Usage Theatre

Usage Theatre is what organisations perform when they cannot measure the value of AI and measure its use instead.

Hirji's shorter name for the same thing is AI pretend: "Move from AI pretend to operating reality. Most teams do not need more tools." Two words for what the rest of this page takes a paragraph to describe.

Adoption dashboards go up. Licence counts become KPIs. Employees learn to perform the metric rather than improve the work. Much of the underlying adoption is real; what is being performed is the usage, because usage is the thing that can be counted by Friday.

The measure of an AI programme is whether decisions got better. That is harder to count, so few organisations count it.

SuperSkills uses the term to describe this pattern. No claim of first use is made: the phrase is not documented here with a dated first publication.

What gets measured instead of capability#

Seats issued. Monthly active users. Prompts per head. Percentage of staff onboarded. Training modules completed. Hours reportedly saved.

Every one of those can rise while nothing changes about what the organisation can do. They are measures of activity, and activity is not the same thing as capability, though it is considerably easier to put in a board pack.

The measure that actively misleads#

Self-reported time saved is not a weak proxy. In the one randomised trial that checked it, the number had the wrong sign.

METR ran 16 experienced developers across 246 real tasks, randomly assigning AI permission. Participants forecast a 24 per cent speed-up. They were measured as 19 per cent slower. Afterwards, having lived through the slowdown, they still estimated AI had made them about 20 per cent faster.

Updated 28 August 2026. METR withdrew this as a signal of the current effect on 24 February 2026. Their second study estimates a speed-up of 18 per cent for returning developers, confidence interval -38 to +9, and they believe developers are likely faster with AI in 2026 than in 2025. They also say their own data is weak evidence, because 30 to 50 per cent of developers declined to submit tasks they did not want to do without AI. The 19 per cent belongs to early 2025 and is quoted here as a historical measurement. What survives untouched is the perception gap this page is built on.

A forty-point gap between belief and measurement, persisting after direct experience. The sample is small and specific to experienced developers on mature codebases, and it should not be generalised to all work. It is more than enough to retire one practice: asking people whether AI made them faster produces a number that may point the wrong way. Most organisations are running exactly that survey.

Why organisations do it anyway#

Not stupidity. Three rational pressures.

Capability is genuinely hard to measure, and adoption is easy. Given a quarterly reporting cycle, the easy number wins.

The easy number is flattering. Nobody is rewarded for reporting that a large investment has not yet changed anything.

There is now a compliance incentive. Article 4 of the EU AI Act has required a sufficient level of AI literacy since February 2025, at every risk tier, with enforcement from August 2026. A completion rate looks like evidence of compliance and is not evidence of capability. See what AI literacy means for leaders.

Adoption metrics are not worthless, and nobody has tested the rest#

That adoption metrics are worthless. They tell you whether people have access and whether anyone is using it, which are real prerequisites and worth knowing.

Nor that organisations measuring this way perform worse. Nobody has tested it. The claim is narrower and harder to dodge: an organisation measuring only usage has bought a dashboard that cannot detect its most serious risk, and will keep reporting green until something breaks.

What drift looks like on a dashboard#

Usage theatre is the measurement expression of drift. Nobody decided to measure the wrong thing. The available number was collected, reported, and became the definition of progress by repetition.

The consequence is specific. If usage is the metric, the rational employee response is to use the tool more, including on work where they should not. If capability is the metric, the rational response is to get better at judging output. Those produce very different organisations within about two years, and only one of them can pass an Article 14 audit.

The awkward corollary for leaders: the number you would most want to know, whether your people can still do the work unaided, is the one nobody collects, because collecting it means taking people off productive work to sit something resembling an exam. That cost is real. It is also the only measure that answers the question.

What to measure instead#

• Unaided capability, sampled. Twice a year, on real work. The only measure that detects the thing everyone claims to be worried about.
• The pairing against the better half. Does human plus system beat the better of either alone? Most deployments have never checked.
• Individual-level outcomes, not averages. Effects run in opposite directions between people, so an average conceals who is being helped and who is being harmed.
• Where the time went. If AI saved hours, name what they became. See the unclaimed hour.
• Overrides. Zero in a quarter is evidence of an untested right, not a good system.

The full version is in how to measure AI adoption properly.

Related SuperSkills research#

On the parent pattern, drift versus design. On readiness claims, the AI readiness lie. On what erodes underneath, capability debt. On the board conversation, what should a board ask about AI. On the workforce plan, AI workforce strategy.

Development of the idea#

Related argument in the Box of Amazing essay This Is Zombie Work (1 June 2025) and in The Cult of Productivity is Breaking People (22 June 2025). The commercial version is in Irish Tech News, July 2026.

Key sources

• METR (2025). Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity.
• Vaccaro, M., Almaatouq, A. and Malone, T. (2024). When combinations of humans and AI are useful.
• Yu, F. et al. (2024). Heterogeneity and predictors of the effects of AI assistance on radiologists.
• Humlum, A. and Vestergaard, E. (2025). Still Waters, Rapid Currents.

About this research#

Written by Rahim Hirji, author of SuperSkills (Kogan Page, 2026), keynote speaker on AI and human capability, and founder of The SuperSkills Intelligence Company.

How this research works · Reviewed quarterly · Found an error? Tell me and it is corrected on the page.