How do you tell if your team is actually good at using AI?

Not by counting logins. The behaviour that separates people who get useful work out of AI from people who do not is whether they keep going after the first answer: reading it properly, saying what is wrong with it, and asking again. That is a habit, not a talent, which means it can be taught. The number to watch is not how much your team uses AI. It is what they send back.

Why seat counts and prompt volume tell you nothing

Most owners measure AI adoption the way they measure a phone plan. How many licences are we paying for, how many people logged in this month, how many prompts went through.

It is the easiest thing to count and it is close to useless, because a person can use AI constantly and use it badly. Volume tells you money is being spent. It says nothing about whether the output is any good, whether anyone read it, or whether it went out to a client unchecked.

Worse, a busy dashboard is reassuring. It looks like adoption. It can sit there looking healthy for a year while the actual quality of the work never moves.

What the research says separates good AI users from the rest

In February 2026 Anthropic published what it called the AI Fluency Index, built on 9,830 anonymised Claude conversations from a seven-day window in January 2026. The question it set out to answer was not whether AI is capable. It was whether people are using it well.

The headline finding is unglamorous, which is why it is useful. Iteration is the strongest signal. Roughly 86 per cent of conversations showed some gradual refinement, and the people who refined their requests questioned the model's reasoning about 5.6 times more often, and spotted missing context about four times more often, than people who accepted the first answer.

Read that again as a hiring or training statement rather than a technology one. The difference is persistence and specificity. Not cleverness, not technical background, not being young. Being willing to say that is not quite right, here is what is wrong with it, try again.

That is a coachable habit. Anyone in your business can learn it this week.

The finding that should change what you watch

The second finding is the one I would actually act on.

When the output looked finished, people checked it less. In conversations that produced polished artefacts, like a document or a small working app, fact-checking dropped by about 3.7 percentage points and questioning of the argument dropped by about 3.1 percentage points.

Those are small numbers describing a large problem. Polish is doing the work that evidence should be doing. The nicely formatted answer with confident headings reads as more trustworthy than the scrappy one, and it is not.

So the risk in your business is not the obviously rough draft, because nobody trusts a rough draft. It is the tidy one. The quote that arrives properly formatted, the client email that reads well, the summary with neat bullet points. Those go out the door without a second pass precisely because they look like they do not need one.

What to measure instead

Here is the framing we use at Handiwork, and it fits on one line.

Measure what came back, not what went out.

An AI request that produced something and was accepted immediately is a request nobody supervised. An AI request that went three rounds is a request somebody was actually responsible for. The second one is what good looks like, even though it takes longer and looks less impressive on a usage report.

In practice that gives you three things worth watching, none of which need software:

One. Does anyone push back on the output? In your next team meeting, do not ask what people used AI for. Ask what it got wrong, and what they did about it. The people with a story are fine. The people with nothing to report are either not using it, or not reading it, and you want to know which.

Two. Is the finished-looking work getting checked? Pick the one AI-assisted thing that leaves your business most often, usually a quote or a client email, and look at whether anyone reads it before it goes. If the answer is that it looks fine so it goes, you have found the gap.

Three. Can the good habits move? If one person is clearly better at this, the question is whether anyone else could describe what they do. If nobody can, it is not yet a capability. It is one person.

Why this is a training problem, not a tooling problem

Almost every owner I talk to can name the person in their business who is good with AI. Almost none can say what that person does differently.

While it stays a fact about a person, there is nothing you can do with it. You cannot hire for it, coach it, or spread it. It just sits there as an observation, and the business stays dependent on one person's instinct.

Naming the behaviours changes that. Once you know the thing to practise is to read the output properly and be specific about what is wrong, it becomes a habit you can ask for, model, and reward. It costs nothing and it does not need a new tool.

That is also why we start a first session by looking at how people are working rather than what they are working with. Most of what holds a small business back is not the software, it is that nobody has ever said out loud what good use looks like. The same reason your team is probably already using AI without telling you, and the same reason setting the tool up properly matters more than prompting it harder.

If you want the fuller picture of how a small business moves from curiosity to something that sticks, that is the whole adoption ladder. And if you want the commercial version of this question, it is what number your AI should actually be moving, which is a different question from how much it is being used.

You can see how we work through this with a business if you want the practical shape of it.

Ready to find out where you stand?

If you want an outside read on how well your team is working with AI rather than how much, our free AI Readiness Check takes about five minutes: AI Readiness Check.