Hold Your Horses! AI Agents Can’t Freestyle Your Business
Why deterministic process docs are the real brains behind every AI agent you're building.
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.
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.
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!
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:
Pick a process to overhaul for AI agents to perform
Evaluate each step for effectiveness
Determine in a world without humans how the steps could be performed
Modify the process flow and SOP to account for AI agents using APIs, MPC or other computer controls
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.
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.