31 July 2026 · 3 min read

Where AI earns its place

The useful question is not whether AI works. It is which jobs it is actually good at, and which ones it only appears to be good at until someone checks.

The argument about whether AI is useful has mostly finished, and it finished in a boring place: it depends what you point it at.

That is not a hedge. The gap between where it earns its place and where it merely looks impressive is wide, predictable, and worth learning, because the failures are quiet ones.

The shape of the good cases

The jobs it does well have something in common. The work is genuinely hard to start, easy to check, and cheap to be wrong about.

Drafting is the clearest example. A first version of almost anything, a reply, a spec, a policy, a summary, is expensive to produce and quick to judge. You know within seconds whether the draft is close. If it is wrong you delete it and you have lost almost nothing.

The same shape appears in reading rather than writing. Going through a long document to find the parts that matter, comparing versions, pulling structure out of something unstructured. You can verify the answer against the source, which is what makes it safe.

• Work where starting is the expensive part
• Work where a wrong answer is obvious to a person who knows the subject
• Work that is repetitive enough to be tedious but varied enough to resist a simple rule

Anything sitting in all three is where it earns its place, and the gain is not marginal.

The shape of the bad cases

The failures share a shape too, and it is close enough to the good one to catch people out.

The dangerous case is work that is hard to start and hard to check. Anything where being subtly wrong looks the same as being right. Figures nobody will recalculate. Claims about what a regulation requires. Summaries of documents the reader will never open.

That is not a technology problem. It is a process problem. The output is plausible, plausible is what gets approved, and the error surfaces much later attached to somebody's name.

The second failure is quieter. Using it for work that did not need doing. It is very easy to generate a great deal of competent material nobody asked for, and mistake the volume for progress.

The judgement it cannot supply

There is a category worth naming separately: decisions that are genuinely yours.

Who to hire. Whether to take the client. When to tell someone something they will not want to hear. What the business is actually for.

It can help you think. It can argue the other side, list what you have not considered, or say plainly that a plan rests on an assumption you have not tested. That is real and it is useful. But the decision carries consequences you live with and it does not, and that asymmetry does not disappear because the reasoning sounded good.

The failure mode is not that it gives bad advice. It is that outsourcing the thinking slowly makes you worse at it, and you do not notice until a decision arrives that nothing can be delegated on.

How we actually use it

The rule we work to is simple enough to say in one line: use it where a person still checks the output, and be honest about whether that person really will.

That second clause does most of the work. Everyone intends to check. The question is whether the process makes checking natural or makes it an extra step somebody skips at the end of a long day. If it is the second, the safeguard is decorative, and you should assume the unchecked version is what goes out.

Used that way it is genuinely one of the better tools to arrive in years. Used as a way to produce more things faster without anyone reading them, it is a way to be wrong at scale.

The technology is the same in both cases. The difference is entirely in what you asked it to do.