Twenty years of being early

I have usually been early. I see a technology, I can feel where it is going, and I back it before most of the people around me are ready to.

And then, almost every time, I am surprised by how long it takes to actually land. The capability arrives on schedule. The impact turns up years later, usually after everyone has stopped talking about it.

At Reed Business Information (now RELX) we had a name for that job. Futurist. It sat inside innovation and corporate development, and the useful part of it was never predicting the technology, which is comparatively easy. The hard part was predicting the lag: the distance between something becoming possible and something changing a P&L.

Predicting the technology is the easy part. Predicting the lag is where I have spent most of my career.

I have spent a good part of this year talking to business leaders, technology leaders, owners and investors about the future of their organisations, and for the first time in twenty years I am not confident I know how long that lag is any more.

What can actually be measured

I am wary of "everything is accelerating" as an argument. People have said it at every point in my career and it has never been quite true, so it is worth being careful about what has changed.

On the capability side, the most useful measure I have found comes from METR, who track how long a task a frontier AI agent can complete unaided, at a fifty per cent success rate, on software and technical work. That length doubled roughly every seven months from 2019. Since 2024 the doubling time has been closer to four.

It is a narrow benchmark. It measures coding rather than the full range of knowledge work, and METR say themselves that the top of the range is getting hard to measure reliably. Even so, a doubling time that halves is not noise.

On the adoption side the numbers are easier to read. Researchers at the St. Louis Fed and Harvard looked at how quickly people started using generative AI, at work and at home, and compared it with how quickly people took up the PC and the internet after each first went on sale. At work it is spreading about as fast as the PC did. Overall, faster than either.

They point out the comparison is not perfect, because nobody had to buy new kit or wait for a connection this time; AI turned up on devices people already had. I would argue that is part of the point. The infrastructure was already there, so the usual delay did not happen.

Put those two together and that is what is new to me. I have watched plenty of technologies get dramatically better very quickly. What normally happens next is that businesses take ten years to catch up. This time they are not waiting, and I cannot fully explain why yet.

What I hear in the room

Every conversation gets to AI within five minutes. I usually widen it back out to "technology", because AI is brilliant and transformational and also, quite often, a sledgehammer for a nut.

Some of the most commercially valuable work I have been part of was prediction on incomplete data a decade before anyone said GenAI. I remember using natural language generation to take data from spreadsheets and turn it into commercial reports that were sold into the aviation industry, and building predictive models for where ships were going based on partial signals. Nobody called it AI. The technologists might have called it ML. It was the precursor to everything we see today.

Whatever the label, the same worry keeps surfacing, though people do not always have the words for it. Even the most mature organisations, the ones rolling AI tooling out to their whole workforce, are uneasy about how much of their future is being bet on one provider's roadmap and pricing.

What they are asking for, underneath, is not "which model should we use" but "how do we build so that it does not matter".

Own the memory, rent the model

The answer I have settled on is this. Models are replaceable, and getting more so every quarter. Your organisation's memory should not be.

Your data, your knowledge, your workflows and the decisions you have already made are the durable layer. They accumulate and they are yours. Models are something you connect to that layer, probably from more than one vendor, and probably a mix of frontier cloud models for hard reasoning and smaller local ones for anything sensitive or high volume. Choose by scenario, not by brand.

Most people nod at that and then keep their knowledge inside a vendor's product anyway, because leaving is theoretical. The contract says you can switch. What makes switching real is holding a test set built on your own documents and your own past decisions, versioned, so that when you move to a new model you can show it still performs against your work rather than against a public leaderboard.

Almost nobody I speak to has this. It is not a large piece of work. It is the difference between being model-agnostic on paper and being model-agnostic in practice.

Saving time is not the same as saving money

None of it matters, though, until the productivity turns into money, and this is where most of the programmes I see stall.

Where I have measured this on real workflows the pattern has been consistent: the uplift on repeatable work is several times what you get on judgement work. That is useful, but it is only half the sum.

What AI gives you is time. Something that took a person a day now takes an afternoon. That is real, but it is not money yet. It only becomes money when you do one of two things with the time: give that person more work that earns something, or take the cost out.

Most programmes never make that choice, because it is an uncomfortable one and nobody clearly owns it. So the saving sits in a slide as a percentage, the headcount stays where it was, and a year later the CFO asks where the money went.

So what is scarce now?

If producing code is no longer scarce, and intelligence can be rented by the token, the question I keep coming back to with leadership teams is what their company actually holds.

The honest answer is usually shorter than people expect. Proprietary data. Judgement, and the tests that encode it. Distribution and relationships. The knowledge of how to turn a capability into something a customer will pay for, which is a much rarer skill than the capability itself.

Building reliable systems is still expensive, and anyone who tells you otherwise has not run one in production. But the expensive part has moved, from making things to knowing what to make and proving it works.

I may be wrong about whether the lag has gone for good. I have been wrong about timing before, in the other direction, and I would not be surprised if some of this takes longer to bite than it feels like it should. But I would not build a five-year plan on the assumption that the old gap is coming back.

---

Sources: METR, Task-Completion Time Horizons of Frontier AI Models. Bick, Blandin and Deming, "The Rapid Adoption of Generative AI", Management Science, 2026 (summary at VoxEU).