Written at the Wrong Altitude

Somewhere in your organisation is an AI strategy, approved by the board about eighteen months ago, that everyone has quietly stopped citing.

It wasn't wrong. Most of it still rings true. But it named models that have been superseded multiple times, priced compute at rates that no longer exist, and committed to an architecture the market has since moved past. So, the executive team is circling a familiar question: do we rewrite, or push on?

It's the wrong question. A strategy that can be made obsolete by a model release wasn't lapped by the technology. It was written at the wrong altitude.

One document, three expiry dates

Many AI strategies fail structurally before they fail substantively, and the pattern is remarkably consistent: they combine strategic intent and technical selection into a single document, approve them as a single decision and give them a single expiry date.

The board signs off on "we will lead our sector in customer intelligence" in the same breath as a named vendor, a named model family, and a unit-cost table. Twelve months later the cost table is fiction, and the model is two generations old. Here's the damage: the technical layer's expiry discredits the whole document. The intent, which was probably right, and probably still is, loses its authority because it shares a page with numbers that are now visibly wrong. The organisation concludes the strategy failed. Then it commissions a rewrite, welds the layers together again, and books the same crisis for eighteen months' time.

The rewrite-or-push-on dilemma is not a hard choice. It's a symptom of a document that should never have been one document.

Three layers, three half-lives

A durable AI strategy is three layers, deliberately separated, each with its own clock and its own owner.

1. Strategic Intent. Where AI creates competitive advantage for this business. Which value pools you're pursuing, and which you are not. Risk appetite. Build vs. buy posture. It should identify the enduring capabilities the organisation needs to develop. Horizon: three to five years. Accountability: CEO and executive team; board approval and oversight. This layer should not depend on a particular model, vendor or implementation architecture. The test is simple: if a frontier lab shipping a new model tomorrow would change a sentence in this layer, that sentence is written at the wrong altitude. Move it down.

2. The Operating Model. Who holds decision rights over AI systems that act on customers, money, or risk. Where capability lives. Which work is human-led, human-supervised, or agent-led. How the investment is funded and measured. How roles change. Horizon: twelve to twenty-four months. Accountability: nominated executive owner; executive-team governance. This layer moves on organisational time, reviewed on a rhythm, revised on triggers.

3. Technology Choices. Models, platforms, frameworks, vendors. Horizon: monthly to multi-year, depending on the component. Model versions may change monthly; enterprise platforms and enduring architecture principles may last considerably longer. Accountability: CIO, CTO or CDO under delegated architecture and risk governance. This layer should be designed to be fluid, written as design patterns, decision principles and revisit conditions, not commitments. And it does not belong in a board paper at all. A board that approves model selections is a board that has agreed to re-approve its strategy every quarter, whether it knows it or not.

The eighteen-month-old strategy on your shelf is usually eighty per cent sound at layer one, absent at layer two, and expired at layer three. That's not a rewrite. That's a decomposition: re-validate the intent, build the operating layer that was never built, and demote the technology layer to where it can churn safely.

The layer that was never built

The middle layer, the AI operating model, is where many strategies are weakest. Proofs of concept rarely spend much time on it. They test whether the technology works, whether employees will use it and whether a task can be completed faster. Those are necessary questions, but they do not establish whether the organisation can convert the capability into sustained business value.

Self-reported productivity is useful evidence, but it is not the same as realised value. Time saved only becomes valuable when the organisation changes capacity, throughput, service levels, risk outcomes or the allocation of work. Without changes to processes, roles, performance measures and funding, productivity remains distributed across individual employees rather than captured by the enterprise.

Consider an organisation introducing AI-generated summaries in a contact centre. A proof of concept may show that employees complete after-call work faster and strongly prefer the tool. But unless workforce planning, quality assurance, case routing and service targets also change, the saved time remains fragmented across thousands of interactions. The technology may have succeeded while the organisation has yet to realise a measurable improvement in capacity, service, risk or financial performance.

Organisations commonly fall into one of two traps. They either delegate AI to existing technology and governance processes, or they avoid the more difficult decisions required to embed it into the business. Neither approach is sufficient. Existing controls may manage the technology, but they rarely resolve who owns the outcome, how work should be redesigned, which decisions may be automated, when human intervention is required, how benefits will be measured or who is accountable when performance falls short.

That is the role of the AI operating model. It defines outcome ownership, decision rights, funding, delivery structures, human oversight, workforce changes, benefit measurement and the path through which successful experiments become part of normal operations. It also addresses the organisational friction early: competing priorities, displaced responsibilities, budget ownership and the politics of changing established ways of working.

This is the layer that connects strategic intent to technical delivery. Without it, the strategy remains aspirational and the technology remains experimental. With it, the organisation can make deliberate choices about where AI should operate, how value will be captured and what must change for that value to endure.

When does a layer need to change?

There is a discipline that makes the three layers durable, borrowed from an unlikely place: modern software delivery.

Modern engineering teams do not rely on an annual declaration that a system still works. They make critical assumptions visible, instrument the system and continuously test whether expected conditions still hold. Strategy deserves the same discipline. A strategy should be capable of being validated and invalidated. Its critical assumptions should be explicit, assigned to an owner and linked to evidence or indicators. We assume this value pool holds. We assume the regulator lands roughly here. We assume inference costs keep falling on this curve. We assume our competitors cannot replicate this data position within three years.

The review question then stops being, “Has a model release triggered another rewrite?” It becomes, “Which assumptions have changed, and which layer do they affect?”A model release breaks layer-three assumptions; architecture governance absorbs it without a board paper. A regulatory shift or a collapsing value pool breaks a layer-one assumption; that, and only that, triggers a genuine strategic reset.

Your strategy document should ship with its own test suite. Most ship with none. Which is why nobody can say whether the strategy failed or was simply never checked.

Enduring alignment

None of this abolishes rhythm. It assigns it. The board re-ratifies intent annually. It becomes a short conversation if the assumptions hold, a longer one if they don't. The executive team reviews the operating layer quarterly against its triggers. Architecture governance runs continuously, because its layer does.

By connecting the three layers effectively, it allows decisions to be aligned to both operational and technology delivery. This increases the opportunity to capture value previously locked within organisational rigidity in the middle layer.

One caveat. When a layer is so vague it commits to nothing and never expires. Durability is not the same as emptiness. The intent layer must still make falsifiable choices: which value pools, which businesses, what you will not do (trade-offs made and their justifications). A strategy that cannot be wrong is not a strategy. The aim is a set of choices that survive a model generation, not choices that survive because they were never really made.

So: rewrite or push on? Neither. Validate what was right, address the tough decisions in the middle, and stop asking the board to underwrite decisions with a six-month shelf life. The technology will lap your document again and again.