When Should You Not Put AI in a Process? Run the Exclusion Test

Updated: 4d
The short answer
No process should get AI until someone has written down what it would be allowed to decide. Census Bureau data put AI use at 19.8% of U.S. businesses as of May 3, 2026, and under 20% among firms with four or fewer employees. The gap is not access. It is an unwritten decision.
Why does nobody ever write down an AI "no"?
Because nothing in the system asks you for one. Every vendor notice, trial offer and default-on announcement pushes in a single direction, and none of them requires a written answer. A decision not to automate leaves behind no invoice, no login and no confirmation message, and a no with no trace is indistinguishable from never having considered the question.
That matters because the next notice overwrites a silence. Pew Research Center found that 70% of Americans who have heard of AI have little or no trust in companies to make responsible decisions about how they use it in their products. Your clients are inside that number. A no that nobody wrote is not a no. It is a blank the vendor fills in for you.
It is easy to spend a year receiving a few dozen notices a month that each switch something on by default, and to reach December having written down a decision about none of them.
What does the adoption data actually show about smaller firms?
It shows a size gap, not an interest gap. The U.S. Census Bureau reported AI use hovering between 17% and 20% of businesses from December 2025 through May 2026, with the increase concentrated in firms of at least 20 employees: 37% among firms with 250 or more, and no significant change among firms below 20. The sector spread is wider still, at 39.7% in Information, 33.9% in Finance and Insurance, and roughly 14% in Retail Trade.
The more useful finding is what the data does not contain. That same Census supplement measures AI use across fifteen business functions, including finance, human resources, customer service and marketing, and publishes no usage rate for any of them. The SBA Office of Advocacy counts 36,207,130 small businesses, 99.9% of all U.S. firms, employing 45.9% of the private-sector workforce. Everyone is counting adoption. Nobody is counting the processes deliberately left alone, which is the number that says whether owners are choosing or drifting. We made the drift case in the adoption gap is a decision gap.
Which four conditions disqualify a process? The Exclusion Test
Run these against one process before you automate any part of it. A process has to clear all four. Tripping one is not a verdict against AI, it is a diagnosis, and the trip names what to fix first.
The input nobody has verified. The process runs on a number, a list or a feed whose accuracy nobody has confirmed in writing. AI does not improve a bad input. It applies it faster, in more places, and with more confidence. If the measurement feeding a process was last checked by whoever installed it, that process is disqualified until someone checks the instrument.
The output nobody will act on. Name the person who reads the result and the decision it changes. If you cannot name both, automation will produce more of something nobody uses, faster. This is the test we applied to recurring reports in the Reader Test: a process whose output changes no decision does not need to be cheaper, it needs to be retired.
The judgment nobody has written down. Most processes carry a rule that lives only in the owner's head: who approves an exception, what number is too high. A tool asked to run an unwritten rule will invent one, and that invention quietly becomes your policy. Pew found 67% of Americans say they understand little or nothing about what companies do with their personal data, up from 59%. Your clients will not ask what rule your automation follows. They will decide what they think it is.
The failure nobody owns. When the output is wrong, who notices, how fast, and what do they do? A process with no named owner for its errors does not need AI, it needs an owner. Pew found 73% of Americans feel they have little or no control over the data companies collect about them, which is what being inside somebody else's unowned process feels like.
The gate takes discipline: until a process clears all four conditions, the answer is "not yet," written down with a reason and a date to revisit. Three of four is not a pass. It is a specific repair with a specific owner.
What do you do with a process that trips a condition?
You fix the condition, which is almost always cheaper than the automation would have been. An unverified input is one afternoon of checking the instrument against a hand count. An unread output is one conversation that ends in retiring a report. An unwritten judgment is three sentences in a document: who approves, what the threshold is, what the exception is. An unowned failure is a name and a response window.
What to fix before you buy a CRM makes the same argument about systems of record, and it holds here: a tool installed on top of an undefined process inherits the definition problem and adds a vendor to it. The technology decisions that actually get made are the ones where somebody outside the business was holding the door, naming one artifact and declining to proceed without it. The Exclusion Test is how an owner holds that door for themselves.
Why does the vendor's default become your policy?
Because the vendor is the only party in the transaction with a deadline. Notices arrive with cutover dates, terms updates, feature flags and expiring credits, each one a small, friction-free yes. The corresponding no requires you to stop, write a sentence and file it, with nobody waiting on it.
That asymmetry is expensive in a growing market. The Economic and Business Research Center at the University of Arizona's Eller College projects Phoenix-area employment growth of 1.0% in 2026 and 1.1% in 2027, with personal income growth of 5.7% in 2026. More activity means more processes, and more processes mean more defaults you did not choose. Evaluating the vendor is a separate exercise, covered in how to evaluate a software vendor when you are not technical. The Exclusion Test runs first, on your own process, before any vendor is in the room.
Frequently asked questions
Is the Exclusion Test an argument against using AI? No. It is a filter that sends AI to the processes where it will hold, and keeps it off the ones where it would scale an error. Most owners who run it end up automating something, just not the process they started with.
How long does the Exclusion Test take on one process? About twenty minutes per process, and the writing is the work. The four answers fit on one page, and that page is the artifact: a verdict you did not write is one the next vendor notice can overwrite.
What if a process trips a condition but the feature is already switched on? Then you have a live policy nobody chose, and the first move is to write down what it currently does. You can decide whether to keep it after that; you cannot decide anything about a rule you have not stated.
Who should own the Exclusion Test in a small business? Whoever owns the process, with the owner signing the verdict. The signature is what converts a default into a decision, and gives you something to revisit on a date instead of rediscovering it after a client asks.
Where CINCO fits
We run this with owners of established Phoenix businesses being sold automation faster than they can evaluate it: name the process, test the four conditions, write the verdict, fix what trips. See Digital for the systems side, Growth for the strategy side, or get started with the one process you suspect runs on an unchecked input.
Sources: U.S. Census Bureau, Large Firms With at Least 20 Employees Biggest AI Users, May 26, 2026. SBA Office of Advocacy, Frequently Asked Questions About Small Businesses, 2026, February 3, 2026. Pew Research Center, How Americans View Data Privacy, October 18, 2023. Economic and Business Research Center, Eller College of Management, University of Arizona, Phoenix MSA Third Quarter 2026 Forecast Flash, August 7, 2026.
Comments