Nobody is short of AI pilots. They are short of an agenda.

There is a meeting I have sat in more than once, at more than one firm, and it always runs the same way. The AI pilot is somewhere in the middle of the agenda, after the budget review and before any other business. It has been running for a few months. It works, in the sense that the demo does what the demo was supposed to do. The discussion takes about ten minutes and ends with an agreement to look at a second use case next quarter.

Nobody in the room thinks that is slow. I do, and I have spent a good part of the last year trying to work out why so many well-run firms treat AI the way those committees do: one careful experiment at a time, each one waiting for the previous one to finish.

The pattern

The shape is familiar. A firm picks one or two small, safe ideas. A document summariser, a support-ticket classifier, something in finance. The idea gets validated, usually by a small team that is good at this and has some slack. Once it works, the door opens to a few more experiments. Progress is real but sequential, and each step waits for the last one to earn permission.

I assumed this was a small-firm habit. The data says it is broader than that, and also that the large-firm version has the opposite symptom.

Among European mid-market firms, the sequential version is the norm. Harmonised central-bank surveys put generative AI use in Italian firms at 5% in 2024, rising to 24% a year later, and in every country surveyed fewer than 4% of firms make intensive use of it. Most adopters are light adopters. One tool, one team, one use case.

Among large enterprises the problem inverts. McKinsey's 2025 survey found 88% of firms using AI somewhere and nearly two thirds not yet scaling it anywhere. Deloitte found that only a quarter of firms have moved even 40% of their pilots into production. One vendor survey of 835 firms found that 60% had run fifteen or more pilots and 84% had at least one that never shipped. IDC's version of the same story: four proofs of concept out of thirty-three reach production.

Banking is the clearest case. Evident tracks the fifty largest banks and counted their disclosed AI use cases tripling in a year, from 56 to 117, while more than 70% of those banks still report no return from any of them. More pilots did not produce more value. IBM asked 2,000 CEOs about this earlier in 2025 and half of them said the pace of AI investment, driven by fear of missing out, had left them with "disconnected, piecemeal technology".

So the large firm has fifty pilots and the mid-sized firm has one, and neither has an agenda. The shy version and the scattered version are the same failure. Nobody has decided what AI is for in this business, so every initiative has to justify itself from scratch, and most of them never get the chance.

Why sequential is expensive now

The usual defence of the cautious approach is that AI is moving fast, so it is prudent to wait for things to settle. I think that gets the risk backwards.

It is true that the technology layer will not hold still. Inference at GPT-3.5 quality cost around 20 dollars per million tokens in late 2022 and about 7 cents two years later, a 280-fold fall according to Stanford's AI Index. Anthropic has retired every model in its 3.5, 3.7 and 4 families twelve to thirteen months after release, and its notice period has shrunk from six months to sixty days. Anything built tightly around a specific model in 2024 has already been rebuilt or is running on borrowed time.

But that is an argument against building tightly around a model, not against building. The part of an AI investment that goes obsolete is the model layer, and if the architecture is right, swapping it is a configuration exercise. The gap between the best open-weight models and the best closed ones has narrowed from 8% to under 2% on the standard benchmarks, so the model is closer to a commodity than it has ever been. I wrote about that in the last issue. The part that does not go obsolete is everything the firm learns along the way: which of its data is actually usable, which processes are worth touching, who in the organisation can evaluate a probabilistic system and sign it off, what the compliance function will accept. None of that transfers from a vendor. All of it compounds.

A firm running one pilot at a time is buying that learning at the slowest possible rate, while its competitors buy it in parallel. That is the cost of not getting on the train, and it does not show up in any budget line. BCG's survey of 1,250 firms in September 2025 found that the 5% it calls "future-built" are growing revenue at 1.7 times the rate of the laggards, with 3.6 times the shareholder return over three years. The ECB's own survey of 20,000 euro-area workers this summer put the time saved by those using AI at three hours a week, close to 8% of working time, which is a number a CFO can do something with. I would treat consultancy multiples with some care, and Goldman Sachs is right that at the level of the whole economy the productivity effect is still hard to see. But at the level of the firm the direction is consistent across every survey I have read, and I have not found one that shows the cautious firms catching up.

What an agenda actually is

Not a strategy deck. Every firm I have seen that is stuck has one of those.

An agenda is three things a firm can point to, and most cannot.

A backlog. A ranked list of candidate AI use cases across the business, with a rough size and a rough value against each, maintained the way a product backlog is maintained. Not a wish list from a workshop, a working document that someone owns and that gets pruned. Pruned is the operative word. BCG's AI Radar found that the firms capturing value prioritise three or four use cases at a time; everyone else runs six. Capgemini found in 2024 that 6% of retail banks had an enterprise roadmap for AI at scale. Banks are not unusual here.

A cadence. Use cases enter the backlog and leave it on a rhythm, monthly or quarterly, independent of whether the last one succeeded. This is the single change that breaks the sequential pattern. If the second experiment can only start when the first one has proven itself, the firm has accidentally built a serial pipeline with a five-month cycle time.

An owner with a budget. Someone senior enough to say no to a pilot that has drifted, and to fund the unglamorous prerequisites, chiefly data readiness and evaluation design, that no individual use case will ever pay for on its own. The surveys are unambiguous on this. Accenture, looking at 2,000 companies, found that C-suite sponsorship makes an AI programme 2.4 times more likely to beat its return expectations, and that scaling even one strategic bet end to end makes a firm three times more likely to. The firms McKinsey classes as high performers are three times more likely to have redesigned the workflow around the AI rather than bolting it on.

None of this is new thinking. It is what any firm does with its product roadmap or its capex plan. The odd thing is how rarely it is applied to AI, where the argument for it is stronger than almost anywhere else, because the technology decays fast and the organisational learning does not.

The honest caveat

An agenda does not make a bad use case good. Plenty of firms with a beautifully maintained AI roadmap are shipping things nobody asked for. The point of the backlog is not to do more, it is to decide faster what not to do, and to stop each decision from having to wait for the last one.

Most of those firms have since started a second pilot. It sits a little higher up the agenda now.

Which item is yours?