The Half-Life of AI Strategy

Most organisations still write AI strategy the way they wrote technology strategy a decade ago: a polished multi-year roadmap, a board-approved business case, a target operating model, and a programme to deliver it. That document is built on assumptions about what models can do, what they cost, what regulators will permit, and what competitors will copy.

The trouble is that many of those assumptions now decay in months, not years. A 36-month plan built on a 9-month assumption is not a plan. It is a forecast with a delivery schedule attached.

The failure mode is not the absence of strategy. It is false certainty. Leaders mistake the confidence of a finished document for the durability of the assumptions inside it, and they fund fixed multi-year execution against assumptions that have already started to expire. It is the strategy version of the old jibe that the slide deck outshines the thinking behind it.

The fix is not to abandon long-horizon strategy. It is to separate the part that should stay fixed, the direction, from the part that must keep moving, the assumptions, and to run the second as an adaptive system underneath the first.

The gap the data already shows

The evidence that adoption has outrun design is well established. McKinsey's 2025 global research found that 78% of organisations use AI in some form, yet only 21% have fundamentally redesigned even some workflows, and only 1% describe their deployment as mature.^[1] Gartner forecasts that more than 40% of current agentic AI projects will be cancelled by the end of 2027, citing escalating costs, unclear business value, and inadequate risk controls.^[2]

Set those numbers against the speed of the underlying capability. The Stanford AI Index found that the cost to query a model performing at the level of GPT-3.5 fell from about $20.00 per million tokens in November 2022 to about $0.07 per million tokens by October 2024, a reduction of more than 280-fold in under two years.^[3] On the SWE-bench coding benchmark, systems solved 4.4% of problems in 2023 and 71.7% in 2024.^[3] Independent analysis by Epoch AI puts the decline in inference price for a fixed performance level at a median of roughly 50-fold per year, though it varies widely by task.^[4]

Read those two sets of numbers together and the problem comes into focus. A strategic assumption written at programme kick-off, that a workflow is uneconomic to automate, or that a step needs a human in every instance, can be false before the programme has mobilised. The gap between adoption and impact is not only an operations problem. It is a problem of strategy that ages faster than the cycle built to deliver it.

This is not a case against long-horizon strategy, and it is not a call to plan in 90-day increments. Boards and regulators are right to want multi-year direction, accountability, and a stated risk appetite, and a firm that re-plans every quarter has no strategy at all.

What has changed is the shelf life of the plan beneath the direction. The detailed plan that once stayed usable for years now ages in quarters, and many organisations are still planning at the old tempo. The direction can still run to 36 months; what cannot is the fixed execution plan bolted to it.

It is a case for pointing the strategy at what AI will be able to do, not only at what it can do today. Capability is compounding rather than improving in a straight line, as the cost and benchmark curves above already show. A direction anchored to today's frontier is obsolete before the programme mobilises, while a direction aimed at where capability is heading stays usable as the frontier moves.

Strategy Half-Life

Strategy Half-Life is the period over which a strategic assumption stays true before AI capability, falling cost, regulation, or competitive diffusion erodes it. It is not a property of the strategy as a whole. It is a property of each assumption the strategy rests on, and different assumptions carry radically different half-lives.

The term borrows from radioactive decay deliberately. A half-life is not a prediction that an assumption is wrong; it is a measure of how quickly the probability that it still holds is falling.

Some assumptions in an AI strategy have half-lives measured in years, such as the firm's regulatory obligations or its customers' fundamental needs. Others have half-lives measured in months, such as the cost of a given inference task or the boundary of what a model can reliably do. The error is committing fixed, multi-year execution to an assumption with a short half-life and never checking the clock.

Four forces drive the decay, and each moves at its own pace. The first is model-capability cadence: the frontier of what models can do reliably is moving fast enough that “AI cannot do this yet” is among the shortest-lived assumptions a strategy can hold, as the SWE-bench jump shows.^[3] The second is the falling cost curve: an automation that is uneconomic today can cross into economic within a single budget cycle, as the inference-cost collapse demonstrates.^[3][4]

The third is regulatory motion: rules that constrain or permit a use case shift on their own timetable, sometimes slower than capability and sometimes faster. The fourth is competitive diffusion: an advantage built on a capability competitors can procure in weeks has a shorter half-life than one built on an operating discipline they would need years to replicate.

How false certainty fails

False certainty rarely looks like a bad strategy. It looks like a good one, finished and confident, with the decay hidden inside it. Seven signals tell you a strategy has committed to assumptions with a shorter half-life than the plan around them.

Zillow shows what happens when a short half-life assumption is scaled as if it were durable. The company built an algorithmic home-buying operation that depended on forecasting near-term house prices accurately enough to buy, lightly renovate, and resell at a margin at volume. In the third quarter of 2021 it bought 9,680 homes and sold only 3,032, and in November 2021 it wound the business down with a write-down of about $569m and cut roughly 25% of its workforce.^[5]

The chief executive's own explanation was that “the unpredictability in forecasting home prices far exceeds what we anticipated.”^[5] The intent was sound; the forecasting assumption had a short half-life and was scaled before it was re-validated against changed market conditions, so scale amplified the error instead of averaging it out.

IBM Watson Health shows the same failure on a longer horizon. Launched in 2015 as a healthcare moonshot, it grew through more than $4bn of acquisitions on a single fixed thesis: that one platform could read medical evidence and recommend treatment at scale.^[6] The thesis was committed as a multi-year plan and an acquisition spree rather than a portfolio of revisable bets, with no cadence that forced it to be re-tested as the data, the clinical workflow, and the regulatory reality diverged from the original bet.

IBM sold the Watson Health assets to Francisco Partners in a deal announced in January 2022.^[6] A long-horizon AI strategy with no mechanism to re-test its central assumption is not patient conviction. It is a wager held past the point the evidence turned.

Directional strategy, adaptive execution

The posture that answers the half-life problem is directional strategy with adaptive execution. The direction is fixed and patient; the execution is a portfolio of revisable bets governed by evidence. This is not a new idea in management research, which is precisely why it is durable enough to build on.

Seven shifts separate it from the conventional planning posture.

The mechanism behind the right-hand column is well established. Rita McGrath and Ian MacMillan's work on discovery-driven planning argues that in high-uncertainty arenas a plan should not be judged by how close outcomes come to the original forecast, because the parameters change as information arrives; instead, capital is released against checkpoints as assumptions are validated, to “learn as much as possible as cheaply as possible.”^[7] Timothy Luehrman's work on strategy as a portfolio of real options frames a strategy as “more like a series of options than a series of static cash flows,” where some moves are made now and others deliberately deferred so the choice can be optimised as conditions evolve.^[8] Rita McGrath's later work on transient advantage makes the underlying claim explicit: advantages, and the assumptions beneath them, expire faster than they used to, so disciplined exit becomes a routine rather than a failure.^[9]

There is a serious objection here, and it is loudest, and most legitimate, in regulated financial services: you cannot run a bank on 90-day bets, because boards and supervisors require committed, multi-year, board-approved plans. The objection is correct about what regulators want and wrong about what threatens it. The G20/OECD Principles of Corporate Governance hold the board responsible for “reviewing and guiding corporate strategy, major plans of action, risk policy” and for “monitoring implementation and corporate performance,” which is a description of direction, accountability, and continuous re-testing, not of frozen execution assumptions.^[10]

The machinery a regulated firm already runs supports this rather than resisting it. Three lines of defence, model-risk governance, and change control exist to test continuously whether a control still holds as conditions move. A directional strategy with a dated assumption register gives those functions what most AI programmes deny them: an explicit, owned statement of what the plan is betting on and when each bet falls due for challenge.

A 36-month directional intent supported by a dated register of assumptions and a regular cadence of review satisfies the substance of board oversight better than a static roadmap, because it gives the board live evidence of which assumptions still hold rather than an annual restatement of a plan that may have silently expired. Real options and discovery-driven funding reduce risk; they retire bad assumptions cheaply before they harden into expensive committed programmes.

Directional strategy supplies exactly the multi-year direction and accountability that boards and supervisors want, while adaptive execution keeps those commitments deliverable as conditions change. The fixed part is real and named. What flexes is the bets and the assumptions, not the direction.

The five components of a strategy that adapts

A strategy that adapts has five components. Each is a concrete artefact or mechanism, not a posture, and together they form the system that runs underneath the fixed direction.

The first is strategic intent: the durable direction that does not move. It names where the firm's cognitive transformation is heading, the way it intends to decide, remember, and coordinate once AI is woven through the work. It is the small set of commitments that survives whichever way the four decay forces turn, written as a few simple rules rather than a thick plan, in the spirit of Kathleen Eisenhardt and Donald Sull's finding that advantage in high-velocity markets comes from a few simple rules and key processes rather than elaborate position.^[11] Intent has the longest half-life in the system, and it is the only part the firm refuses to revise on a quarterly clock.

The second is a portfolio of bets, funded as options rather than programmes. Each bet receives enough capital to test its central assumption, not enough to deliver the whole vision, with further capital released only as the assumption is validated.^[7][8] This is the structural opposite of the single committed wager that sank Watson Health.

The third is an adaptive cadence: a regular loop, set here at 90 days, that reviews assumptions rather than milestones. A milestone review asks whether the work is on schedule. An assumption review asks whether the thing the schedule was built on is still true. The cadence is owned by the people who set the direction, which is what keeps it a strategy mechanism rather than a delivery ritual.

The fourth is an assumption register: the named artefact that lists each strategic assumption, dates it, estimates its half-life against the four decay forces, names an owner, and records the evidence that would prove it false. It is the diagnostic the Strategy Half-Life concept makes possible, and it is the single artefact most AI strategies lack. Without it, decay is invisible until it is expensive.

A register row is deliberately plain. It states the assumption, the force most likely to erode it, a first estimate of its half-life, and the evidence that would prove it wrong.

Three rows of an assumption register. The value is not the estimate; it is the argument the estimate forces.

The fifth is a scale-or-kill mechanism: the decision point that converts a validated bet into committed execution or shuts it down. When the assumption holds, the bet scales into Intelligent Ops, where the workflow is redesigned around human-agent collaboration and run as a production capability, and what that system learns flows back into the assumption register so execution keeps testing the strategy. When the assumption breaks, the bet is killed before it hardens into a committed programme; in a regulated setting that kill is a decommissioning event with audit, remediation, and notification consequences, which is why the kill criteria are written before the bet scales. Killing on pre-agreed evidence is not failure of the system; it is the system working.

Amazon shows the portfolio and the kill discipline working. The Fire Phone launched in mid-2014, took a write-down of about $170m disclosed that October, and was discontinued within roughly a year.^[12]What matters is what held and what moved: the strategic intent, to own a place in the customer's home, stayed fixed, while the specific bet, a phone, was killed quickly on evidence, and the team and its capabilities were redeployed into the Alexa and Echo line that succeeded. The example is from 2014, which is the point; the kill discipline is older than the current AI cycle, and it travels.

Klarna shows the corrective behaviour the thesis prescribes, mid-course and in public. After an aggressive push into AI-driven automation that cut headcount and vendor relationships, chief executive Sebastian Siemiatkowski told Reuters ahead of the firm's September 2025 listing that it had “over indexed a little bit” on cost-cutting and was working to “course correct” back towards service quality.^[13]

The assumption that automation could substitute broadly for human service had a shorter half-life than the strategy assumed, and the firm re-tested and adjusted rather than holding the line. That is recoverable. Zillow and Watson Health, with no cadence to force the re-test, were not.

GM Cruise shows why the scale-or-kill mechanism must gate scaling, not just permit a late exit. Cruise scaled driverless robotaxi operations while the edge-case safety assumptions were still maturing, and in October 2023 a vehicle struck and dragged a pedestrian who had been thrown into its path, after which regulators halted operations and the fleet was grounded.^[14] In December 2024 GM announced it would stop funding Cruise's robotaxi development.^[14] The sequence, scale, then incident, then halt, then kill, is the exact failure the scale-or-kill discipline exists to prevent: the bet was scaled into production before its assumption was validated, so the kill arrived only after the human, regulatory, and reputational cost had been incurred.

Centralise the guardrails, decentralise the bets

The five components raise a real tension in how to organise. Centralise too much and AI transformation slows to the pace of a single overloaded strategy team that cannot hold the domain knowledge of the whole firm. Decentralise too much and it fragments into shadow AI, duplicated bets, and no shared learning. Neither extreme runs an adaptive system; one cannot move and the other cannot remember.

The resolution is a federated operating model that splits the two cleanly. The centre owns the parts that must be common: the strategic intent, the assumption register, the adaptive cadence, and the governance guardrails. The edges own the bets, placed where the work and the domain knowledge actually sit. The centre sets a small number of durable rules and the firm-wide learning loop; the edges capture the local opportunities the centre could never see in time.

Governance in this model is guardrails, not gates, and that is where it answers the second objection: that this is all just agile under a new name. Agile manages delivery; this manages strategic assumptions, and the difference shows in the governance layer. The discipline it needs is human-on-the-loop oversight: people monitoring live systems and intervening on exception, rather than approving every output in advance. That means continuous monitoring, structured feedback, and a mechanism to disengage a system that drifts from its intended use.^[15]

A gate checks a bet once, at launch, and then looks away. A guardrail monitors a live bet continuously and feeds what it learns back into the assumption register, which is exactly the loop that approval-only governance lacks. The artefact, the assumption register reviewed at board altitude, is what makes this a strategy method rather than a sprint cadence with better branding.

The regulated firm has the strongest reason to adopt this shape, not the weakest. A federated model gives supervisors what they ask for, common guardrails, a documented risk posture, and continuous monitoring, while letting the bets move at the pace the domain demands. The centre holds the line that protects the institution; the edges learn fast enough to keep the institution competitive.

The leadership agenda

The work of a board, chief executive, chief technology officer, or chief transformation officer is not a programme to launch. It is a change in how leadership holds the strategy, from authoring a plan and defending it to running a system and interrogating it. The board's job stops being the annual approval of a roadmap and becomes the standing question of which assumptions still hold.

That begins with a harder honesty about what the firm actually knows. A 36-month direction is a statement of conviction; a 36-month plan built on six-month assumptions is false certainty dressed as conviction. The task is to hold the first boldly and treat the second with suspicion, and most operating rhythms are built to do the reverse.

The most revealing move a leadership team can make is also the cheapest. Write down the assumptions the strategy is betting on, date them, and ask how long each survives contact with falling cost and rising capability. The discomfort that follows is the point, because an assumption no one will date is usually one no one wants to test.

The competitor can buy the same models in weeks and copy the same demo in a quarter. What they cannot copy quickly is two years of a working adaptive system: a dated assumption register, a disciplined cadence, a portfolio with real kill criteria, and the institutional habit of changing its mind on evidence. That system is the durable advantage, and it compounds, because every bet retired cheaply and every bet scaled into Intelligent Ops makes the next decision faster and better grounded than the last.

The question for the board is not whether the AI strategy is good. It is sharper than that. Which assumption in your current AI strategy has the shortest half-life, who owns it, and what is the evidence that would tell you it has already expired? If no one in the room can answer, the strategy is already running on a clock no one is watching.