People and culture: the AI nobody uses

What this layer actually measures

It measures sustained use, not training coverage.

The metric the industry reports — employees trained, licences deployed, sessions delivered — correlates with nothing. The one that matters is how many people are still using the tool eight weeks later without being reminded, and what they actually use it for.

What it does not measure: enthusiasm. Initial enthusiasm is deceptively high and decays with almost clockwork regularity.

Why mid-sized companies get it wrong

Because they confuse permission with adoption.

The pattern goes like this: a tool is bought, announced, a ninety-minute training session is held, it is opened to the whole workforce, and everyone waits. Eight weeks later a handful of people use it, almost always the ones who would have found a way on their own. The conclusion drawn — "people resist change" — is comfortable and almost always wrong.

What is happening is simpler: the work did not change. A tool was added on top of an identical process, with the same deadlines, the same deliverables and the same quality criteria. Using it is extra work until someone folds it into the flow, and nobody has time for extra work.

There is a second, subtler mistake: training everyone equally. General AI literacy has cultural value, but it does not produce use cases. Use cases come from a small group of people who perform a specific process and whose process gets redesigned.

The unit of work is the task, not the job

Daniel Susskind argues that current automation does not eliminate whole professions but substitutes tasks within them, and that the challenge is not mass unemployment but the transformation of professional profiles.

That reading is useful here because it changes the unit of work: you do not "deploy AI in the department", you redistribute specific tasks within a specific job. And it has a practical consequence: if nobody has looked at the job task by task, no redesign is possible, only superimposition.

We develop this in the piece on his thesis.

Diagnosis: where you are

Stage 1 · Tool available. There are licences and training has been delivered. Use is voluntary and uneven. Observable signal: nobody can say who uses it; you have to request a report from the vendor.

Stage 2 · Redesigned process. A specific process has been rewritten around the tool: deadlines, deliverables or review points have changed. Observable signal: if the tool were removed tomorrow, that team would have a real operational problem.

Stage 3 · Distributed competence. Several areas have been through the same thing, people teach other people, and judgement transfers without formal training. Observable signal: uses appear that nobody designed from above, and they are good.

The stage 2 test is the most honest in the series: if removing the tool does not hurt, there is no adoption.

The first 90 days

Weeks 1-2 · Pick the group, not the workforce. Five to fifteen people who perform the process of the chosen case. General literacy can run in parallel, but it is not this.

Weeks 3-4 · Task map. Break the job into tasks and mark which change, which disappear and which appear. This is usually where you discover the saving forecast in the value layer was allocated to the wrong place.

Weeks 5-8 · Process redesign. New deadlines, new review points, new quality criteria. This is operations work, not technology work, and it is the part that almost never happens.

Weeks 9-12 · Measure sustained use and collect friction. What gets abandoned and why. Reasons for abandonment are the best improvement material there is, and almost nobody asks for them.

What NOT to do yet

• Do not roll out to the whole workforce. Broad deployment before a redesigned process turns a good tool into noise and burns the budget.
• Do not measure training. Measure use at eight weeks.
• Do not make adoption a personal objective. Putting AI use into performance reviews produces simulated use, which pollutes the data and destroys trust.
• Do not promise there will be no impact on roles if you do not know. If the project changes people's work, saying so clearly and explaining how is more effective than reassurance you cannot guarantee.

What skipping it costs

It costs the use case, even when everything else is done well. A technically correct system, on correct data, with correct governance, that nobody uses, produces exactly zero.

And it costs the next round. The organisation learns that "the AI thing" was a tool that got announced and changed nothing, and that memory is expensive: the second project starts with accumulated scepticism and with the most capable people less willing to invest their time.

Fifth layer of the series on the seven layers of AI maturity.

Frequently asked questions

Why is nobody using the AI tool we bought?

Almost always because the process did not change. If deadlines, deliverables and quality criteria stay the same, using the tool is additional work. Adoption is not achieved by training: it is achieved by redesigning the work of the specific people who perform that process.

Should the whole workforce be trained in AI?

General literacy has cultural value but does not produce use cases. Use cases come from a small group — five to fifteen people — whose specific process gets redesigned. It is worth not confusing a culture programme with an adoption project.

How do you measure AI adoption?

By sustained use at eight weeks without reminders, and by what it is actually used for. The definitive test is simpler: if the tool were withdrawn tomorrow, would that team have an operational problem? If not, there is no adoption.

Should you tell people AI will change their job?

If it will, yes, and explain how. Promising nothing will change when the project exists precisely to change something destroys the credibility of the whole programme. Profile transformation is the more likely scenario, more so than substitution.