The verification habit: checking AI work before it ships
Fluent output is not checked output. Five fast verification moves that catch most AI mistakes before a customer, a client, or a regulator does.
AI output fails in a specific, predictable way: it is wrong while sounding right. That single property should shape how you check it. You are not proofreading for typos. You are auditing a confident draft from an assistant who never says “I am not sure.”
The good news is that verification is a habit, not a project. Five moves cover most of the ground, and most of them take under a minute.
Move 1: sort by the cost of being wrong
Before checking anything, ask what happens if this is wrong. A brainstorm list can ship unread. A customer email deserves a read-through. A price quote, a legal-ish claim, or anything with numbers that move money gets the full treatment. The catalog in what AI still gets wrong is effectively a map of where the full treatment applies.
Most people check everything equally, which means checking everything badly. Budget your attention where the failure hurts.
Move 2: trace the claims, not the prose
Read the draft once and underline every checkable claim: every number, date, name, citation, and “studies show.” Ignore the connective prose around them. The prose is almost always fine; the claims are where the failures live.
Then trace each underline to a source you trust. Not “does this sound plausible” but “where did this come from.” A claim with no findable source gets cut or rewritten as opinion. This is exactly the discipline this site applies to its own briefings, and it transfers to any business document.
Move 3: recompute every number
Language models predict text, including text that looks like arithmetic. They are much better at this than they were, and still not a calculator. Any number that survives into a final document, a total, a percentage, a date difference, gets recomputed by you or by an actual spreadsheet.
The trap to respect: a model can produce a correct method with a wrong result, or a wrong method with a confident explanation. Checking the explanation is not checking the number.
Move 4: test instead of read where you can
For anything executable, testing beats inspection. Code gets run against cases you chose. A formula gets tried on a row where you know the answer. A process document gets walked through once, literally, step by step. Ten minutes of testing finds what an hour of reading misses, because reading shares the model’s blind spot: things that look right.
Move 5: the second-opinion pass
For high-stakes output, ask a model, in a fresh conversation, to find what is wrong with the draft: missing cases, unsupported claims, arithmetic to re-check. A fresh conversation matters because the original one is anchored to its own answers and will mostly defend them.
Treat the critique as leads, not verdicts. The point is a cheap list of places to look, produced by something that does not get tired or polite.
Making it stick
A habit needs a trigger. Tie verification to the moment of sending: nothing AI-assisted leaves the building without its underlines traced and its numbers recomputed. Put it in your checklist, your template, or your team’s definition of done.
And keep the workload honest. If checking a task’s output regularly takes longer than doing the task yourself, the AI is not earning its place in that workflow. That arithmetic, the end-to-end time including the checking, is the honest metric from how to choose AI tools for a small business, and it is the one to live by.
The habit pays compound interest. Every error you catch at your desk is one that did not reach a customer with your name on it.