Same Deck, New Nouns: Why Most AI Strategies Are Cloud Strategies Relabelled
A strategy that counts seats will always mistake distribution for change.
The meeting where the deck looked familiar
There is a particular silence that settles over a steering committee when someone in the room recognises the slides. The programme director is presenting the organisation’s artificial intelligence strategy. Five workstreams: a platform at the centre, a capability-building track down one side, a governance function down the other, an operating-model workstream underneath, and a prioritised pipeline of use cases feeding the middle. It is a competent deck. It is also, almost line for line, the cloud strategy the same organisation adopted a few years earlier. The workstream names have changed. The architecture of the argument has not. Somewhere in that room, more often than not, sits the person who ran the earlier programme, wearing the expression of someone hearing a song they cannot quite place.
I have sat in that room more than once, and the recurrence is now consistent enough that I have stopped filing it under coincidence. Most of what organisations are currently calling an AI strategy is a cloud strategy with the nouns replaced. That is not, by itself, a scandal — reaching for a template that worked is usually a sign of maturity, not laziness. The difficulty is subtler and more expensive than plagiarism. The parts of the new situation that refuse to fit the old template are precisely the parts that carry the value and the risk, and a recycled strategy is, by its very construction, blind to exactly those parts.
The tell
You can recognise a recycled strategy by its silhouette before you read a word of the content. It centres on a platform, because the last decade taught us that the platform is the thing you build first and defend hardest. It carries a “centre of excellence”, because that is how we housed scarce cloud skills and it is how we now propose to house scarce model skills. It has an adoption curve, an enablement track, and the word “democratise” somewhere on slide four. It frames the destination as migration: we are here, the future is there, and the strategy is the set of workstreams that moves the organisation across the gap.
Laid side by side, the vocabulary substitution is almost mechanical.
| The cloud strategy said | The AI strategy says |
|---|---|
| Migrate the estate to the platform | Adopt the models across the enterprise |
| Cloud centre of excellence | AI centre of excellence |
| Lift and shift, then optimise | Pilot and scale, then embed |
| Consumption and FinOps discipline | Token cost and inference discipline |
| Upskill engineers on the new stack | Upskill everyone on prompting |
| A pipeline of workloads to migrate | A pipeline of use cases to deliver |
| Landing zones and guardrails | Governance frameworks and guardrails |
The word “guardrails” survives the translation entirely intact, which ought to tell us something. When the language of a new strategy can be produced by find-and-replace on the old one, the thinking has usually travelled the same route.
Why the template keeps winning
It would be comfortable to attribute this to laziness, or to consultants selling last year’s methodology at this year’s rates. That is too easy, and it mistakes the mechanism. The recycling is structural, and understanding why it happens is the only way to stop it.
Consider who writes the strategy. It is written, in most large organisations, by the same constituency that wrote the cloud one — the enterprise architects, the transformation office, the integrator on the framework agreement. These are capable people, and they are reaching, honestly, for the most relevant precedent they have. In their lived experience the most relevant precedent is the cloud programme: another organisation-wide technology shift, another platform, another adoption problem. Pattern-matching to it is not negligence; it is expertise doing what expertise does.
Consider, too, the machinery the strategy has to pass through. A large organisation can only metabolise a technology initiative in a small number of shapes. The business-case template wants a capital line, a benefits curve, and a payback period. The governance apparatus knows how to review a platform build and a set of adoption targets. Procurement knows how to buy capacity and licences. Confronted with something genuinely unfamiliar, this machinery does not pause; it reshapes the unfamiliar thing into the nearest shape it already knows how to process. The strategy comes out looking like a cloud programme partly because a cloud programme is the shape the organisation’s own digestive system imposes.
And consider the incentive on the person at the front of the room. Declaring a strategy is a demonstration of control. In a moment when boards are anxious and competitors are making announcements, a five-workstream deck that resembles the last successful programme is enormously reassuring — to its author and its audience alike. It says: this is under control, we have done something like this before, here is the map. The recycled strategy is, at bottom, a way of appearing to have decided.
The cloud template is not chosen because it fits. It is chosen because it is the shape an anxious organisation already knows how to hold — and it converts a genuinely open question into a programme that looks answered.
That is the real function the template performs. It is a defence against ambiguity. When we are not yet sure what the right questions are, borrowing a structure that supplied good answers before lets us feel we have addressed the problem without having to sit in the discomfort of not knowing. It is precisely because the template is reassuring that it is dangerous.
The number that gives it away
Abstractions are easy to argue with, so let me be concrete about how the pattern shows up in the data an organisation actually collects about itself.
A programme of this kind will typically report progress in the currency of provisioning. It will tell you that twelve thousand assistant licences have been deployed, that forty use cases sit in the pipeline, that the platform is live in three regions. These are real numbers and they are all, quietly, measures of supply. Ask a different question — how many of those twelve thousand people used the tool in the last week, and changed a decision or a piece of work because of it — and the figure collapses, often to fewer than one seat in ten. Ask how many of the forty pipeline use cases are in production and materially altering a cost line eighteen months on, and the answer is frequently three.
Anyone who lived through the migration years will feel the déjà vu in the stomach. This is the lift-and-shift number wearing new clothes. We moved the workloads to the platform and reported the estate as migrated, and only later discovered that migrating the workload and changing the economics were two different projects, and that we had budgeted for the first while promising the board the second. The gap between provisioned and adopted, between moved and modernised, is the single most reliable finding of the cloud era. The recycled AI strategy reproduces that gap faithfully, because it inherited the metric that hides it. A strategy that counts seats will always mistake distribution for change.
What is actually different — and why that is the whole point
Here the honest objection deserves its strongest form, not a caricature. A capable reader will say: you have proven too much. AI is not cloud. It is a genuine step change — non-deterministic where cloud was deterministic, a system that produces judgement rather than merely storing and serving data, and now, with agents that take actions rather than return answers, something that reaches out and does things in the world. Of course it needs its own strategy. Treating it as just another platform is the error, not treating it as a new one.
That objection is correct in every particular, and it is the reason the recycling matters rather than the reason it does not. Everything genuinely new about this technology sits exactly where the cloud template is silent. A cloud programme’s hardest questions were about capacity, cost, and migration — how much, how expensive, how to move it safely. Those questions have answers, and the template is very good at pursuing them. But a system whose output is non-deterministic raises a question the migration playbook never had to ask: not did it move but is it right, and how would we even know. Evaluation — the discipline of measuring whether a probabilistic system is doing what you need, continuously, as the models and the world underneath it shift — has no ancestor in the cloud deck. Neither does the question of delegation: a platform stores what you give it, but an agent acts, and deciding which decisions an organisation is willing to hand to a system that is usually-but-not-always right is a governance problem of a completely different kind from provisioning a landing zone.
“A cloud programme asked whether the workload had moved. An AI programme has to ask whether the judgement should be trusted — and the old template has no box for that question.”
So both things are true at once, and holding them together is the entire task. It is a recycled strategy, and the recycling is the failure — not because borrowing is wrong, but because the new substance will not fit the old container, and the places where it spills over the edges are the places that decide whether the whole endeavour is worth anything. The template optimises attention onto the risks we already know how to manage and directs it away from the risks that are actually new. It is not merely unhelpful. It is actively misdirecting.
What a strategy that had not been written before would foreground
If the test of a real AI strategy is that it could not have been produced by editing the cloud one, then a handful of things have to move from the footnotes to the front page.
- Evaluation as a first-class workstream, not a governance sub-bullet. The organisation needs a standing capability to measure whether its models and agents are producing acceptable outputs, against defined criteria, and to keep measuring as everything drifts. In a deterministic world this was testing, done once. Here it is continuous, and it is central. If evaluation is not one of the largest workstreams on the page, the strategy has not understood what it is dealing with.
- A language of delegation, not adoption. The useful question is not how many people have access, but which decisions and actions the organisation is prepared to let a system take on its own, under what supervision, with what ability to intervene. “Adoption” counts seats. “Delegation” forces the leadership conversation that actually matters, and it scales the trust deliberately rather than declaring it by licence count.
- An operating model built for things that are usually right. Deterministic systems fail loudly and rarely; these systems fail quietly and plausibly. That inverts a great deal of received wisdom about controls, escalation, and where a human belongs in the loop. None of it can be inherited from the platform playbook.
- A metric of changed decisions, not provisioned capacity. Retire the seat count as the headline. Report, instead, the number of decisions, workflows, or cost lines that are demonstrably different because of the technology. It is a harder number to gather and a far more honest one, and it is the only figure that would have caught the lift-and-shift illusion the first time.
None of these is exotic. What is striking is how completely absent they are from the recycled deck — not because anyone argued against them, but because the template that produced the deck had no slot for them, and a strategy tends to contain only what its template has a slot for.
The comfort worth giving up
The recycled strategy endures because it is genuinely comforting. It lets a leadership team feel it has decided at the precise moment when the most valuable thing it could do is admit what remains undecided — how far to trust a system that reasons, which judgements to hand over, how to know when it has quietly begun to fail. Those are uncomfortable questions, and the cloud template’s great gift is that it lets us skip them and still walk out of the room with a plan.
We should be more suspicious of that comfort than we are. When a new strategy can be generated by search-and-replace on the last one, the honest reading is not that we have found a durable pattern. It is that we have not yet done the thinking the new situation demands, and have reached for the old map to cover the fact that the territory has changed. The remedy is not to abandon everything the migration years taught us; much of it — the discipline about hidden costs, the scepticism about adoption theatre, the insistence on realised benefits — transfers and should be defended. The remedy is to notice where the map stops describing the ground, and to have the composure to stand in that gap and ask a question we do not yet have a workstream for. That composure, not the platform, is the thing an AI strategy is really made of.