tied to toil and hours—is now economically obsolete.” — Genesis: Artificial Intelligence, Hope, and the Human Spirit

Subscribe now

The Uncoupling: Efficiency is No Longer Value

The philosophy of DeepMind is a stark reality check for every business leader: if a problem can be solved by intelligence, AI will soon solve it at a fraction of the cost, moving the intellectual value of labor toward zero marginal cost. As Eric Schmidt, Craig Mundi and Henry Kissinger observe in Genesis; Artificial Intelligence, Hope and the Human Spirit, the historical model of compensation—reward tied to toil and hours—is now economically obsolete. AI has broken the direct link between human effort and corporate value creation. The new value must be found in uniquely human activities: Wisdom, Creativity, Judgment, and Empathy. This uncoupling of compensation from toil frees human cognitive capacity, unlocking a creative dividend that can be channeled into designing novel business models, holding continuous reimagination sessions, and pioneering entirely new markets and products.

The continued flawed assumption is that companies are looking for the cost center they can reduce via AI. The opposite question is what we should really be asking:

Cost reduction will come as a function of productivity enhancements per person, but you literally cannot grow your top line if you’re focusing all of your business model changes on bottom line impacts. This is the great turning point. We can use this immense productivity to free humanity from drudgery.

“The greatest revolution of the 21st century will not be in technology, but in the definition of what it means to be human.” — Yuval Noah Harari

Subscribe now

The Fatal Flaw: Individualism or self-wealth mindset

If the gains of AI’s productivity flow only to capital, the result is a catastrophic concentration of wealth. A small class of super-rich investors will prop up “trillion-dollar” companies that have minimal workforces and astronomical margins. However, these companies will ultimately face a terminal crisis: a collapsed consumer base. If mass job displacement eliminates the paycheck and purchasing power of the middle class, the very customers needed to buy the products and services these efficient companies generate will disappear. The cycle of layoffs leading to lost customers will cause a full-system implosion. This is why a new economic contract—one that shares the “Creative Dividend” of AI with the workforce—is not merely an ethical imperative, but an act of corporate and governmental self-preservation.

5 things you can do today in your company, large or small

1. Redefine Your Employee Value Proposition (EVP): Stop rewarding employees primarily for volume of execution (which AI excels at). Start rewarding them for Wisdom, Creativity, and Judgment—the human activities that are inherently better than machines.

2. Mandate a Creative Day: Pilot a formal 4-Day Work Week structure where the fifth paid day is ring-fenced for unrelated personal growth, deep thinking, or internal innovation “Hackathons” focused on creating entirely new markets for the company.

3. Implement a Core Profit Share: Design a Mutualized Profit Pool to distribute a predefined percentage of annual profits back to the workforce, ensuring that the wealth generated by your company’s AI is shared and stabilizes the broader consumer economy.

4. Shift L&D to EQ/Teaming: Immediately re-allocate your Learning & Development budget to training programs focused on Emotional Intelligence, Ethics, Systems Thinking, and Human-Machine Teaming, moving away from pure technical skill instruction.

5. Require Customer-Fronted Field Study: Embed a mandatory, recurring requirement for all senior and strategic employees to spend time on direct, ethnographic field studies with customers to discover unsolved problems and generate human-led, empathy-driven new product strategies.

Subscribe now

Here we go…

On the one hand, this will be a very stressful time. The next few years are going to be painful as our entire society either comes together to figure out a new system that supports continued human development or falls apart due to the overwhelming pull of human greed. It’s simply no longer acceptable to continue down the path we’re on and all of the major players are sending similar signals through memorandums to / from governments and books hitting shelves every day.

Will we have the courage to rise above our internal, self-worldviews to achieve societal change?

Share

📌 THE POINT IS: Everyone says ‘bad data,’ but they’re describing different problems. Define data quality by component and owner, then build a plan to improve trust, because agentic AI will operationalize your data at scale.

Subscribe now

The Plinko Problem: You can only guess where a “bad data” conversation will land you

Lately I’ve had a lot of conversations with people who end the same way: a business or technology leader shrugging their shoulders in defeat about “bad data.” In a conversation with colleagues not long ago, I stopped them and started asking, “But what does bad data mean…to you…?” The term “bad data” is being used to describe several actual problems or phenomena, but unless you dig into what the person is actually experiencing, it’s like watching a Plinko ball fall down the pegboard during a “The Price Is Right” episode: you just don’t know where the conversation will land.

Confusion usually comes from two places: either enterprises don’t have a clear definition of “data quality” or that definition includes too many components. Good data quality in the analytical environment is achieved when the data faithfully matches the upstream systems. I over-index on that to build trust in the environment.

“3 out of every 5 data scientists… spend the most time cleaning and organizing data.” — CrowdFlower Data Science Report (2016)

Upstream—where data is created—CTOs and application leaders own quality. To an application owner, good data is data that’s captured accurately within any required business bounds that can then be used for any use case required later on. As long as the data is created correctly using input controls and other audit strategies, the possibilities are endless.

Build the Plinko Board: the 4 buckets “bad data” usually falls into

Data quality typically falls into a handful of buckets. Each one could be an article, but there are patterns, at varying levels of effort and investment, that technology and business teams should apply to increase quality.

1. We don’t have the data (not captured, not retained, not accessible)

2. We have it, but people don’t understand it (definitions, timing, joins, grain)

3. We have it, but it’s malformed at creation (input controls, free text, weak validation)

4. We have it, but the platform breaks trust (fidelity gaps vs source, missing lineage, silent transforms)

“Data quality refers to the usability and applicability of data used for an organization’s priority use cases…” — Gartner

Increasing data quality used to be a very hard business case to write. Sometimes with legacy tools, adding one input control or validation would cost $1MM! Business teams would scratch their heads because it seemed so simple, but whether large enterprise IT processes caused the price to balloon, or vendor costs to do custom work added to the ticket, it was very hard to justify the amount.

Nowadays two things are changing the equation very quickly:

1. The importance of high-quality data in operational systems that an agentic AI will depend on.

2. The rise of AI coding tools that significantly reduce both the coding time and the testing time required to make changes.

Suddenly the business case is simultaneously becoming much more important and much easier to attain for organizations. This will change the game.

Subscribe now

“Quality is made at the factory”: input controls are the shift-left move

“Finding and fixing flawed data soon becomes a permanent fixture.” — Thomas C. Redman, Harvard Business Review

Agentic AI will operate systems, which means data quality becomes operational safety, not reporting hygiene. They’ll also be swimming in the analytical data lake to surface insights that drive business decisions and performance. Interns won’t be able to “fix numbers” before being seen by an executive. Input controls in the operational environment will function much like Knowledge Graphs will in the analytical one.

“Shifting left” to ensure superior logic is built into applications will be paramount. I asked a group of CTOs recently: How many web forms in the past week let you type letters into a phone number field? The answer is most likely zero, because JavaScript input controls have been around for decades! The same needs to be thought about in our operational environment and AI coders will close the gap on that in the near term future.

The business cases should write themselves.

“Poor data quality costs organizations at least $12.9 million a year on average…” — Gartner

That quote is from 2020, predating generative AI. That means that they were focusing on reporting, analytics, and machine learning use cases. Take this estimate and scale it for a rapidly accelerating world of AI and business outputs.

Subscribe now

Six Actions that Executives Should Consider Today

Here’s where you should start if you’re ready to dive into tuning up your enterprise’s data:

1. Force the definition upfront: Every “bad data” complaint must declare which bucket it’s in and which use case is harmed. (Fitness-for-use framing.)

2. Shift-left with input controls: Put validation rules where data is created; stop funding downstream bandages as a “strategy.”

3. Make fidelity measurable: Reconciliation checks between source and Trusted Environment for key KPIs (especially “executive numbers”).

4. Invest in observability + lineage like it’s production monitoring: because it is.

5. Kill the master-sheet incentive: publish “trust scores” (lineage, freshness, known issues) so people don’t feel safer in spreadsheets.

6. Treat AI as a trust multiplier: no agent rollout without defined data contracts + controls + monitoring.

When someone says, “we have bad data!”…

Don’t argue! They’re probably right. However, dig deeper with a couple of probing questions. Get them to explain what they actually mean and are experiencing so that you can “route the call” appropriately.

Meanwhile as you’re looking at opportunities for future investments, push your teams to bring at least one idea that tunes up the source systems so that data is produced well so that when it flows downstream, everyone benefits. Taking this “root cause approach” will pay dividends on the investment when it comes to reduced time to produce reports, build machine learning or other AI models, train AI agents to do quality work, and expand your executive worldview through highly accurate and automated business metrics and insights.

Share

📌 THE POINT IS: AI gets exponentially more useful when you stop asking it to produce answers and start asking it to sharpen your thinking — one question at a time — until your real objective reveals itself.

Having a partner is way better than having a digital intern

In today's AI world, people have learned to ask their AI tool of choice to just execute something. This makes sense as early adoption patterns encouraged people to ask AI to write the first draft. However, in his book, “The Ai-Driven Leader,” Geoff Woods leans into advising readers to think of their favorite AI tool as a thought partner. When you tell the assistant to ask questions, it forces you to refine your point, get ready to write that first draft, and provide editorial guidance. This results in a much more inspired dialogue.

“How can AI help me do this?” - Geoff Woods

The innocent question that saved me hours of wasted effort

You can try this today: next time you're about to ask ChatGPT to build a business plan, strategic plan, coaching plan, etc., ask it to act as your thought partner and help you think through that activity. Give it background information, like you would any human advisor. Then tell it to ask you additional questions, one at a time, to help it fill in any other relevant blanks that you didn't think of (I recommend giving it a cap so that it doesn't get stuck in an endless question loop). Notice how different the process is; how many things pop into mind as you're answering the questions; and the quality of the output that it provides.

Subscribe now

For me, this was very relevant last week when I spent about 5 minutes writing a prompt (it was a long prompt because of the significant context that I provided). After I asked my AI tool to ask me questions, one at a time to act as my thought partner and agent co-builder, it immediately began with a clarifying question:

I stared at the screen…

Then out of the blue, I had this epiphany. I didn't need an agent to tell me I should delegate more or protect calendar time for strategic thinking. I already knew that. I was doing this for the exercise of building an agent, but this agent wasn't going to have any ROI.

Had I not asked the AI tool to be my thought partner, it would have dutifully started telling me how to build this agent. But by asking it to help me think, it very innocently asked me a question that opened my eyes. After sharing the epiphany with the AI, it told me that we should not continue building the agent at all.

Based on its question, I asked it to help me review productivity in relation to my strategic plan for the year. I got meaningful ideas in a one-page report and the activity took far less time! It was also a much more strategic exercise.

“Stop looking at [AI] as an assistant and… start looking at it as a thought partner.” – Geoff Woods

Subscribe now

Questions don't just improve prompts. They help leaders have those ah-hah moments that lead to inspiration

This idea of asking the AI agent to think about the context and activity first, then create its own questions for you to go deeper before presenting ideas is a strategic unlock. Like using ChatGPT or other tools in Study Mode, the AI tool now moves from a mindless answer machine that may be totally, confidently wrong to a thought-inducing tool which helps you discover insights you may have been overlooking.

In a recent exercise I did with ChatGPT, by asking it to help me think through a business problem in order to make a decision, the questions that it created for me were logical, changed based on my answers, and were mostly multiple choice in order to force me to decide (or give it a write-in answer but only if I had additional details to support it). The next question was always related to the answer that came before it, almost like the model was following a pre-defined decision tree that ultimately helped me realize where my own biases, preferences, and experiences were taking me.

It may be good to note here that using a prompting framework like C-R-A-F-T is particularly useful. Prompt frameworks help structure your thinking so you don't omit critical instructions to the AI.

Subscribe now

How leaders can apply these techniques to up their game with their AI thought partner

Leaders can apply these immediately:

1. Use Thought Partner mode before you build anything. If the goal isn’t crisp, automation just scales confusion.

2. Force prioritization early. “Help me think about the 3 outcomes that matter most.” beats “build me an agent.”

3. Add a “tradeoff question.” “What must be deprioritized for this to work best - give me options to consider?” (prevents fantasy planning).

4. Switch from “more productivity” to “better alignment.” My ah-hah moment above is a great example of this.

5. Log the questions, not just the answers. Over time, your best questions become your leadership operating system.

6. Teach your team these frameworks and patterns. It’s a cultural upgrade: better questions → better thinking → better work.

AI is moving from being your assistant to being your mirror + coach + strategic lens

Thought partner mode with AI tools moves from blind robotics to creative interactions that unlock insights. It doesn't just improve output quality; it improves decision quality. The next time you're sitting down to try something just to try it, think about your time's ROI and ask the AI to help you take your use-case to the next level! You might be surprised what it suggests. Then you can ask it to be your thought partner to work through the next steps and notice how different the experience is working with it. This is more than a productivity hack; it's a decision-quality upgrade.

Share

📌 THE POINT IS: AI agents can only be trusted in operations when their answers are correct, consistent, and explainable. SOPs teach agents how work should happen; knowledge graphs determine what is true, connected, and allowed. Together, they form the deterministic control plane that keeps probabilistic agents from freestyling your business.

Click the “Heart” icon and/or Restack / Share to show me that you vibe with this article! Thanks!

Generative and Agentic AI are confidently wrong without bounds and direction

We've all been there: you ask ChatGPT for the answer to some dilemma and it gives you a very confident and very wrong answer. You point out what's wrong, and it cheerfully admits that it was wrong and you were right and here are five reasons why it was wrong.

AI tools without bounds and a deterministic control plane can get you pretty far, but so will driving 100 MPH in a car with no seatbelts. Especially when we're industrializing AI agents for business usage, we need to remove opportunities for wrong answers, inconsistent ones, and results that are hard to explain.

It’s not about model tweaks, but about AI “infrastructure” that companies need to invest in alongside their key use cases. This isn't new; it's already codified in risk frameworks:

“The characteristics of trustworthy AI are integrated into organizational policies, processes, and procedures.”
– NIST AI Risk Management Framework (Pg 5, “Govern” § 1.2)

Subscribe now

Remember those SOPs? They contain the ‘how’, but lack the business context

In my last article, I discussed how critical it is to have documentation for your processes, but also to do classic process engineering to make sure they're optimized for your AI agents. This work is an essential part of the agent-building process. They will help move the needle on making sure your agents' steps and outputs are done correctly.

“Without a formal process model, it is impossible to reason about the correctness of process executions.”
– Wil M. P. van der Aalst, Process Mining: Data Science in Action

While getting the steps down should be done first, telling the agent how to do a process doesn’t guarantee that it will get the right answer or that it'll do that process the same way every time.

I’ve created a couple of ChatGPT CustomGPTs and have given them specific instructions on how to perform activities. I find that they'll adhere to my instructions several times, but eventually they'll drift and add a different format or flair to the activity. That drift is harmless in personal workflows, but it's not harmless in regulated ones.

Knowledge Graphs are the control plane for Agents

Enter Knowledge Graphs (KGs). When your business partners build a KG for their department or processes, they're creating a model of the entities, relationships, and authoritative sources that are critical for successfully answering questions correctly, consistently, and in an explainable manner. They are the deterministic control plane for your agents that bound, guide, and give them context about your business that they need.

Here's what that looks like in practice:

• Correct: answers are grounded in authoritative sources and versions (tables, databases, etc.)

• Consistent: the same path to the same answers every time, regardless of how you ask the question

• Explainable: if asked, humans can look at the reasoning path along the KG chain and explain to an auditor why certain sources are being used

What's important to note is that the SOP tells the agent how to perform a process. The KG then tells it where to go to perform the various steps. The map gives everyone confidence that it's bringing back the right answers based on trusted data repositories.

Subscribe now

The effort brings multiples of benefits back

It takes a village of business partners to help build the nodes and relationships between the data. I'm not trying to hide that; while Tech plays a role, the business does the important work. The good news, though, is that the benefits multiply every time you connect KGs together to create a larger map that links the enterprise together.

With a unified KG, you can answer questions and perform actions related to:

Customer → Account → Product → Policy → SOP → Allowed Action

Each node above is related to the next and questions related to Policies can be tied back to Accounts and Customers. Maintenance can be performed on an Account by identifying the Policy number or Customer name. Customers can be authenticated using Account and Policy numbers (two-step) when requesting something like a last name change. And all of those operations are connected to their proper SOPs.

All of a sudden, when the nodes are connected, exponential numbers of questions can be answered that weren't possible before by getting more and more specific about what's happening in the world. You can build KGs iteratively over time. You don't have to do it all at once! This is important: don't get overwhelmed and table the whole project because it seems too daunting.

As you connect up different pieces (e.g. a KG built by the Claims team and a KG created by the Customer team), the possibilities explode. Suddenly, proactive agents, dashboards, and actions can be set to alert all customers in Northeastern Texas that have an Auto policy when a hail event is detected 30 miles away. And the list goes on the larger the KG eventually becomes.

The best part is, when an auditor or state regulator shows up to ask about why those alerts were sent, the business and Tech teams can open the graph, explain each node, highlight relationships between the nodes, and when viewed side-by-side with the agent's thought chain, the reason becomes easily clear. This is where many AI demos without KGs fall apart at a critical moment.

Subscribe now

5 things you can try today to stress-test your agents, chatbots, or everyday AI tools

You can put these concepts under the microscope at your company by trying these five things:

1. Run the same question three times
   Ask your agent a real customer-service question (e.g., eligibility, coverage, or next-best action) three separate times, ideally across channels or sessions.

2. Compare the answers side by side
   Do you receive the same conclusion each time, or does wording, logic, or outcome drift?

3. Demand the policy or source of truth
   Require the agent to identify which policy, rule set, or authoritative document version was used to generate the answer.

4. Ask for the reasoning path
   Request a clear explanation of how the answer was derived: which entities were involved, which relationships applied, and why alternatives were excluded.

5. Stress-test cross-domain consistency
   Ask the same question from the perspective of customer service, underwriting, or claims. If answers diverge, your enterprise meaning is fragmented.

Look out for these gotchas:

• If the answers come from the wrong source → you lack correctness

• If answers change when asked differently → you lack consistency

• If reasoning doesn’t show you enough detail → you lack explainability

Next, try to ask your agent to answer questions that you know require cross-departmental data. It may be very telling to see how much it hallucinates or assumes relationships that aren’t there. If you have to create huge prompts every time you ask a question to get closer to the answer you need, that’s where knowledge graphs make a huge difference.

Subscribe now

At this point, the pattern should be pretty clear

In order to get correct, consistent, and explainable answers from AI, you need:

• SOPs that show how work is supposed to happen

• Knowledge graphs that capture what is true, what is connected, and what is allowed

• Agents that do the work, using judgment, inside those guardrails

This is not a philosophical debate about AI: it’s an architecture choice.

If you want to maximize investments in AI and reduce unnecessary human-in-the-loop effort, answers need to be correct, consistent, and explainable every time. This is especially true when customers, auditors, or regulators ask questions.

Share

Click the “Heart” icon and/or Restack / Share to show me that you vibe with this article! Thanks!

📌 THE POINT IS: AI agents can’t freestyle your business. Deterministic SOPs are the guardrails that turn probabilistic agents into reliable digital workers. If you don’t have clear process truth, you don’t have an agent strategy—you have a risk strategy.

I've been in a lot of meetings recently where people talk about AI agents as being able to do everything in your business on their own with autonomous thinking. While that's the goal of agents, I think the baby has been thrown out a bit with the bathwater on the thinking here, and that's why many AI initiatives fail.

Subscribe now

Agents are “probabilistic” models meaning that they don't follow a set of instructions step by step. They act like people do. They learn a process, start to understand how something works, and then they figure out patterns, trends, workarounds, and hacks to get the work done, just like people do.

This is a stark contrast to software of yesteryear. Software programs are essentially sets of sometimes millions of “If…Then” logical statements that execute a certain way and can predictively accomplish a task based on instructions. These “deterministic" systems are excellent for storing business logic, guiding a user through steps in a workflow, and capturing data about the process being performed. However, they aren't autonomous, don't think, and generally require a person to operate them.

So why do we need both? Why can't agents just…do things?

The answer is the same as, “why can't we just have people…do things?” People need deterministic workflows, and step-by-step instructions, especially when learning a new process or operation. In business parlance, a set of business rules and instructions for a process is called a Standard Operating Procedure (SOP). It turns out that AI agents also need the same! Many businesses, especially large and established ones, have these in spades.

Subscribe now

I was amazed at how well Bank of America's operating departments had painstakingly documented their processes. They documented the steps, technology used, business rules, roles that do each step, and so on. This is a treasure trove of great data for humans to consume when learning how to perform the Bank's many different transactions.

These are human-centered, deterministic systems! That's right: people can and do utilize deterministic models all the time. Anytime you open a box of furniture parts that need to be assembled, you're using a deterministic system to help guide you through an operation. It's how we learn.

But I thought AI could just…learn on its own?

In a business setting where you want to teach an agent how to perform a standard operation so that it can replicate that transaction in split-second time across thousands of clients, you need a little more say in how and what it's learning. Enter SOPs and the criticality that they bring to your environment. SOPs will help agents think about how to approach situations, how to complete transactions, how to work with tools in your organization. They'll point your agents to the right operational system / application in your tech stack and help it determine which buttons to push (or API values to pass). They're essential!

Subscribe now

Sweet! So I just need to point it to my folder of SOPs and let it rip??

Well…yes and no. I mean you certainly can do that and the agents might be able to learn how to perform transactions with lots of hoops and hurdles in them, because that's how we architected processes relying on dozens of humans to do small parts of them. The real unlock, though, is in good old-fashioned process re-engineering first. It’s time to crack open your Lean Six Sigma book and try the following:

1. Pick a process to overhaul for AI agents to perform

2. Evaluate each step for effectiveness

3. Determine in a world without humans how the steps could be performed

4. Modify the process flow and SOP to account for AI agents using APIs, MPC or other computer controls

5. Run an experiment with full AI agent operation of the process

There's true power in digital workers performing digital transactions. Don't limit yourself!

Once you've reworked and “swizzled" your processes to organize the steps into seamless digital transactions, controlled and coordinated with APIs or an AI Gateway, then you're ready to seriously cut out complexity, time, and the need for intense human oversight.

Don't forget: the applications you use to run your business today are deterministic tools of tomorrow!

If you've spent millions investing in solid processes and systems, your investment will truly pay off! You don't need to create agents and then sunset all of the software you've spent millions of dollars perfecting. They'll still capture critical data elements and guide logic for your promotions, exceptions, and task steps. They'll continue to be used by your probabilistic humans and digital workers.

Subscribe now

Think of deterministic systems / data and probabilistic workers as two strands of a double helix. They're connected by the work and transactions that are performed but they're integral to making sure each other is informed, empowered, and capable of building the lifeblood of your company.

Try this with your team right away:

Step 1: Pick one real, customer-impacting process. High-volume. Low-to-medium risk. Something the business actually runs every day.

Step 2: Pull the SOP. Is it current, explicit, and complete? Or does the process just live in people’s heads?

Step 3: Separate digital steps from human workarounds. Identify what could run end-to-end digitally versus what exists only because humans were needed.

Step 4: Define safe autonomy. Decide where an agent can act independently, where it must pause, and where a human must intervene.

Step 5: Run the experiment. See what happens when you wrap an agent around the process given your revised SOP and the guardrails you put in place.

Did this article strike you as handy? Give it a like by clicking the heart on top!

Share

Click the “Heart” icon and/or Restack / Share to show me that you vibe with this article! Thanks!

📌 THE POINT IS: entry-level jobs won't disappear…a range of experiences will continue to be crucial in any company. But what the entry-level associates do will be 3-5 years more advanced than what they are today. AI will empower a broadening of products offered and work done by people as we get great at identifying human jobs vs. AI jobs.

In my role as the Head of Enterprise Data, I'm involved with developing our plans to build upskilling programs for everyone from entry-level to experienced associates. The more I discuss the role of AI at companies, the more I'm seeing that although today's philosophy and assumptions are that the entry-level roles are being targeted, it's really the middle-manager role that will probably be phased out before that happens.

​“Junior staff might have spent their early careers doing routine work... Many of those tasks are now handled quickly by AI tools.” — Mindset AI Podcast

Subscribe now

Let's take a closer look: 3 year horizon

Right now companies are looking at AI in terms of traditional ROI with the “R" really targeted to increasing efficiency and thus reducing headcount. The problem with that relates to Jevons Paradox. Companies who “do it right" will find that maybe they tighten their belt with some hiring, but really the “R" will be in expanding their market or market impact through new products, offerings, or capabilities.

So what about entry-level talent? Stanford research shows a 13–16% decline in entry-level roles in AI-exposed fields. Meanwhile in the UK, graduate roles in the Tech sector were cut by 46% from 2023-2024. The impact of these cuts on companies who aren't training new entrants to the workforce are going to explode as we think about the “aging workforce” problem that's already starting to breach the surface. Baby Boomers have largely been retiring over the past several years with many more soon to come. Early Gen Xers are also starting to retire now. Very quickly over the next decade or two we'll see most of the leadership teams at today's top companies exit the workforce. While that could result in a “pushing up" of the current workforce today, not having anyone to support the operations of a company will clearly lead to disaster.

Subscribe now

We need to be preparing today to maximize both AI and human workers

The reckoning that we all have to have as leaders now is that there are “human" roles and there are “AI" roles. The sooner we can draw a line in the sand at our companies the better. Here's an example: recently I was shopping at the NFL Pro Shop at Bank of America Stadium in Charlotte. There were people on the floor helping customers find sizes, reach high-up merchandise, and answer questions. When I got to the checkout line, there was one person at the beginning of the line handing out clear bags and instructing customers to put their items into the bag in order to checkout. That sounded backwards, but when I got to an open register, the computer told me to put the bag into a basket. The items were automatically scanned using computer vision and the total appeared on the screen. With a tap of my digital wallet, I walked out having experienced the most seamless checkout process ever. The lesson was clear: humans were on the floor doing human work (sales, sizes), while AI handled the transactional work. It maximized the ‘R’ on their ‘I’ without turning the store into a soulless vending machine.

Even if that's true, what about the longer term impact?

Picture this, your company, 2035…a new batch of marketing freshers are arriving. In the old days (2025…), a brand new marketing associate may spend a lot of time doing data input, web searches, maintaining mailing lists, and other menial tasks that are “important to learn the ropes". It could take 3-5 years for that junior employee to get exposed to more higher-order and higher-impact responsibilities such as performing in-depth market analyses and interacting with customer data through surveys or product research. In that timeframe, by the way, companies lose sometimes up to 80% of a hiring class to other companies who may be willing to pay more just to fill their rosters.

In 2035 brand new marketing associate will start with an AI buddy that onboards them, trains them, and gets them familiar with the tools and information needed to be effective. The AI will do research, data entry, and mailing list management. The first year associates will be empowered by AI to be product owners, product managers, product analysts, brand-value contributors, or design consultants. Jobs that would have taken years to break into will now be part of the entry-level expectations because AI will take care of those menial tasks that don't really “require” humans.

Subscribe now

It's important to hang onto the concept that “humans" should continue to be the “thought leaders” with AI supercharging them as “thought partners".

​“By 2026, 20% of organizations will flatten their structures, removing over half of middle management positions.” — Gartner

With flatter organizations and more associates working on value-added work, relying on AI to handle the tasks that aren't as high-impact, senior leaders will need to be able to be thought leaders that can also direct much larger teams of humans and AI. The thought leader skillset will be crucial because AI associates will bring edge cases, exceptions, and more thought-worthy scenarios to the human's attention. Meanwhile, there will still need to be an array of experience levels across these teams so that humans can direct less-experienced teammates and ultimately guide an ever-widening swath of new products, services, and projects that just weren't able to get off the ground with much more limited, expensive human resources in the past.

This sounds lovely but you still need less people which means higher unemployment

A couple parting thoughts to chew on with this one. First, global birthrates have halved in the last 70 years, meaning the future workforce will naturally shrink. This makes AI supplementation a necessity, not just a choice. It may end up being great timing that we're walking into a situation where AI can supplement workforces considering there will be much smaller ones in the future as this trend continues.

Subscribe now

But there's another thread to pull here: as companies ultimately become a projected 5-10x more profitable per employee, we may start to see the return of defined benefits plans such as pensions return to the stage. The government and social security will not be able to handle the impact from people retiring or simply not getting jobs as the AI workforce expands thus pushing some of the responsibility back to companies. As companies expand, though, too, they will be just as invested in making sure that people have money to spend on the products they develop so that the companies can be profitable. It does seem like a bit of a ouroboros, however I like to think of it as a virtuous cycle instead of a destructive one.

​“Artificial intelligence has the potential to help facilitate more sustainable pensions... unlocking more ways for people to remain engaged in the workforce for longer.” — Mercer

The other reason that companies will end up taking on a role in keeping people funded is that with a shrinking workforce, just imagine how valuable those humans will end up becoming! Competition often drives creative compensation and retention packages to be created and how attractive will it be for new associates to know that they'll be secure later in life when they ultimately go out to enjoy their golden years.

One thing is for sure: the future is coming

There's a LOT we don't know about how AI in the workplace will continue to shape our society. One thing is for sure: it's coming. Companies who figure out the optimum roles, boundaries, and accelerators for humans will win the long race, while others will see a short-term boom and a long-term flop.

Share

I am an AI artist.

Phew! There. I said it!

I’ve used tools like Donna AI and Suno to create music from lyrics that I wrote. Actually, those lyrics started off as poems that I wrote over the past decade or so. I got the idea one night to try to take a poem of mine and throw it into Donna AI to see what it would do with it.

I quickly learned that that doesn’t work.

I went back to the drawing board to turn the poem into a set of musical lyrics. I added instrumental and vocal cues to direct the digital performers on how to create the music. Then after probably 30-50 iterations, tweaks, changes, suggestions, and instructions, for the vocalist and the band, I finally heard a version of the song that sounded like, well a real track (by the way, for anyone who has been in a vocal or instrumental group, I hope I didn’t just give you PTSD!).

The point is that it was my work, my labor, my idea, my intellectual property, my directions, my ears, my brain, and my soul that went into all of that…just like any other songwriter in the world. Why is “songwriter” in italics back there? Simply because most songwriters don’t play a band worth of instruments. Many don’t sing…which is why they work with a vocal artist and a musical group to ultimately bring their vision to reality.

Be sure to click the “heart” on this article and if you’re not already:

Subscribe now

To me this kind of work is uncovering the difference between AI slop and human-with-AI created art.

I can hear it now: “That’s just not the way it goes! AI can’t create art even if your hand is on the brush.”

Let’s take a walk back in time, shall we?

The year is 1982 and this brand new technology is hitting music stages around the world. It’s electronic and odd and makes all kinds of sounds that aren’t real. The synthesizer was so much as banned by the UK Musician’s Union in order to protect jobs. Queen had a “No Synthesizers!” advertising campaign because they were soulless and didn’t create “real” music. By 1988, though, Mannheim Steamroller was releasing one of their famous holiday albums. I was listening to “Carol of the Bells,” on my commute this morning thinking about how powerful that song is and it's completely made with synths.

How about the rise of Auto-Tuners in the 90’s? At one point these handy little tools helped studios configure their equipment to align to the artist’s voice. However artists like Cher and Kanye West started using these tools to distort their voices on purpose, thus making it possible to create new tones and to express their souls more completely in their works.

As we look at musical history, we can see many examples of where a new instrument or tool came onto the scene and met instant rejection by existing artists who felt like their work was more “real” than the artists using these new technologies. We haven’t even talked about the drum machine “panic” in the early 80’s, but I think you catch the beat 😁

By the way this is true for other art forms…did you know that when cameras came on the scene, painters said that photography wasn’t real art because “the machine does the work”! In 1859, the famous poet and critic Charles Baudelaire wrote that photography was

“art’s most mortal enemy’ and a refuge for ‘painters too lazy to complete their studies.”

He sounded exactly like Nick Cave does today.

Yeah but I mean there is such thing as AI slop…right?

Absolutely!

I think the problem with AI in art is the same as the problem with AI in just about everything else: if humans give their agency away, let AI create an artifact, and don’t take ownership of the final product, we often see slop arise. I’ve definitely just told Donna to create a silly song in some style about my dog and let it rip. Now I actually like that song about my dog…but that’s not the point. The song isn’t going to win any awards and if I put it online, it’d get ripped to shreds. But that’s the point: I was playing around, maybe getting some ideas, learning how to use the tool, and I wasn’t trying to pass it off as Beethoven.

Nick Cave came down hard on a fan who asked ChatGPT to generate lyrics in Nick’s voice. He said

“Data doesn’t suffer. ChatGPT has no inner being, it has been nowhere, it has endured nothing... This song is bullshit, a grotesque mockery of what it is to be human.”

…and well, he’s right! In my example above, I wrote the lyrics. I translated my lived-experiences into a workable, rhythmically acceptable (and dare I say enjoyable) set of words that convey deep meaning. I did not give my agency away. I did not give my creativity away. In fact it took days to finesse the work into something that sounds somewhat professional. Just like the early photographers who were told that they don’t draw every line of their photographs so they can’t be called art, I didn’t write every note of those songs, but I framed up the emotions and instructed my tool to help take that out of my soul and into this reality.

Be sure to click the “heart” on this article and if you’re not already:

Subscribe now

This distinction between outsourcing agency (slop) and directing outcomes (art) doesn’t just apply to Suno; it also applies to your GitHub repositories.

It’s always good to make sure we’re lifting our heads up to hear what’s happening in the AI world (or otherwise!) in order to build perspective. This is, at its core, a discussion about how AI tools should be used in our lives.

Anthropic recently said that 90% of the code being written is done by Claude Code. The role of the software engineer is fundamentally changing to a supervisor and a guide over someone who puts fingers to keyboard to belt out a few thousand lines of instructions. Is that wrong? Is it AI slop that’s being created? Are the humans doing just as much work reviewing the code and making sure it works?

These are interesting questions, maybe, because major vibe-coding platforms, such as Cursor, have their own test beds and automation tools built into the engines too. In fact, one of the earlier ways that AI tools for coding (like GitHub Co-pilot) were used was to create test cases and test scripts based on code (in many cases on older code or code that was being updated for a release). So why not just let these babies rip and do all of the work? What do you mean 90%? Why isn’t it 100%?? Will the software engineer truly go the way of the Dodo by next year?

Clearly we’re not at 100% and frankly when I actually put my hands on the keyboard to vibe code, many times it doesn’t even run. So to me there is clearly hype out there; clearly we’re somewhere more realistically around 50% of code generated; that code may or may not be actual, production-grade, enterprise hardened code (probably not); and in fact there’s a much longer tail that we have to look forward to. This isn’t just speculation. Last week during the AI Daily Brief we heard about user stats starting to slow on platforms such as Cursor and Replit because the whole vibe-coding “thing” isn’t going as smoothly as everyone thought. A recent study by METR (Model Evaluation & Threat Research) found that AI coding tools actually slowed down experienced developers by ~19% because they spent so much time reviewing and fixing the code. We’re seeing the “Productivity Paradox” emerge…humans can’t just watch the loop. They need to be in the loop. That’s not to say that code won’t be created by AI-developers, but much like my music example above, the better strategy for now anyway is still to use AI tools as thought partners and let them be “tandem coders” with seasoned software engineers.

Thought partner you say? That’s a new one…what do you mean by that?

I’d hate to steal the thunder from Geoff Woods in his book, the Ai-Driven Leader, because that book is a must read for anyone using AI tools for basically any reason. In that book, though, Geoff presents many eye-opening and simple concepts on how to use AI tools to truly assist you, the thought leader. I’ve tried implementing a few of the prompts that he provides in the book and seeing how ChatGPT and Gemini behave differently when they’re put into this “frame of mind” is nothing short of mind blowing.

We are moving quickly…

…into an era where our everyday toolset is transforming right before our eyes. Whether it’s the new “camera” for current day photographers or the “synthesizer” and “drum machine” of the mid-20’s for today’s audio artists, new tools will define new techniques, genres, and outputs. AI slop is real…but it’s not fair to throw that label out to everything that AI touches. Humans will create works with AI tools just like they started creating new vocal intonations with auto-tuners.

In the workplace, we’ll see people succeed by working with AI tools, using them as tools and not expecting that they’ll take away all of the value that workers provide. Don’t get me started on my discussion about the future of work…that’s for another article. Suffice it to say, though, the faster we learn how to think 10x differently, the smoother we’ll move into the AI-powered future…with human hands hopefully on the wheel!

Be sure to click the “heart” on this article and if you’re not already:

Subscribe now

And / OR

Share

📌 THE POINT IS:
Customers own their data. Businesses are merely custodians. Building a Trusted Data Environment is no longer optional; it’s how organizations earn the right to use that data in the first place.

The Ownership Fallacy

For decades, “the business” in large enterprises have talked about owning their customer data. But the legal and ethical landscape has changed. Regulations like the GDPR, California Consumer Privacy Act (CCPA), Brazil’s LGPD, and India’s DPDP Act all enshrine the idea that the individual owns their personal data. Businesses merely hold it on loan.

Under Article 17 of the GDPR, individuals have the “right to erasure” — the ability to request deletion of their data at any time. If businesses truly owned customer data, that right wouldn’t exist. The very fact that it does underscores the shift from corporate ownership to customer custodianship.

“Trust is the ultimate currency in the digital age.” – Satya Nadella, CEO, Microsoft

When leaders fail to internalize this, the results can be catastrophic. Data breaches, AI bias, and misuse of personal information don’t just create legal risk; they erode trust. And trust, once lost, is nearly impossible to rebuild.

Subscribe now

Trust as the New Asset Class

Analysts at Deloitte and McKinsey consistently find that organizations with high trust scores outperform peers in customer retention, brand equity, and growth. According to Edelman’s 2024 Trust Barometer, 71% of consumers say they’re unlikely to buy from a company they don’t trust to protect their data.

Data, then, isn’t the new oil — trust is. The data itself has little value without a trusted relationship behind it. In the AI era, where companies ask customers to share more personal and behavioral data than ever, trust becomes the foundation of every model and every algorithm.

This brings new paradigms and challenges old assumptions for tech leadership and especially for business leadership teams. In the past, business teams assumed that tech was doing what they needed to in order to protect “their” data. Tech on the other hand often faults the business for not providing the funding required to adequately do so, and even when there is funding, technology is constantly under pressure to “go faster” to deliver business results. This is the mindset and culture at large companies that I’ve worked at and or studied till today. As we fly into a new world of agentic AI, autonomous systems, and intelligent decisioning, though, all of this is called into question. Enter the conversation about Data Products, continuous funding models, cross-functional teams that are not charged to deliver “business results”, but to deliver “customer outcomes.”

The whole way of running businesses powered by technology is turning upside down.

Shared Stewardship Across the Business

Too many organizations still treat data protection as IT’s responsibility. Every business function though — from marketing to finance — touches customer data and therefore shares accountability for how it’s handled.

The Chief Data Officer plays a vital role as educator and architect, but business leaders must lead the cultural charge. They are the ones customers will hold responsible when trust is broken. This shift requires a new kind of leadership: one that sees data stewardship as core to the customer experience, not as a back-office compliance function.

“If data is the lifeblood of AI, trust is the oxygen that keeps it alive.” – MIT Sloan Management Review, 2025 AI & Data Leadership Report

At Nationwide, we often remind ourselves that we’re not just an insurance company — we’re a protection company. That promise extends beyond physical assets to digital ones. The same way we safeguard our customers’ homes and livelihoods, we must protect their information with the same care and diligence.

Subscribe now

That’s one of the reasons I’ve been leading the charge on building what I call a Trusted Data Environment. On the surface, it’s about ensuring that we can trust the results of our analytics and AI systems. But at its core, it’s about something deeper — ensuring that customers can trust us to use their data responsibly, transparently, and securely. Because if they can’t trust us to protect their data, why should they trust us with anything else?

Building this environment isn’t just a technology project. It’s a cultural one. It’s causing us to relook at development practices; cybersecurity policies; the partnership between the business, their tech counterparts, enterprise data, and cyber teams. Everything is getting called onto the carpet as we careen towards a future that demands high-speed throughput by autonomous machines that ultimately we all need to be able to trust as they help us simplify and enhance the customer experience.

But most of all, this all demands that business and technology leaders work together to define what “trust” even means in operational terms — from how data is shared to how models are validated and monitored. Recently we had a discussion during which I reminded my peers and senior leaders that although we’re spending a lot of time talking about building AI solutions, we can’t forget that we also need to think about securely running those solutions over their lifetime. This means investment in runtime monitoring and other automated controls that will maintain trust in our system without relying on humans, who even when they’re in the loop, may not be able to see or respond fast enough to an issue bubbling up.

The AI Trust Loop

AI depends on data quality and provenance. But data quality requires that business teams and technology partners are building the controls up front into their systems. That creates a self-reinforcing loop: when you demonstrate to customers that you can be trusted with their data through good hygiene and system controls, they share better data; better data powers more reliable AI; reliable AI reinforces customer trust.

Conversely, when data is misused or mishandled, the loop collapses. Models degrade, customers withdraw, and regulators step in. Gartner predicts that by 2027, 60% of enterprises will use trust metrics as a core KPI for AI performance. This isn’t a compliance trend; it’s a business survival strategy.

Subscribe now

Leadership Imperatives

1. Replace “ownership” with “custodianship.” The words you use shape your culture.

2. Make trust measurable. Track customer data sentiment and correlate it with retention.

3. Elevate the CDO as a cultural leader, not just a technologist.

4. Treat every data touchpoint as a trust transaction.

5. Embed privacy-by-design into all AI initiatives at every level of the development and environmental stack.

In the end…

The strongest moat in the AI age won’t be proprietary data or faster models. It will be earned trust. Customers are paying attention. They know their data is theirs. The question for today’s leaders is simple: what are you doing to earn and prove that they can trust you enough to let you use it?

References

1. General Data Protection Regulation (GDPR), Article 17 – Right to Erasure. https://gdpr-info.eu/art-17-gdpr/

2. Edelman Trust Barometer 2024. https://www.edelman.com/trust/2024/trust-barometer

3. Deloitte Insights – Building Trust in AI. https://www2.deloitte.com/us/en/insights/focus/trust/ai-trust.html

4. McKinsey Digital – The State of AI in 2025. https://www.mckinsey.com/capabilities/mckinsey-digital/our-insights/state-of-ai-2025

5. MIT Sloan Management Review – Five Trends in AI and Data Science for 2025. https://sloanreview.mit.edu/article/five-trends-in-ai-and-data-science-for-2025/

6. Gartner Predicts 2027 Trust Metrics in AI. https://www.gartner.com/en/newsroom/press-releases

Share

Subscribe now

📌 THE POINT IS: AI doesn’t “bottleneck”—it chokes on bad, incomplete, or stale data. To fill gaps it hallucinates data or it answers with old data and assumptions. At scale, that’s a governance, brand, and financial nightmare waiting to happen. You can proof against that by investing in your Trusted Data Environment, leaning into high-velocity data plumbing that's governed, observable, discoverable, and powered by an intelligent semantic layer.

The AI boom has captured every boardroom’s imagination. Executives are funding pilots and demos, believing that generative models will give their organizations an “intelligence” advantage. Yet most implementations still sit on top of brittle spreadsheets and siloed dashboards. When outputs look plausible, we assume they’re correct. When they’re wrong, we call it hallucination. In reality, bad data is choking AI. Models trained on incomplete or outdated data make up answers to fill gaps. Many organizations don’t even know their data is bad—81 % of surveyed companies trust their AI results despite fundamental data inefficiencies, and the average firm loses 6 % of its annual revenue as a result.

Subscribe now

Why AI chokes on bad data

Generative models are designed to answer questions even when context is missing. When underlying data is incomplete, biased or stale, models guess to fill the void and fabricate “facts.” A recent analysis noted that language models trained on incomplete datasets, biased sources and outdated information generate hallucinations that compromise business operations. When important details are missing, the model tries to “fill in the blanks,” producing plausible‑sounding but fabricated outputs. These problems show up not just in chatbots but also in code generation, marketing content and internal analytics.

Poor data quality isn’t a niche problem. A 2025 survey found that 84 % of IT leaders consider a configuration management database (CMDB) essential for decision‑making, yet only 17 % say theirs is fully accurate; 64 % of IT teams still haven’t adopted AI because of data quality and security concerns. At the same time, junior data workers report that nearly half of their time is spent cleaning data and fixing broken pipelines. More than two‑thirds of technical executives admit that their teams struggle to access all the data needed for AI programs (69 %) and to cleanse it into a usable format (68 %). No wonder 42 % of respondents experience data‑driven hallucinations.

Data stack complexity slows everyone down

Even when data exists, it’s often stuck in a tangle of tools. A 2024–25 report on data architecture found that 85 % of data teams cite tool integration as a top challenge, and 63 % spend more than one day a week maintaining their stack instead of delivering value. Teams manage five to ten tools just to move data around; more than 40 % of their time is spent switching between platforms. As the report notes,

“architecture complexity is the invisible force slowing down most data teams”.

Because current stacks focus on storage and movement rather than context and governance, teams must add lineage and metadata manually. The same study reports that 65 % of professionals believe combining strong data models with business‑ready data products is essential for compressing time to insight.

The result is a vicious cycle: messy data leads to hallucinations; hallucinations erode trust; engineers spend more time firefighting than innovating; AI adoption stalls.

Subscribe now

Real‑time data hubs: beyond dashboards

Most enterprises still think of data as something you analyze in batches. But the high‑speed digital economy requires real‑time, data‑responsive systems. In 2025, companies must move beyond static dashboards and reports and build fast‑velocity data hubs that ingest, process and publish information as it is generated. Analysts at Value Innovation Labs argue that real‑time data solutions are now the foundation for smart enterprises because they enable companies to react instantly to opportunities and risks.

My CIO friends out there are thinking: “Real-time data flows for everything? But that's so incredibly expensive. What is the business case for this? How can we afford it?” That's super fair and in today's reality, maybe enterprises do need to prioritize which workflows warrant this level of streaming. But the days of real-time transaction processing, powered by AI, are coming and they're coming fast. Companies that are at least prepared for this architectural shift will be far better off when they integrate more and more AI agents into their workforce.

Traditional batch processing is no longer sustainable; data must be collected and acted on the moment it is produced. Real‑time platforms comprise ingestion pipelines, stream‑processing engines (such as Kafka or Flink), analytics layers, automated triggers and AI‑driven decision engines. These hubs support operational use‑cases (yes you read that right) such as:

• Customer experience: Live data lets brands personalize interactions on the fly and detect service issues as they occur.

• Operations and logistics: Real‑time routing and proactive inventory management reduce waste and improve service.

• Finance and risk: Continuous monitoring detects fraud instantly and improves forecasting.

• Cybersecurity: Live threat signals trigger automated defenses, which is essential as AI‑driven attacks accelerate.

These just‑in‑time data flows are critical if AI agents are to do more than make recommendations. Agents need high‑context, up‑to‑date data to take actions autonomously and avoid hallucinations.

A related topic that my colleagues and I have been talking about is the inevitability that AI will use data hubs like this for some workflows, but they may be interacting with live, operational systems for true, real-time agentic capabilities. There's still a strong case for prioritization today, but as companies' financial profiles shift in future years, making trade-offs to build more real-time pipelines will become more feasible.

Knowledge graphs: the Rosetta Stone of AI operations

As organizations wire up hundreds of systems, semantic consistency becomes the hardest problem. Fields labeled “subscriber,” “customer” and “account” might refer to the same entity, but AI cannot infer that without guidance. Modern knowledge graphs and ontologies solve this by providing a shared semantic layer that translates between systems. Totogi’s telecom ontology demonstrates how a knowledge graph built by AI can act as a digital Rosetta Stone for enterprise systems, automatically mapping disparate data and business processes into a common model. Once connected, every system interoperates through this ontology, allowing AI agents to understand that different fields refer to the same concept.

Knowledge graphs don’t just unify data; they also power orchestration. On top of the ontology layer, Totogi deploys AI “workers” that read documentation, write code and call APIs across systems. These agents compress release cycles from weeks to days by automating testing and coordination. More broadly, ontology‑based AI allows systems to communicate with each other and infer relationships, functioning as a Rosetta stone that provides richer context for decision making.

For businesses outside telecom, the same principle applies. Seth Earley, a knowledge‑engineering expert, describes ontologies as the “knowledge scaffolding” of the enterprise. By capturing relationships among products, services, roles and processes, an ontology becomes master data management for AI. It not only connects data models but also workflows and business logic, enabling AI to orchestrate tasks end‑to‑end.

Subscribe now

Investing in the plumbing: an urgent priority

The sobering reality is that most organizations are not ready for AI. A survey cited by Databricks reveals that only 22 % of organizations believe their current architecture can support AI workloads without modification. As companies race to build AI agents, infrastructure will be the biggest investment area. Effective agents need to work across diverse data sources and models, and they require an end‑to‑end data platform that unifies data, governance and model evaluation.

Robin Sutara, Databricks’ Field CDO, notes that:

“A successful AI strategy starts with a solid infrastructure. Addressing fundamental components like data unification and governance through one underlying system lets organizations focus their attention on getting use cases into the real world”.

Investing in the unglamorous plumbing will pay dividends as your company grows its AI capabilities. It will reduce maintenance costs, accelerate deployment and ensure that AI acts on high‑quality data. Executives are beginning to recognize the link between governance and AI reliability; unifying metadata and governance across data and AI assets ensures models take actions based on trustworthy, up‑to‑date information.

Conclusion: build the pipes before chasing magic

AI will continue to transform business processes, but it cannot do so on a foundation of bad data. Organizations are losing millions and wasting talent because they treat data plumbing as an afterthought. Modern data stacks are too complex, slow and context‑poor to power AI‑first operations. Building a trusted data environment, high‑velocity data hubs, underpinned by knowledge graphs that harmonize semantics across systems, is essential. This plumbing may not make headlines today, but it is the difference between AI that hallucinates and AI that drives profitable decisions. Invest in the pipes now so that when AI agents take the wheel, they are navigating with clean, contextual, real‑time data.

Subscribe now

Share

References

1. Socialnomics article explaining how poor data quality—such as incomplete datasets, biased sources and outdated information—causes hallucinations in language models.

2. Socialnomics article describing how models fill gaps when training data is missing, leading to fabricated outputs.

3. BetaNews report showing that IT leaders view CMDBs as essential but acknowledge that poor data quality hinders AI adoption; 64 % have not adopted AI because of data quality and security concerns.

4. CDOTrends survey revealing that 81 % of organizations trust AI despite data inefficiencies, losing 6 % of annual revenue; data scientists spend 67 % of their time cleaning data; 69 % struggle to access data; 68 % struggle to cleanse data; 42 % experience data‑driven hallucinations.

5. Modern Data Company report noting that 85 % of data teams cite tool integration as a top challenge and 63 % spend more than 20 % of their time on maintenance.

6. Modern Data Company commentary stating that architecture complexity slows data teams and more tools create integration headaches.

7. Modern Data Company observation that AI needs data enriched with context, lineage and governance; a unified layer ensures data arrives business‑ready.

8. Modern Data Company finding that current fragmented approaches require manual context; 65 % believe combining strong data models with data products compresses time to insight.

9. Value Innovation Labs article describing how real‑time data solutions enable enterprises to react instantly and provide continuous, up‑to‑the‑moment intelligence.

10. Value Innovation Labs article emphasising that the shift from batch processing to real‑time insights is necessary in a high‑speed digital economy.

11. Value Innovation Labs article outlining the components of real‑time data platforms: ingestion pipelines, stream‑processing engines, analytics layers, automated triggers and AI decision engines.

12. Value Innovation Labs article listing operational use‑cases for real‑time data—customer experience, operations, finance and cybersecurity.

13. Totogi blog illustrating a telecom ontology that acts as a digital Rosetta Stone by mapping disparate systems and processes into a common model.

14. Totogi blog describing how AI agents on top of the ontology layer can read documentation, write code, call APIs and compress release cycles.

15. Earley Information Science article explaining that ontologies allow systems to infer relationships and act as a Rosetta stone, providing richer context and enabling systems to communicate.

16. Databricks blog noting that only 22 % of organizations believe their architecture can support AI and that infrastructure will be the biggest AI investment area; a unified data platform is needed for agents.

17. Databricks blog highlighting the link between data governance and AI reliability—unifying metadata ensures models act on high‑quality data and reduces operational costs.

📌 THE POINT IS: Don’t just bet the farm on moonshot business cases for AI. Build a culture where potentially thousands of small, compounding AI wins free up time, fund the big bets, and enable real quality-of-life gains—including maybe even a path to the four-day week for some companies in the US!

Executives keep telling me that AI requires a use-case before they can get started on investments. I get it, you want that sweet ROI, but realistically, ROI on your AI investments will be realized through non-traditional mechanisms as we go forward. The data says durable value comes from a portfolio: pair Everyday AI (task-level copilots, workflow assistants, coding aides) with Strategic AI (new products, agentic systems, business model shifts). Leaders who balance both see results faster—and they keep them.¹²

“Companies that boost their learning capabilities with AI are significantly better equipped to handle uncertainty.” – MIT Sloan Management Review & BCG⁶

Subscribe now

Why Everyday AI must sit beside Strategic AI

Generative AI’s macro potential is undeniable—McKinsey estimates $2.6–$4.4T in annual value and a 0.1–0.6% annual lift to labor productivity through 2040.¹ That value isn’t magic; it materializes when thousands of routine activities get faster, easier, and higher quality. Randomized and field studies back this up: GitHub Copilot users completed coding tasks 55.8% faster in a controlled trial,³ and a large-scale NBER study of 5,179 support agents found a 14% productivity gain on average—34% for novices—when using an AI assistant.⁴ Microsoft’s Work Trend Index echoes this pattern across knowledge work: early Copilot users were 29% faster across writing, summarizing, and searching; 70% reported higher productivity.⁵

“Only 22% of companies have advanced beyond proof-of-concept to generate some value—and only 4% are creating substantial value.” – BCG, Where’s the Value in AI?⁷

Translation: moonshots without a base of Everyday AI often stall overall AI adoption and culture change. Leaders in this field focus on core processes and people, then scale.⁷ Meanwhile, Gartner expects agentic AI to permeate software quickly—by 2028, 33% of enterprise apps will include agentic capabilities—so the operational backbone you build now will compound.⁹

We’re seeing this now in my world. Whether it’s Copilot entering Microsoft’s suite of tools more fully; Databrick AI/BI Genie and Mosaic AI; Snowflake’s Cortex AI; AgentForce in SalesForce; or Workday’s newest AI features…these capabilities are becoming table stakes in large, enterprise applications across the board.

Everyday AI lifts productivity and quality of life

The productivity story matters, but so does well-being. OECD evidence finds most employees using AI at work report better performance, job enjoyment, and improvements to mental and physical health.⁸ And when organizations reinvest time savings into better ways of working, the gains can support schedule innovations. In the UK’s landmark four-day week pilot (61 firms; ~2,900 workers), researchers reported revenue held steady (up 1.4% on average) during the six-month trial, while stress and burnout dropped and retention improved—many firms kept the policy.¹⁰ ¹¹

“Company revenue barely changed during the trial period—rising by 1.4% on average.” – Autonomy & university research team, UK pilot report¹⁰

What’s interesting about this is that they did this pilot before AI was really introduced into companies. We’re just now seeing models that are capable enough to be junior co-workers. Imagine the impact on revenue in the UK pilot with today’s tools. We could have seen a much greater than 1.4% lift on that same 4-day time period.

Subscribe now

A brief, on-the-ground example

Inside my team, we’re building a chatbot agent tuned to our standards, accelerators, and “Golden Path” methods for data pipelines and curated models. In a Trusted Data Environment, this everyday assistant will onboard new developers, guide business data teams, and help contractors ship integrations the right way, faster. It won’t transform our whole business by itself, but it will accelerate everything else—shorter time-to-market for analytics and the next wave of Strategic AI implementations!

“Generative AI assistance disseminates the best practices of more able workers and helps newer workers move down the experience curve.” – Brynjolfsson, Li & Raymond, NBER⁴

The cool part is that we can roll out capabilities bit-by-bit to continuously show incremental value since this is a self-built, Everyday AI agent. It’s a digital co-worker that can be at everyone’s side when building data pipelines and data models. It will help us position our Trusted Data Environment as a rapidly growing platform.

Related whitepaper:

Make them sister initiatives: Everyday AI + Strategic AI

Here are some steps to take to get started and stay in check on both paths:

1. Stand up Everyday AI as a program with a product owner, budget, and backlog. Target the top 10–15 recurring tasks in each function (docs, meetings, search, coding, service). Prove time savings and quality deltas with before/after baselines (cycle time, defect rate, CSAT).³⁴⁵

2. Run Strategic AI in parallel with 1–3 high-value bets (e.g., agentic workflows, new AI-native features). Make sure executives are bought in and understand the commitment, but insist on milestones tied to measurable outcomes. Leaders that do this outperform.⁷

3. Invest in change management and learning loops. Train managers on “work redesign,” not just prompts, but keep the basics of prompt and context engineering at the forefront of your associate training programs. MIT SMR/BCG finds organizations that combine organizational learning with AI-specific learning are 1.6–2.2x better at navigating uncertainty.⁶

4. Measure and reinvest time. Redirect reclaimed hours to customer work and innovation that teams can actually deliver outside of just a couple-day hackathon setting. Where metrics hold—consider how you can give time back to associates in terms of extra, flexible time off, floating holidays, flex time for associates who perform work over weekends, or even if you’re bold, four-day workweek pilots in eligible teams.¹⁰ ¹¹

5. Strengthen the data backbone. Everyday AI without trusted, well-governed data is risk, not value. Often times this requires a whole different set of data hygiene controls, tools, and policies since this type of AI will often interact with your user-generated, office-document processes. Don’t skip out on making sure you have good guardrails setup here to minimize risk!

6. Ready the stack for agents. Gartner’s agentic wave is coming; design guardrails, identity, observability, and human-in-the-loop for any use cases now.⁹

Bottom line: Everyday AI compounds; Strategic AI differentiates. Together, they create the financial, cultural, and operational surface area to go faster and work better.

Subscribe now

Share

References

1. McKinsey Global Institute, “The economic potential of generative AI: The next productivity frontier.” https://www.mckinsey.com/capabilities/mckinsey-digital/our-insights/the-economic-potential-of-generative-ai-the-next-productivity-frontier (McKinsey & Company)

2. McKinsey Global Institute (PDF), “The economic potential of generative AI.” https://www.mckinsey.com/~/media/mckinsey/business%20functions/mckinsey%20digital/our%20insights/the%20economic%20potential%20of%20generative%20ai%20the%20next%20productivity%20frontier/the-economic-potential-of-generative-ai-the-next-productivity-frontier.pdf (McKinsey & Company)

3. GitHub (arXiv), “The Impact of AI on Developer Productivity: Evidence from GitHub Copilot.” https://arxiv.org/abs/2302.06590 (arXiv)

4. NBER Working Paper, “Generative AI at Work” (Brynjolfsson, Li, Raymond). https://www.nber.org/papers/w31161 and PDF https://www.nber.org/system/files/working_papers/w31161/w31161.pdf (NBER)

5. Microsoft Work Trend Index Special Report, “What Can Copilot’s Earliest Users Teach Us…?” https://www.microsoft.com/en-us/worklab/work-trend-index/copilots-earliest-users-teach-us-about-generative-ai-at-work (Microsoft)

6. MIT Sloan Management Review & BCG, “Learning to Manage Uncertainty, With AI.” https://sloanreview.mit.edu/projects/learning-to-manage-uncertainty-with-ai/ (MIT Sloan Management Review)

7. Boston Consulting Group, “Where’s the Value in AI?” https://www.bcg.com/publications/2024/wheres-value-in-ai (BCG)

8. OECD, “The impact of AI on productivity, distribution and growth” (2024). https://www.oecd.org/content/dam/oecd/en/publications/reports/2024/04/the-impact-of-artificial-intelligence-on-productivity-distribution-and-growth_d54e2842/8d900037-en.pdf (OECD)

9. Gartner, “Intelligent Agents in AI Really Can Work Alone. Here’s How.” https://www.gartner.com/en/articles/intelligent-agent-in-ai (Gartner)

10. Autonomy, “The Results Are In: The UK’s Four-Day Week Pilot” (full report PDF). https://autonomy.work/wp-content/uploads/2023/02/The-results-are-in-The-UKs-four-day-week-pilot.pdf (The Autonomy Institute)

11. University of Cambridge (Dept. of Sociology), “New results from the world’s largest trial of a four-day working week.” https://www.sociology.cam.ac.uk/news/new-results-worlds-largest-trial-four-day-working-week (sociology.cam.ac.uk)

Share Your Suggestions

I’m excited that dAIta POINTS has taken off ever since migrating over to the Substack platform earlier this year. Seeing new people subscribe each week after publishing and article and/or whitepaper is super energizing!

Now it’s time for you to write to me! Well, not really…I’m actually hoping you’ll take a quick survey to give me some feedback. The survey will take ~2 minutes to complete…no really, I swear!

2-Minutes I swear!

Please click the button and share your thoughts so I can keep this growing and focused on things that are of interest to you.

Have additional feedback, certainly send me an email! I’d love to hear more anecdotal, even constructive, feedback from readers!

Leave a comment

Thanks,

M@

Last Chance! :)

📌 THE POINT IS: On the surface, personal accountability seems to have vanished over the past few decades, but it hasn’t. Instead, it’s been diffused across socio-technical systems. Psychology trends, cultural incentives, and AI companions make it easier than ever today to offload blame, a natural human tendency. Leaders can reverse this by designing workflows for ownership, measuring process (not just outcomes), and building auditing into every AI-touched workflow.¹⁴

Personal accountability is not a fossil from the pre-AI era; it’s a casualty of diffusion that’s started well before AI became popular. While I don’t often veer from the professional world, this article is heavily influenced by the recent discussion about the emotional impact that chat bot models have had both on adults and children, in the workplace and at home. It’s a call to action that we as a society need to double down on relying on each other for support and companionship over virtual partners…it could be the difference between life and death.

Subscribe now

There’s a history to unpack related to the human nature around externalizing our actions

Across decades, young Americans’ locus of control has shifted outward—toward external forces. As one cross-temporal meta-analysis put it, “locus of control scores became substantially more external” between 1960 and 2002.¹ Combine that with the self-serving bias (credit successes to me, blame failures on the situation or someone else) and you have a cognitive slipstream away from personal ownership of our decisions and actions.²

Sociologists have described a growing “victimhood culture,” where moral status accrues to those most aggrieved.³ Parenting research points in the same direction: overcontrolling. Overly-involved or “helicopter parenting” at age 2 predicts poorer emotional regulation by age 5 and downstream social/academic issues. These are skills required to learn how to own one’s mistakes, and they start from early development.⁴ Meanwhile, grade inflation has marched upward while ACT scores stagnate or fall, muddying the signal that effort and learning still matter.⁵ Having said that, though, impressively, despite caricatures about “lawsuit-happy America,” civil trials have actually plummeted—from 11.5% of federal civil dispositions in 1962 to 1.8% in 2002, meaning that people take to social media to air their grievances about each other more often than they go to the courtrooms (something that mediation clauses in disclaimers, contracts, and waivers also play a role in).⁶

AI companions: anthropomorphism at scale

With that backdrop, let’s examine how we’re primed to humanize software (animals, toys, etc.). The ELIZA effect explains why we see modern chatbots as real people, especially given their language fluency and 24/7 availability.⁷ Character.AI alone has reported ~20 million monthly active users;⁸ one national survey found 1% of young adults say they already have an AI friend, 10% are open to it, and 25% believe an AI partner could replace a human relationship.⁹

These attachments have teeth. When Replika pulled erotic role-play, “users [were] in crisis,” describing genuine grief.¹⁰ University of Connecticut researchers chronicle teens redefining love as “easy, unconditional, and always there,” warning of social withdrawal and skill erosion.¹¹ Lawsuits against Character.AI in 2024 and OpenAI in 2025 allege chatbots worsened or even coached self-harm, underscoring real-world stakes.¹² ¹³

This is underscored by a very recent story in the NY Times about a 16 year old California boy who committed suicide. It was found that ChatGPT advised him to seek help, but ultimately facilitated his suicide by answering questions related to him hanging himself. OpenAI is now making changes to how the models respond to similar situations and questions, attempting to safe-guard against future similar situations.

Subscribe now

Why accountability gets scrambled in human–AI work

Two patterns matter for business:

“Automation bias—the tendency to over-rely on automated recommendations—has emerged as a critical challenge in human–AI collaboration.”¹⁴

Reviews across healthcare, aviation, and public administration show people over-trust suggestions, miss model errors, and under-report their own.¹⁵ Then, when something breaks, we fall into moral crumple zones:

“The moral crumple zone protects the integrity of the technological system, at the expense of the nearest human operator.”¹⁶

In one incident, a clinician blames “the algorithm.” In another, compliance blames the clinician for “not verifying.” Either way, accountability is obscured—great for avoiding heat, terrible for learning and trust.

But closer to home, school, and our offices, we should learn lessons from the 16 year old mentioned above. The algorithm didn’t make him do it, and it’s fundamentally not ChatGPT’s fault, but is there a responsibility that organizations have to curtail how these bots reply to questions or is the responsibility on the person who performed the act? If we’re going to seek answers outside of the victim, why aren’t we pointing the spotlight to the people who were in the boy’s life as real actors who could have intervened.

These are tough and controversial questions. I only mention them to provoke thought and dialog as a society so we can zoom out on the larger implications of these tools in young people’s (and adult’s) lives.

The answer to the above questions are: it’s not that simple. There isn’t anyone to blame and although adding guardrails to the chat bots will help in some circumstances, we as a society have to figure out how we empower each other more to prevent these activities from happening. We fundamentally won’t be able to code for every circumstance and situation that people will engage with chat bots over…there has to be a human-centric change.

The executive angle: how this shows up in our office buildings

Knowledge workers are now decision editors over model outputs—drafting emails, pricing endorsements, segmenting customers, flagging fraud. That creates three failure modes:

1. Responsibility gaps: RACI charts say “the team” owns it; logs say “the model” suggested it. No single throat to choke.

2. Invisible drift: Models are updated silently; prompts mutate; responses change over time as new facts are introduced…how do we keep up from a governance perspective (HINT: this goes back to having a trusted data environment to start from).

3. Metric theater: Dashboards praise cycle-time gains while error costs hide downstream in rework, complaints, or regulatory findings.

   ---

   Subscribe now

   ---

Pragmatic fixes to investigate

1. Mandate human sign-off where it matters. Require named ownership for consequential actions (pricing overrides, denials, escalations). Make the signer visible in the record. Tie incentives to accuracy after review, not just speed.¹⁴ ¹⁵

2. Instrument the last mile. Log prompts, versions, and human edits. Store the decision packet (input → model output → rationale → final action) for audit and retro. This shrinks post-mortem guesswork.¹⁴

3. Design for friction at decision boundaries. Insert verification steps when confidence, novelty, or impact exceed thresholds (e.g., new policy, out-of-distribution inputs). Don’t add friction everywhere; add it where regret is expensive.¹⁵

4. Kill blame diffusion with RACI+AI. Extend RACI to RACII (Responsible, Accountable, Consulted, Informed, Interpreter). The Interpreter is the human who attests they understood the model output and constraints.

5. Teach parasocial and automation literacy. Roll short, mandatory modules: anthropomorphism, automation bias, calibration, escalation etiquette. Not vibes—case-based drills with failure examples.⁷ ¹⁴

6. Set red lines for companionship features. Create a system where whether you’re a parent, teacher, friend, sibling, or other adult in a child’s life, there are clear changes in behavior and other prompts that you can use to engage in dialog with those who seem to be lost or in trouble. Let’s not leave it up to a virtual companion. We need training as a society that could help prevent the loss of life in a way that’s much more effective than child-to-bot friendships.

Bottom line: Accountability didn’t die; we let it atomize. Reassemble it with named owners, logged decisions, calibrated friction, and teams trained to question polished answers. That’s how you get speed and responsibility in the AI era.

Share

References

1. Twenge, J. M., Zhang, L., & Im, C. “It’s beyond my control: A cross-temporal meta-analysis of increasing externality in locus of control, 1960–2002.” https://pubmed.ncbi.nlm.nih.gov/15454351/ PubMed

2. SAGE Encyclopedia of Social Psychology. “Self-Serving Bias.” https://sk.sagepub.com/ency/edvol/download/socialpsychology/chpt/selfserving-bias.pdf SAGE Journals

3. Campbell, B., & Manning, J. “The Rise of Victimhood Culture.” https://www.researchgate.net/publication/323181753_The_Rise_of_Victimhood_Culture ResearchGate

4. Perry, N. B., et al. “Childhood self-regulation as a mechanism…” (PMC). https://pmc.ncbi.nlm.nih.gov/articles/PMC6062452/ PubMed Central

5. ACT Research. “Grade Inflation a Systemic Problem in U.S. High Schools.” https://leadershipblog.act.org/2022/05/grade-inflation-past-decade.html ACT

6. Galanter, M. “The Vanishing Trial: An Examination of Trials and Related Matters in Federal and State Courts.” https://api.law.wisc.edu/repository-pdf/uwlaw-library-repository-omekav3/original/0b9f361000c04494e8cff30f04b3afeb486193d4.pdf Wisconsin Law API

7. Nielsen Norman Group. “The ELIZA Effect: Why We Love AI.” https://www.nngroup.com/articles/eliza-effect-ai/ Nielsen Norman Group

8. Business of Apps. “character.ai: revenue and usage statistics (2025).” https://www.businessofapps.com/data/character-ai-statistics/ Business of Apps

9. Institute for Family Studies/YouGov. “Artificial Intelligence and Relationships.” https://ifstudies.org/blog/artificial-intelligence-and-relationships-1-in-4-young-adults-believe-ai-partners-could-replace-real-life-romance Institute for Family Studies

10. VICE. “‘It’s Hurting Like Hell’: AI Companion Users Are In Crisis.” https://www.vice.com/en/article/ai-companion-replika-erotic-roleplay-updates/ VICE

11. UConn Today. “Teenagers Turning to AI Companions…” https://today.uconn.edu/2025/02/teenagers-turning-to-ai-companions-are-redefining-love-as-easy-unconditional-and-always-there/ UConn Today

12. Associated Press. “AI chatbot pushed teen to kill himself, lawsuit alleges.” https://apnews.com/article/chatbot-ai-lawsuit-suicide-teen-artificial-intelligence-9d48adc572100822fdbc3c90d1456bd0 AP News

13. SFGATE. “California family sues Sam Altman, OpenAI over son’s suicide.” https://www.sfgate.com/tech/article/chatgpt-california-teenager-suicide-lawsuit-21016916.php SFGATE

14. Romeo, G., & Conti, D. “Exploring automation bias in human–AI collaboration.” AI & SOCIETY (2025). https://link.springer.com/article/10.1007/s00146-025-02422-7 SpringerLink

15. Vered, M. et al. “The effects of explanations on automation bias.” University of Melbourne (2023). https://psychologicalsciences.unimelb.edu.au/__data/assets/pdf_file/0019/5252131/2023Vered.pdf psychologicalsciences.unimelb.edu.au

16. Elish, M. C. “Moral Crumple Zones: Cautionary Tales in Human-Robot Interaction.” DOAJ record. https://doaj.org/article/97ff6743ea7a44a5ade2a04fd2c57a3c Directory of Open Access Journals