The Most Overlooked AI Skill Is Not Prompting. It Is Verification.

The document looked perfect. That was exactly the problem.

A fluent summary, a confident tone, plausible sources. And in the middle, a regulatory obligation that does not exist. This scene plays out in companies every week. Not because people are careless, but because nobody taught them what checking an AI output actually involves.

The skill nobody trains

For two years, companies have invested in prompting training. That is useful, and it remains necessary. But while teams were learning to write better requests, the models kept improving. Outputs became more fluent, better structured, more convincing. And that is where the risk changed in nature.

A crude error catches your eye on its own. A plausible error, embedded in impeccable prose, sails through. The better the model writes, the more fluency looks like reliability. The brain reads clean text and concludes the substance is solid. The two have nothing to do with each other.

The capacity to produce with AI has exploded. The capacity to evaluate what AI produces has not kept pace. That gap is what creates incidents, not the technology.

The wrong conclusion

When an error slips through, the usual reaction follows the same pattern: the tool is unreliable, usage should be restricted, or we should wait for a better model.

That is a diagnostic error. A human colleague also makes mistakes in deliverables. Nobody concludes that colleagues should be banned. We built review, validation and sign off processes instead. So the right question is not: how reliable is the model. The right question is: what does review look like when the first draft comes from an AI.

Verification is a skill. It can be taught, practiced and structured. And today it is the weakest link in most professional uses of AI.

What verifying actually means

Four principles are enough to transform a team's practice.

1. Match the effort to the stakes. Not everything deserves the same level of scrutiny. An internal brainstorm tolerates approximation. A client deliverable requires a factual review. Regulatory or contractual content demands verification at the source, point by point. Defining these three levels, and building the reflex of asking which level you are at, already changes a great deal.
2. Verify claims, not vibes. Rereading an AI text while asking yourself whether it seems right verifies nothing. The useful method is to separate what is factual from what is reasoning. Numbers, names, dates, obligations, citations: each of these is a claim that can be true or false. Reasoning and framing are debatable, and they are checked in a different way.
3. Make verification visible in the workflow. If everyone assumes someone else checked, nobody checks. A deliverable produced with AI should have a named reviewer, like any important document. The initial prompt can help too: asking the model to cite sources or flag its own uncertainty guarantees nothing, but it gives the reviewer something to hold on to.
4. Train it like any other skill. In our training sessions, the exercise that stays with participants is never the one where the AI shines. It is the one where they have to fix a flawed output, find the error, and argue their case. This is precisely the Knowledge dimension of the MAKIA framework: a team's capability is not measured by what it produces with AI, but by the quality of its judgment about what AI produces.

Trust is built somewhere else

Trust in AI inside an organisation will not come from a better model. It will come from people who know when to check, what to check, and how to do it fast. A team that masters this can use AI on ambitious work without taking reckless risks. A team that does not will stay stuck between two bad options: distrust that blocks everything, or blind confidence that eventually gets expensive.

Try this week

Take one recent deliverable your team actually produced with AI: a summary, an important email, an analysis. Open a document and set a 20 minute timer. Highlight every factual claim: numbers, names, dates, obligations, references. Mark each one V (verified), P (plausible but unchecked) or U (unknown). Then check at the source the three claims that would do the most damage if they were wrong. When the timer ends, count the P marks. That number is your current risk exposure.