Activity Is Not Capability: The Six Levels of AI Adoption

Experimentation is easily mistaken for transformation.

Knowledge teams are natural early adopters. Their work is language, synthesis and judgement, and the output looks useful immediately. That is exactly why AI looks more mature in language-heavy work than it is. The output is public. The method behind it, what was entered, what was rejected, what was corrected and why, stays private with the individual.

The numbers most firms report reflect this. They show use. They do not show adoption. This post sits inside the operating model and looks at one part of it in detail: how to tell activity apart from capability, and why so many programmes stall before they get there.

Easy to count is not the same as worth knowing

Adoption is being measured where it is easiest to see, not where it creates value. In a 2024 review, the Bank of England and FCA found that 46% of UK financial services firms using AI only partly understood the technologies they were running, while 75% of firms surveyed were using it. Across the wider economy, DSIT reported that 77% of adopters had seen no change in revenue, and fewer than half of staff in adopting firms actually used the tools.

The metrics that are easy to count are seductive because they move:

  • Licences issued
  • Active users
  • Prompts shared
  • Time reportedly saved
  • Pilots launched
  • Positive sentiment

None of these answer the questions a leader actually has to answer:

  • Which tasks are suitable
  • Whether the method is repeatable
  • Whether output quality is consistent
  • Whether savings survive review and rework
  • Whether scaling is justified
  • Whether risk remains acceptable

Scaling access before establishing what good looks like does not scale capability. It scales variation: more people producing work in more different ways, with usage itself signalled as the desired outcome.

Visible activity is evidence of use. It is not proof of operational adoption or fit.

Six levels, from tool access to measured transformation

Adoption is best understood as a transfer of ownership. It moves from the individual, to the team, to the organisation. There are six levels.

Individual owns it.

  1. Tool access. People can use an approved or unapproved tool. The firm knows who holds a licence, not what it changes.
  2. Individual experimentation. Prompts tested, outputs compared, personal techniques developed. Quick, informal, invisible from outside.
  3. Personal productivity. Regular users get faster at chosen tasks. The gain depends on their experience and their ability to spot weak output.

Team owns it.

  1. Shared team practice. Methods leave personal chat histories. Templates, examples, standards and boundaries get shared.
  2. Operational integration. AI sits inside a defined workflow. Inputs, review points, ownership, escalation and quality measures are explicit.

Organisation owns it.

  1. Measured transformation. Repeatable improvement in productivity, quality, outcomes or cost that no longer depends on a few enthusiasts.

The first three levels need almost no leadership. Someone who finds that an hour of rewriting now takes twenty minutes does not need a programme to keep doing it. They will keep doing it because it helps them. Every metric in the easy-to-count list can look healthy while every use case still sits at level three.

Why firms stall at level four

The step to level four is different in kind, not degree. It means converting tacit judgement into an explicit method, and deciding which parts of one person's approach transfer and which depend on expertise no template captures.

This is where most organisations stall.

The reason is structural, not a lack of effort. The prompt is visible, so it gets the focus. The surrounding judgement does not. An experienced user knows which tasks to avoid, how much context the model needs and when an answer merely sounds convincing. Ask that person to publish a prompt library and you transfer the visible technique while leaving the reasoning behind. Colleagues copy the instruction without knowing why the task was chosen, which source was trusted, or how much editing was normal.

Formal adoption, practical understanding and operational control mature at different speeds. A tool can improve the first stage of a task without improving the workflow, and a person can get faster without the team getting more capable. Levels one to three can be reached by individuals. Levels four to six can only be reached by design.

What crossing the line actually requires

Moving past level three is not about better prompting. It is about making a method reproducible, owned and measured beyond one user. That requires deciding which tasks are suitable in the first place, then wrapping the suitable ones in a sequence that does not depend on anyone's memory.

Two questions follow directly from this one. The first is which tasks belong to AI at all, which we cover in where AI earns its place, and where a rule is safer. The second is how to hold a method in place once you have chosen it, which is the subject of governed workflows.

Adoption has happened when the method survives the departure of its most enthusiastic user. Until then, you have activity, not capability.

If you want the full argument in one place, download the full field note. If you have a workflow stuck at level three or four and want a plain read on what it would take to move it, send us the use case and we will tell you honestly whether it is worth the step.