I spent Saturday morning on a Teams call with the head of digital at a 3,500+ colleague European financial services firm, watching her open Microsoft Copilot and ask it, in front of me, to summarise an email she had already read.

It was the third time in twenty-two days I had watched a senior person do something like this. The pattern is unmistakable. The licences are deployed, the rollout was on schedule, the prompt-a-thon went brilliantly, the AI Champions network sends out a weekly tips email that almost nobody reads, and the active-user dashboard looks lovely. But when you actually watch what people do, Copilot is being used as Google with a slightly more verbose interface.

The industry has a name for this now: the Prompting Plateau. A sizeable consulting industry has sprung up to tell you that the cure is a more sophisticated training programme – prompt-a-thons, internal AI Champions networks, communities of practice, hands-on workshops, gamified onboarding. Microsoft themselves now offer a Copilot Prompt-a-thon Workshop as a productised service through their training partners. There are kits. There are kits with worksheets. There are now kits with worksheets in nine additional languages.

I have a problem with all of this. Not because prompt-a-thons don’t work – they sort of do – but because they mostly fix the wrong thing. Both sides of this argument deserve a proper hearing before I get to that.

The case for the literacy-led approach

The advocates aren’t wrong. Passive video training for AI tools is genuinely useless. Research going back to 2024 consistently shows the same thing: watching someone else use a model is not how anyone actually learns to use one. Harvard Business Review’s work on Learning to Learn Together finds that teams who learn collaboratively adopt new technology roughly three times faster than teams who don’t. Microsoft’s own Work Trend Index has teams learning AI together reporting 35% higher satisfaction and productivity than those using traditional training routes. The UK government’s civil service pilot – over twenty thousand users, found an average of 26 minutes a day saved with Copilot, which extrapolates to roughly two working weeks per person per year, and is not nothing.

If your starting position is the four-hour video module HR built two years ago that nobody ever finished, a prompt-a-thon is a meaningful improvement. The hands-on bit matters. The collaborative bit matters. Picking three internal champions, giving them time off to actually become good at Copilot, and letting them teach their teams is materially better than the alternative of nobody being any good and nobody teaching anyone.

I’m not the person telling you to scrap the prompt-a-thon. The mistake is thinking the prompt-a-thon is the whole strategy.

The case against

Here’s the bit nobody seems to want to put in a slide deck.

MIT’s Media Lab published research this year finding that 95% of enterprise AI initiatives have produced no measurable P&L return. McKinsey’s most recent state-of-AI work has 87% of executives reporting active AI use and 12% reporting measurable strategic value – a 75-point gap between what people are doing and what’s showing up in the accounts.

The New Yorker has drawn the obvious parallel to the 1980s “productivity paradox”, when computers were visible everywhere except in the productivity statistics. The industry has even started using a phrase for what’s actually happening: efficiency theatre. The dashboards look impressive, the seconds-saved metric is up and to the right, and the P&L doesn’t move.

If your prompt-a-thon programme is moving the active-user count without moving anything that survives audit, you’ve built a slightly more expensive version of the already-agreed-to-be-less-than-effective four-hour video module. You’ve substituted the appearance of training for the appearance of literacy. You’ve made the dashboard look better. You haven’t made the organisation more productive.

Both views are right. The prompt-a-thon advocates are right that the format works better than passive training. The 95-percent-of-AI-initiatives-fail crowd are right that none of this is showing up in productivity. The two camps are mostly not talking to each other, which is partly why the question almost nobody seems to be asking is this: how can both of those things be true at the same time?

The actual diagnosis

Here’s what I think is going on, having watched this from inside a fair number of UK and European organisations over the last eighteen months, and forgive me if it offends the standard playbook.

The Prompting Plateau is not, principally, a literacy problem. It’s an incentive problem dressed up as a literacy problem.

Most knowledge workers in most organisations aren’t using Copilot for deep cognitive work for an extremely rational reason: nobody’s asking them to. Their performance review is built on the artefacts they produced last year – emails sent, decks made, hours billed, slides delivered, tickets closed. Their manager checks whether the deck looks right, not whether they used a model to interrogate the underlying assumptions. The path of least resistance is to use Copilot to draft the email faster, polish the slide, summarise the meeting, and call it a day. Everything else is uncompensated risk: novel ways of working that might go wrong, that the manager doesn’t understand, that nobody else on the team is doing, and that don’t show up on any review framework anyone is using.

In that environment, telling employees to use Copilot for analysis rather than email drafting isn’t a literacy intervention. It’s asking them to volunteer for additional risk, for no additional reward, to make a dashboard nobody on their pay grade has ever seen look better. They’re entirely rational not to bother.

Run a prompt-a-thon into that environment and you’ll get a brilliant 90-minute session in which everybody’s energised, a handful of genuinely good prompts written down, and within ten working days the active-user count quietly returns to baseline.

What actually works

Three things, in roughly this order, and they’re uncomfortable because none of them are simple training programmes.

First, rewire the incentives before you rewire the skills. This is the bit that gets handed to HR and quietly forgotten. If you want your team to use Copilot for analysis rather than email, the analysis has to be what gets reviewed, recognised and promoted. That means rewriting job descriptions, manager 1:1 templates, OKRs, and at least one ladder rung of the career framework. It’s hard, it cuts across organisational politics, and it’s the bit your Copilot consultant won’t bring up because they don’t own it and can’t bill for it. It’s also the only thing on this list that materially changes behaviour.

Second, train depth on the bits that matter, not Copilot in general. The prompt-a-thon, run properly, should be aimed at one team doing one workflow that actually matters to the business – the procurement team reviewing supplier contracts, the underwriters assessing case files, the marketing analysts producing the weekly trading report. Generic Copilot literacy programmes scale beautifully on a slide and regularly deliver almost nothing. Twenty hours of intensive depth on one team’s actual work, with their actual data, and with a measurable before-and-after on their actual deliverable, delivers far more than three hundred hours of generic training across the organisation. The reason organisations don’t do this is that intensive depth is expensive, hard to staff, and looks much less impressive in a quarterly board update than “we ran twenty-three prompt-a-thons reaching 4,200 employees”.

Third, stop measuring active users. Active-user counts are the vanity metric of AI adoption, and we have two recent and very British examples of what happens when they become the target.

Take the UK NHS App rollout. During various phases of digital transformation, massive emphasis was placed on user-acquisition and active-login counts to prove to Whitehall that the digital investment was working. To hit deployment targets, local trusts started essentially forcing active usage – patients directed to log into the app while standing at the reception desk to book an appointment that could have been done verbally in three seconds. The dashboards showed a beautiful spike in active users. The reality was millions of people logging in once a year, completely frustrated, using none of the deep health-management features the app was actually built for.

Or the UK civil service and the Teams green light. When hybrid work tracking arrived, the active-user status indicator on Microsoft Teams became the metric people were monitored against. This directly created a booming UK market for mouse jigglers and software scripts designed to keep the status icon green. Active-user counts and time-online metrics were flawless. Actual output fell through the floor because staff were optimising for presence rather than performance.

Active-user counts are easy to instrument, they go up, and they tell you absolutely nothing about whether an organisation is getting any value. Replace them with two metrics that actually matter.

First: time taken to complete a specific business workflow, before and after AI is embedded.

Second: number of workflows where AI is now in the critical path, weighted by their financial materiality.

Both are harder to measure. Both will go up more slowly than the active-user count. Both are defensible in a board meeting in a way that “78% active user rate” emphatically is not – especially in front of an auditor who’s read the same MIT note you have.

The future

The organisations that benefit most from AI in 2027 won’t be the ones with the best prompt-a-thons or the largest AI Champions networks. They’ll be the ones whose performance management systems finally caught up with the technology – where being good with AI is measurably rewarded, where the deep cognitive uses are what gets recognised, and where the dashboard the board looks at measures productivity rather than activity. I haven’t seen many of these yet. I’ve seen several boards that think they’re the one because their active-user count is 78%.

The Prompting Plateau is not, in the end, about prompting. It’s about incentives, measurement, and the unglamorous business of changing how an organisation actually evaluates the work that gets done inside it. None of that fits onto a kit with worksheets. All of it is what your AI Champions, bless them, can’t fix from a Tuesday lunchtime session.

The organisations that crack this now will have a roughly two-year head start, because the rest of the industry’s going to spend that time perfecting the prompt-a-thon. Almost nobody seems to have noticed that this is what they ought to be doing instead.

I’d push you to be one of the few who do.

Andy McGurk is a Fractional Chief AI Officer and the founder of AMVEN. He helps UK and European organisations build AI strategies that survive contact with reality, and is increasingly grumpy about adoption metrics.