When to Say No to AI

We spend most of our time helping businesses use AI well, so this one may read as off brand. It is not. The fastest way to lose your team’s trust in a new tool is to point it at the process where being wrong is unforgivable. One bad call in the wrong place undoes a year of quiet wins.

Harrison likes to say that AI is the dumbest it will ever be today, and that is true. It is also not an argument for using it everywhere right now. He also says to treat AI like a junior staff member, and that framing arrives with limits built in. There are things you would never hand a brilliant new hire in their first month. Those are where you say no.

The Test: Who Answers for This?

Most no decisions come down to one question. When this goes wrong, who stands in front of the client, the regulator, or the judge and explains it?

The answer is never the software. A court made the point cleanly in Mata v. Avianca in June 2023, where attorneys submitted a brief containing what the court described as non existent judicial opinions with fake quotes and citations created by an AI tool. The opinion states that technological advances are commonplace and there is nothing inherently improper about using a reliable artificial intelligence tool for assistance, then makes clear that existing rules impose a gatekeeping role on attorneys to ensure the accuracy of their filings. The accountability never moved.

So the test is not whether AI can produce a plausible output. It almost always can. The test is whether a human will own the result and knows enough to have caught a bad answer. If nobody would have known, the process is not ready.

Five Places We Say No

These are the categories where we keep a person in the chair. Not reviewing a draft. Doing the thinking.

Final client communication during a crisis. Outage, breach, missed deadline, safety issue, a mistake that cost someone money. The tone matters more than the words, and the person receiving them can tell accountability from a generated apology. Use AI to organize your facts. Write the message yourself.
Hiring and firing decisions. Employment decisions carry legal exposure, and the appeal of automation is exactly the risk. A system that scores people at volume applies the same flawed judgment thousands of times before anyone notices. Europe’s AI Act treats certain employment uses as high risk because of what they do to livelihoods, and even if that law does not reach your business, the reasoning does. The decision is yours.
Anything where a specific person must be accountable. Signing off on financials, approving a payment, attesting to a compliance answer, certifying that work was completed. If your name goes on it under penalty of something, you need to have looked.
Regulated advice. Legal, medical, tax, financial, engineering, safety. These fields have licensing requirements for a reason, and the requirement follows the advice, not the tool that drafted it. AI can help a licensed professional work faster. It cannot stand in for one.
Judgment calls with no undo button. Deleting data, terminating a contract, publishing something permanent, moving money, taking a system offline. The common thread is not difficulty. It is that there is no cheap way back.

How to Tell the Difference

You do not need a policy binder. Run any process through four questions.

Is it reversible? If a bad output can be caught and fixed before it does harm, the risk is low and you should probably try it. If harm happens the moment the output is used, slow down.
Would a competent person catch a wrong answer? Review only works when the reviewer can spot the error. NIST warned in its 2024 generative AI guidance that these systems present erroneous content confidently and even supply logic or citations that appear to justify a wrong answer. If your reviewer cannot evaluate the substance, you do not have a review, you have a signature.
Does someone have to answer for it personally? A regulator, a court, a client contract, a professional license, or your reputation with a customer of nine years.
Does the relationship depend on it being human? A condolence, an apology, a promotion, a hard piece of honest feedback. Automating those does not save time, it spends trust.

The European approach is a useful checklist even far outside its reach. Article 14 of the EU AI Act requires that people overseeing high risk systems be able to correctly interpret the output, decide not to use the system or to disregard, override, or reverse its output, and interrupt it through a stop procedure. It also requires overseers to remain aware of the tendency to automatically rely or over rely on the output, the failure mode we see most in practice. If your process does not let a human do those things, that is your answer.

What Saying No Does Not Mean

A no on the decision is almost never a no on the whole process, and treating it that way throws away the value.

In a crisis, AI can assemble the timeline, pull contract clauses, and flag which clients are affected. You still write the letter. In hiring, it can build consistent interview questions and organize your notes. You still choose the person. In regulated advice, it can draft the explanation your licensed professional corrects and approves. In irreversible operations, it can prepare the change and write the rollback plan. A person still presses the button.

Teach it as a sentence: AI does the preparation, humans make the decision. It preserves most of the speed and all of the accountability. It also avoids the opposite failure, a blanket ban that pushes people toward tools you cannot see, which is how you end up with an AI shaped version of the problem we described in our article on shadow IT.

Write It Down So It Actually Sticks

A verbal understanding lasts until the first busy week. One page is plenty.

List the no zones by name. Use your actual processes and job titles, not abstract categories. People follow rules they can see themselves in.
Say what is encouraged. Name the places you want people using AI aggressively. A document of only prohibitions gets read as anti technology and ignored.
Name an owner for exceptions. Someone who can say yes to a case the list did not anticipate. Otherwise the list becomes a wall people climb over.
Revisit it on a schedule. Twice a year. Some of today’s no zones will move, and the document should move with them.

If you want a formal structure, NIST published its AI Risk Management Framework in January 2023 as a voluntary framework organized around four functions: govern, map, measure, and manage. You do not need to adopt it wholesale, but the sequence is sound. Decide who is accountable, know where AI is in use, have a way to tell whether it is working, and act on what you find.

The Bottom Line

Saying no in the right places is what makes your yeses believable. When your team knows there are lines that do not move, they stop worrying that every new tool is a slow replacement for their judgment, and they start using AI where it is genuinely good. Draw the lines deliberately, revisit them as the technology improves, and be generous everywhere else. That is the same balanced approach we take in preparing teams for AI without the hype.

If you want help deciding where AI belongs in your business and where it does not, we will give you a straight answer rather than a sales pitch. We work with small and mid sized businesses across Denton County and North Texas. Contact us today.

Sources:

AI Policy Artificial Intelligence IT Strategy Small Business IT

Comments are closed