When Not to Use AI: A Practical Checklist for Operators.

A tool has a shape. You learn it by where it fails, not where it shines, and AI shines so brightly in the demo that nobody films the part where it misses. Knowing when not to reach for it is the whole of using it well. We run a checklist with clients before a single prompt gets written, ten shapes the work can take where AI is the wrong thing in your hand.

A wrong answer is expensive and invisible.

Medical advice, a legal reading, a security call, money moving out the door. AI hands you something fluent and sure of itself, because fluent and sure is what it is built to be. The danger is not that it is wrong. It is that the person on the receiving end cannot tell, and neither can you. Picture it missing this one, and ask who notices before the damage is done.

The answer is already a rule.

If you can write the logic in a spreadsheet, write it in a spreadsheet. A rule that runs the same way a thousand times is a gift. Do not trade it for a system that improvises, then spend your weeks coaxing the improviser back toward the rule you already had. It happens constantly, almost always because somebody decided the project needed AI somewhere and went looking for a place to put it.

Nobody has time to read the output.

Bolt AI onto a team that is already underwater and you have not bought them time. The work does not vanish, it moves, from doing to checking, and checking a machine that is confident and occasionally wrong is its own full day. When a team is underwater, more output is not relief. It is more water.

Nothing tells you it is working.

If you cannot say how you would know the AI is doing well, you will not catch the day it quietly stops. Drift is patient. The numbers slip a little, then a little more, and one morning something breaks and the trail is cold. Every deployment owes a plain answer to one question: how do we know this is working? “We don't” is a reason to wait.

The work is rare and it matters.

AI earns its keep on the thing you do a hundred times a week, where the pattern is thick and the stakes per instance are thin. Flip both and the logic inverts. A decision that lands twice a year and carries real weight does not want a workflow. It wants a person who slows down and weighs it.

Someone is paying for a human.

The premium service, the briefing in the room, the account where the whole product is that a real person spent real hours on your problem. Slip AI scaffolding under that and you have not made it cheaper to deliver, you have made it something else, and the buyer feels the difference even when the sentences look right.

The data is a mess nobody will admit.

Most AI failures are not model failures. They are data failures wearing a model's coat. Feed it inputs that are inconsistent, half-empty, or quietly fought over by two departments, and it broadcasts that mess back at you, louder and surer. The fantasy is that the project will force you to finally clean the data. It will not. Clean it first, or do not start.

The call is about a person.

Hiring, performance, who is about to churn, which ticket gets escalated. The legal and ethical and reputational exposure is real, and these systems pick up our worst habits in ways that are hard to see from the outside. Use AI as an input here, never as the judge. Let it summarize and lay out the options. Do not let it rank human beings and hand you the verdict to sign.

The upkeep costs more than the work.

An AI workflow is not free because no one is invoicing for it. Somebody still updates the prompts, watches the outputs, refreshes the data, patches the edge case that surfaces at month four. If it saves less time each week than it takes to keep standing, you have not built a tool. You have adopted a chore.

Everyone else is doing it.

The worst reason of the ten, and the most common. When the honest answer to “why are we doing this” is “because we should be doing AI things by now,” the project is theater, and theater fails on schedule and embarrasses everyone in the cast.

Saying no in the wrong places is what makes the yes worth anything. A practice that recommends AI for every workflow is not advising you, it is selling you. The most useful thing we do is point at the place AI does not belong and say so, and it is what clients remember a year on, when the team is intact and the budget did not vanish into a stalled build. The tool is fine. It has a shape, like every tool before it, and most of your work was cut to a different one. Learning where it does not fit is not the timid version of this work. It is the craft.