Intelligence is not the same thing as intellect

The useful question about AI is not where you could use it. It is where intelligence is actually the bottleneck. Plenty of stalled pilots were never held back by an underperforming model: the real constraint was a broken handover, an integration nobody owned, or a decision no one wanted to sign off. Putting a model on top of any of those does not fix them. It makes the mess run faster.

But that framing has a failure mode, and it is worth addressing directly, because the word "intelligence" is doing a great deal of work in it.

When people hear it, they think of strategy and analysis. They picture Einstein. Then they scan their operations, find nothing that looks like a physics problem, and conclude the argument does not apply to them. It usually does. They were looking for the wrong shape.

The test isn't difficulty

Intelligence, in this context, is not brilliance. It is perception, attention, and consistent judgement applied to messy input.

Anyone on the loading bay can tell you the delivery note doesn't match what's on the pallet. Almost nobody can write down how they know. That is not a gap in their ability; it is the nature of the knowledge. It is tacit, and it resists the interview that would be needed to turn it into a rule.

The pattern is everywhere once you stop looking for cleverness:

• Does this photo show the damage being claimed?
• Which six of these forty inbound emails are different?
• Is this the right document, and is it the current version?
• Does the site photo show the work that was signed off?
• Is this invoice line reasonable for this job, or does someone need to look at it?

None of that is intellectually demanding. All of it resists a rules engine. And all of it stalls when the one person who is good at it is on leave.

The tell is a thin procedure

The signature is a thin standard operating procedure. Not thin from neglect — thin because it has been thickened before, and the rules did not survive contact with the work.

That is as true of perception as it is of judgement. The knowledge is not unwritable. It is unavailable to the interview. Polanyi's examples were perceptual from the start: recognising a face, reading an X-ray, hearing that an engine is wrong. Nobody thinks the radiologist's judgement is undocumented because someone was lazy.

The second version: the human is the throughput limit

There is a variant where nobody is struggling at all.

One person doing this check is fine. One person doing it four hundred times before lunch is the bottleneck — and by item three hundred they are worse at it than they were at item ten. Attention is a consumable.

That decay is asserted from experience, not measured here. Treat the shape as illustrative rather than as a figure you could put in a business case.

Tedium is a capacity problem wearing the costume of a skill problem. That distinction matters commercially, because tedious work never makes it onto the transformation roadmap. It is too small, too boring, too obviously "just admin" to justify a slide. It is also, very often, exactly where the queue forms and where the errors that cost real money get made.

Widening the definition doesn't lower the bar

This is where the argument needs care, because it is the argument every vendor would love to be made on their behalf: look, the list of AI opportunities is bigger than you thought.

It is bigger. But the qualifying questions still apply, and two of them get harder once you move from judgement into perception.

Is the data available?

For a judgement task, that usually means records you already keep. For a perception task it means labelled examples of the cases that go wrong — and those are rare by definition. Rare is why they are valuable. Rare is also why nobody kept them.

Will the people doing the work trust it?

Someone has to act on the output. If the checker still has to check everything, you have added a step and removed nothing. That is not a model problem, and no amount of accuracy fixes it on its own.

So the qualified list is longer than most boards think and shorter than most vendors claim. Both are true at once, and the gap between them is where the money goes missing.

Where to look

Not at the impressive work. At the work where a capable person does something in two seconds, several hundred times a day, and cannot explain how.

That is where intelligence is genuinely the constraint — not because the task is difficult, but because it has never been possible to hand it to a machine, and never been worth hiring six more people to do it.

And when a capability does not pass the qualifying questions, that is not a disappointing answer. That is a seven-figure initiative you didn't start.