The 6% Club

I ask two questions in every executive briefing. First: who has generative AI in use? Nearly every hand goes up. Second: keep your hand up if you can state the return in Swiss francs. Around the table, the hands come down.

McKinsey's State of AI survey (2025) measured the same gap at scale. Six percent of companies report a significant EBIT effect from AI. The other 94 percent are still looking for one. Most of them have models in production, a strategy document, and a growing inventory of pilots.

The six percent form a club, and membership has little to do with technology access. Every bank can buy the same models from the same vendors, at prices that fall every quarter. What the members share is operational. Three patterns recur.

Breadth beats the perfect pilot

The 94 percent concentrate effort on a single showcase: one process, heavily engineered, presented to the board twice a year. The members of the club deploy across many processes at once and accept that several attempts will fail. They run AI as a portfolio and judge it on expected value across the whole book. An institution running 30 use cases generates 30 streams of evidence about where value actually sits. An institution polishing one pilot generates a demo.

No use case is ever finished

A deployment tuned in January runs on January's models, January's prompts, and January's process assumptions. By June, all three are stale. Club members treat every use case as permanent work in progress: re-tuned as models improve, re-scoped as adjacent steps become automatable, re-measured against the same financial baseline. The 94 percent staff AI initiatives as projects with an end date. The six percent budget continuous sharpening as standing capacity.

What does not pay is stopped

The least glamorous habit is the most discriminating. In most organisations, terminating a use case is recorded as a failure, so nothing is terminated. Pilots accumulate: politically alive, economically dead. The club kills without sentiment. If a use case cannot demonstrate value in francs after a defined period, it stops, and the capacity moves to one that can. Breadth without the discipline to stop is just more pilots.

Underneath the three habits sits a structural difference. The 94 percent deploy AI as a tool: something an employee opens, uses, and closes. The club deploys it as a decision layer: agents acting within designed guardrails, embedded in end-to-end processes, with humans at the edges where judgement and accountability live. Tools produce convenience. Decision layers produce EBIT effects.

Two mental models explain why so few institutions get there.

Thirty steps

Take 30 ordinary steps, at a walking stride of 70 centimetres, and you cover about 21 metres. Now take 30 exponential steps, doubling the distance each time: 70 centimetres, then 1.4 metres, then 2.8. After 30 doublings you are more than 750'000 kilometres out. The moon is 384'400 kilometres away. Thirty doublings take you to the moon and most of the way back.

Everyone computes the first number instantly. Almost nobody's intuition survives the second. Human planning is calibrated to linear change, and this technology does not move linearly: model capability, cost per token, and the scope of what agents can be trusted to do have compounded for three consecutive years. A budget cycle that treats next year's AI as this year's AI plus ten percent is mis-specified from the first line. Today's model is the worst model you will ever work with.

The dip is the plan

The second model explains the abandonment rate. Erik Brynjolfsson, Daniel Rock, and Chad Syverson call it the productivity J-curve (American Economic Journal: Macroeconomics, 2021): general-purpose technologies depress measured productivity before they raise it, because the early years go into complements that standard accounting treats as pure cost. Process redesign, data plumbing, skills, governance. Output dips while the foundation is poured.

Applied to AI in a bank: the first months of an agentic deployment cost money, slow teams down, and produce awkward metrics. This is the point where the 94 percent stop, which means they pay the entry price and decline the payoff. The dip is not a flaw in the plan. It is the plan. Business cases that promise payback by month three guarantee abandonment in month four.

The members of the club size the dip, tell the board about it in advance, and hold course through it. Endurance here is a governance property, and it is decided before the first franc is spent.

Doing is believing

The third habit is the least comfortable, because it is personal. Executives who experience agents only through steering committees calibrate on slides. Slides do not convey what it means when an agent researches, drafts, critiques, and finishes a piece of work while you sit in a meeting. Judgement about this technology is built through exposure, and exposure cannot be delegated.

The practical version: use an agent on your own work this week. Draft the board paper with one. Let one prepare the client briefing and then mark its errors. Leadership teams in the six percent tend to do this routinely, which is why their risk appetite is calibrated rather than borrowed.

The urgency is generational as much as competitive. The next cohort of clients will arrive already fluent: my daughter will be agent-native before she is old enough to open a bank account. When every client sends an agent, institutions still running AI as a pilot programme will negotiate with counterparties that finished their J-curve years earlier.

The entry price

Locating your institution takes five minutes: the 15-question assessment measures readiness across strategy, operations, and governance. The pattern it most often reveals is the one this post describes: technology in place, habits missing.

The 6% Club has no waiting list and no membership fee. It has an entry price: the dip, taken deliberately, across a broad portfolio, with the discipline to stop what does not pay. Most of the market is still deciding whether to pay it.