Where Does AI Fit? Put It Where Mistakes Are Cheap

"Where does AI fit in our business?" is the question I hear most as an AI advisor, and almost everyone answers it the same way: point AI at the most valuable thing we do. It sounds like strategy. It is usually how AI initiatives die.

After 35+ years building software and security systems, and now working as a fractional CTO across my own companies, I use a different filter, and it has saved more AI projects than any model upgrade ever has. Do not ask where the payoff is biggest. Ask where a mistake is cheap.

Every task has a price tag for being wrong

Here is the uncomfortable truth about the current generation of AI: it will be wrong some percentage of the time, and nobody can tell you exactly when. That is not a reason to avoid it. It is a design constraint, the same way latency and downtime are design constraints. You do not eliminate it. You decide where it lands.

Every process in your business has a price for a wrong answer. A bad first draft of an email costs the thirty seconds it takes someone to fix it. A bad number in a customer proposal costs the deal, and maybe the relationship. Those two tasks might look equally automatable in a demo. They are not remotely the same decision.

So the real return on an AI placement is not the payoff. It is the payoff minus the cost of wrong answers, times how often they happen. Most spreadsheets I review as an AI consultant only contain the first half of that equation. That is why so many projects that looked brilliant in the planning meeting get quietly shut off three months after launch: one expensive mistake in front of the wrong person, and leadership loses its nerve.

Cheap mistakes, huge volume: that is the sweet spot

Run the full equation and the answer to "where does AI fit" changes shape. The glamorous, high-stakes placements fall to the bottom of the list. The boring ones rise to the top: first drafts, triage, summaries, classification, internal lookups, data cleanup. Work where a wrong answer costs seconds to catch, gets caught by a human who was already going to touch it, and happens thousands of times a year.

That volume is the point. A task that saves ninety seconds and runs two hundred times a day is worth far more than a task that saves an hour and runs twice a month, and it carries a fraction of the risk. Inside my own companies this is the standing rule: AI writes the first pass, a human owns anything a customer will feel. Not because the models are bad, but because that is where the math actually works.

You can also engineer the price of wrong downward

The second half of the job, and the part where experience earns its keep, is that the cost of a mistake is not fixed. You can design it down.

Years ago I built AI for a sports video company called Egoli, using facial recognition to find marketable moments in football matches. Run against a database of every Premier League player in history, it produced garbage. At one point it confidently tagged a player who had been dead for twenty years. The fix was not a better model or more training data. We constrained the problem to the twenty-two players actually on the pitch, and accuracy went through the roof. Same AI, same footage. We just shrank the space it could be wrong in.

That is what good AI system design looks like: narrowing inputs, adding review steps where stakes jump, routing low-confidence outputs to a human. Where the price of being wrong starts out high, you can often engineer it down until the placement makes sense. But you have to know the number first.

Two questions before you automate anything

So before you green-light your next AI project, skip the technology evaluation for a moment and put two numbers next to the task. How often will the output be wrong? What does one wrong output cost you? If you cannot answer the second question, you are not ready to automate that task, no matter how good the demo looks.

Boring tasks with cheap mistakes will come out of that exercise looking like gold. High-stakes tasks with expensive mistakes will come out looking like what they are: places where AI might belong eventually, behind guardrails, and never first.

The companies winning with AI right now are not the ones with the best models. Everyone rents the same models. They are the ones who put errors where errors cost almost nothing, and kept human judgment where it is the most valuable thing in the building.

If you want help mapping where mistakes are cheap in your business, and where AI actually fits, I take a few clients at a time. Book your free 15-min AI Fit Call →