An AI performance of progress

Most of what is sold as AI strategy is a performance of progress

Why so many AI programmes are busy, expensive, and quietly going nowhere — and what an honest look would find instead.

And yes — before anyone reaches for the comment box — the em-dashes are mine. I had them long before the machines made them fashionable, and I'm not about to surrender my punctuation simply because a machine now uses it too. Which, as it happens, is more or less the argument that follows.

Let me concede the impressive part at once, so we needn't keep circling back to it.

The models are extraordinary. They write, summarise, translate and reason their way through tasks that humbled software a few short years ago, and they improve on a timetable that would embarrass most of the people reading this. If the question were is the technology real, there would be nothing to discuss and no article to write.

But that was never the question your organisation actually faces. The decision came first. The strategy was assembled to flatter it.

Notice how the commitment to AI was usually arrived at — someone spends a long evening in conversation with a language model, comes away faintly electrified, and concludes that AI is the future. A strategy is then assembled to ratify a decision that had, in truth, already been taken in a moment of private wonder. The analysis is summoned to flatter the verdict.

There is a name for this, and it is not strategy. It is decision-based evidence-making — and once you have seen it, you will not stop seeing it.

It is enforced by a particular kind of sentence. Companies that fail to adopt AI will be left behind. This is dressed as a forecast. It operates as a threat: surrender your attention and your judgement now, or be counted among the doomed. As the Oxford philosopher Carissa Véliz argues in Prophecy, a prediction of this kind does not so much foretell the future as bully the present into building it — its real power is over behaviour, not knowledge. A great deal of what passes for AI strategy is simple obedience to a prophecy no one in the room was brave enough to examine.

Where it actually fails — and it is almost never the model

Here is the most reliable finding of the era, and also its least advertised: where AI disappoints, it almost never disappoints on the model. MIT's study of enterprise AI found that the overwhelming majority of corporate pilots deliver no measurable return — and that the cause is rarely the technology itself, but the "learning gap" between a capable model and an organisation that has not changed to use it.

It disappoints on the data that was never fit for the purpose it was abruptly conscripted into. On the governance improvised at speed once something had already gone wrong. On the decision rights nobody remembered to revise. On the workflow that quietly routed around the clever new tool because the old way still paid the bonuses.

The model arrives brilliant and context-free. It lands in an organisation built for a different kind of work — and the organisation, being an organisation and therefore a creature of magnificent inertia, politely declines to rearrange itself for the technology's convenience.

This is the oldest finding in the study of institutions, and the present moment is rediscovering it at spectacular expense: a tool changes nothing on its own. The organisation has to change to use it — and changing an organisation is political, effortful, slow, and nobody's idea of a quick win. It is also, with tedious reliability, the entire question of whether the thing works. As researchers at Harvard and HKU put it, most firms struggle to capture value from AI "not because the technology fails — but because their people, processes, and politics do."

Hence the cruellest law in the field: the demonstration never survives contact with the org chart. A demo is a controlled environment. An organisation is the living opposite of one — indeed the controlled environment is the single thing it has never once possessed. Strip a company back to its essence and you don't find technology, or even data; you find a structure of commitments between people who hold the authority to make them, and that is precisely the layer a model cannot reach.

The promise that quietly reversed itself

There is a darker comedy underneath all this.

AI was promised to the workforce as relief — the tedium automated, the evenings returned. What frequently arrives instead is the reverse. The tool absorbs the easy work, raises the expected output, and leaves everyone busier than before. The thing that saves time becomes the thing that makes you spend more of it — a dynamic Harvard Business Review researchers christened "workslop": polished-looking AI output that lacks the substance to advance the task and quietly generates more work for whoever receives it.

By the time anyone notices, the people who must actually execute the strategy are exhausted, and the people who announced it have moved on to announcing the next one.

A small tell worth recording: the executive who went all-in on AI rarely goes looking for an honest second opinion. It tends to be the direct report — the one handed the strategy, the deadline and the fatigue — who goes looking for someone willing to say the quiet part. It is seldom the person who held the pen. Wharton's researchers describe the same shape as a "donut hole": the C-suite invests, the youngest staff adopt natively, and the middle managers actually charged with changing how the work happens are the ones left to absorb the strain.

Why no one is telling you this

You might reasonably hope the advisory industry would wander over with the bad news. It will not, and pretending the reason is mysterious would be its own small dishonesty.

Much of that industry is not, on any honest accounting, in the business of helping you use AI well. It is in the business of keeping you subscribed to the worry — and a worry, unlike a problem, has far too much commercial sense to ever get solved and stop paying.

So the urgency is delivered before the diagnosis. The framework arrives before the inquiry that might have embarrassed it. The deliverable's real purpose, beneath the handsome binding, is to establish the indispensability of the next deliverable. Everything is counted in motion — pilots launched, tools procured, roadmaps produced — because motion is what photographs, and whether the organisation actually improved is the one question the people being paid have arranged not to ask.

Beneath that obvious conflict sits a subtler and more flattering one. A firm that arrives with its methodology already loaded is not examining your situation; it is confirming its own. It has brought the answers and gone looking for a context to apply them to — which is faster than understanding, costs the same, and photographs identically. As the complexity theorist Dave Snowden puts it, the practitioner who arrives with a framework already running is not attuning to the system in front of them; they are confirming. The market rewards this, because the market, like the rest of us, would rather be reassured than informed.

Confirmation is a comfort. It is also, with some regularity, simply wrong.

And the conflict has now dispensed with even the courtesy of discretion: the makers of the models are buying, funding, and entangling themselves with the firms and infrastructure that recommend them. One need not allege a conspiracy — the arrangement is far too well-mannered for anything so energetic. The advice and the product merely share an income statement now, the way a wine critic might quietly share a cellar with the vineyard. A balance sheet exerts its gravity on everyone in range, whether or not a single soul intends it to.

What an honest look would do instead

It would examine before it prescribed. And it would be willing to conclude that you should do less.

A practice paid to act will always, somehow, find action necessary. The more useful posture is to find what is actually true and let it recommend whatever it recommends — including patience, including a smaller programme than the one already announced to the board, including, on the occasions that warrant it, the word stop.

None of this is scepticism about the technology. It is real, and frequently remarkable; the failures are seldom the model's doing. The problem is the theatre that has grown up around it — the performance of progress, sold most enthusiastically by the people you turn to for the truth.

So before the next pilot, the next platform, the next working group with transformation in its name, it is worth asking the only question that ever actually mattered:

Has any of this changed how the place works — or have we simply been very, very busy?