The Business Case for Enterprise AI

A business case is usually written as though its only job is to be correct. Its real job is to still be persuasive after the person who wrote it has left, the CFO has been replaced, and the quarter has gone badly.

Eighteen months after approval, the AI programme comes up in a portfolio review. Somebody pulls the original case out of the shared drive, and it opens onto a title slide reading building the foundation for an AI-first enterprise, followed by a section on competitive positioning, a maturity curve with a dot showing where the company sits today, and a benefits page whose assumptions are laid out with real care. It is a good document, persuasive enough to have won funding against several other things that year. The trouble is that the director who wrote it moved divisions nine months ago, the executive who sponsored it has left the company, and the finance lead now running the review has no memory of the approval meeting and no personal stake in whether it was a good idea. She reads it the way a stranger reads anything, looking for the sentence that tells her why this belongs in next year's plan, and what she finds is a case for a technology, written by people no longer in the room, at a moment when the technology no longer feels new. The programme does not get killed for failing. It gets killed for being unclaimed.

This happens constantly, and it is almost never diagnosed correctly afterwards. The postmortem talks about value not materialising fast enough, or the pilot not scaling, or the vendor underdelivering. What actually happened is more structural. The business case was treated as an analytical artefact, a thing whose quality is a function of how sound its reasoning is, when in practice it functions as a political instrument whose quality is a function of how well it survives being read by people who owe it nothing. Those are different design goals, and optimising for the first while ignoring the second is how well-reasoned programmes die.

Every case eventually meets a reader who owes it nothing

Approval is not the finish line for a business case. It is the first of many readings, and by far the friendliest, because at the moment of approval the author is in the room, the sponsor is spending their own credibility on it, and everyone present has just agreed on a shared story about the future. Every subsequent reading happens under worse conditions: a budget cycle where something has to be cut, a reorganisation that moves the programme under a leader who inherited it rather than chose it, a bad quarter in which the question is not whether the thing is good but whether it is more defensible than the four other things competing for the same money. By the time a multi-year enterprise programme reaches steady state, the case has typically been reread by people who had nothing to do with writing it more often than by anyone who did.

The uncomfortable implication is that the persuasive force of the original document is largely non-transferable. A sponsor's conviction is not stored in the deck; it lives in the sponsor, and it leaves when they do. The private context that made the numbers feel reasonable — the operational anecdote everyone in that meeting had lived through, the frustration with a process that had failed twice before, the sense that this particular team could pull it off — none of it is written down, because at the time it did not need to be. What is written down is the part that felt objective: the benefit model, the roadmap, the technology rationale. The technology rationale is exactly the part that ages worst.

The aggregate consequence shows up in how these programmes end. Gartner has predicted that more than forty percent of agentic AI projects will be cancelled by the end of 2027, naming escalating costs, inadequate risk controls, and unclear business value among the causes. That third phrase is usually read as an accusation about analytical rigour, as though sponsors had failed to do arithmetic. It is worth reading instead as a statement about legibility over time. Value that was clear to the person who wrote the case, and to the executives who approved it, and never made legible to anyone who arrived afterwards, is functionally the same as no value at all when the review comes around, because the programme cannot defend itself and nobody's reputation is served by defending it.

Novelty is a claim with an expiry date; an unsolved problem is not

The most common structural flaw in an AI business case is that its load-bearing premise is the strategic significance of the technology itself. Cases in this shape argue that the organisation cannot afford to fall behind, that the capability will be foundational, that competitors are moving, that the board expects a position. Every one of those claims is true at the moment it is written, and every one has a shelf life, because each depends on the technology continuing to feel new. Novelty decays in two directions and both are fatal. If the technology becomes ordinary, the urgency evaporates without anyone needing to argue against it — nobody writes a strategic case for having email. If the wave disappoints, the case collapses faster still, having borrowed its credibility from an enthusiasm that has since become scepticism. Worse, it has no defender once the sponsor is gone, since the thing it addressed was never anyone's problem in the first place. It was an opportunity, and opportunities do not have owners.

Contrast that with a case whose premise is a problem the organisation was already failing to solve before anyone mentioned AI. The claims backlog that has been over capacity for three years, the compliance obligation the business keeps missing by a margin small enough to survive and large enough to appear in every audit, the service level that holds fine at current volume and breaks predictably every time the business grows, the reconciliation process that two previous programmes tried to fix and did not. Problems like these have three properties that make a case durable in a way that no amount of analytical polish can substitute for. They have a named owner who existed before the programme and will exist after it, so the case is never orphaned. They have a history of failed attempts, which means the organisation has already conceded that the problem is real and hard. And they are described in the business's own vocabulary, so a reader who thinks AI is overhyped can still follow the argument from beginning to end without having to accept a single claim about technology.

That third property yields the sharpest test available. Take the case and mentally delete every mention of AI, agents, models, and platforms; if what remains is a coherent account of something the organisation needs fixed, with an owner and a history, the technology is doing what it should in the document, serving as the mechanism rather than the premise. If what remains is an empty frame, the case was about the technology all along, and it will not survive a reader indifferent to the technology.

Write it to be inherited, not to be approved

The practical shift this implies is that the case should be drafted for its third reader rather than its first. The first reader shares your context and can be persuaded in conversation; the third has only the document, and writing for that person changes what goes into it. It means stating the problem in the terms the business already uses to complain about it, including the internal name people give it, so that the case is searchable by the vocabulary of the problem rather than the vocabulary of the solution. It means recording why previous attempts failed, not defensively but as institutional memory, because a later reader trying to decide whether a programme is struggling or merely mid-course has no other way to tell the difference. It means naming, explicitly, who owns the underlying problem independent of the programme — the person whose job gets harder if the work stops — because that name is the closest thing a business case has to a successor clause.

It also changes how the technology is described. A case that commits itself to a category — that the organisation will adopt autonomous AI workers, or standardise on a particular platform — has tied its survival to the continued fashionability of a phrase, and phrases in this field turn over quickly. A case that commits itself to a business outcome and treats the platform as the current best mechanism can absorb a change of vendor, architecture, or terminology without being rewritten. This is not an argument for vagueness about implementation; it is an argument about which sentence carries the weight. When a system like StudioX appears in a durable case, it belongs where mechanisms belong — how the backlog gets worked, how the compliance gap gets closed, how the coordination nobody has hours for finally gets done — rather than in the premise, where its presence quietly turns the whole document into a bet on a name. The body of analysis now accumulating around autonomy as an operating model for the enterprise is useful in the same way: as evidence about mechanism, not as a substitute for having a problem worth solving.

None of this makes a bad idea good. A case tied to a real, owned, previously unsolved problem can still be wrong about whether this particular approach will solve it, and durable is not the same as right. What it changes is the failure mode. The novelty case fails silently and permanently, dropped in a budget cycle by people who never understood it, and nothing about the organisation changes as a result because nothing was really at stake. The problem case fails loudly and recoverably: the problem is still there, still owned, still embarrassing, and the organisation tries again with a better mechanism.

So the mental model worth carrying is that a business case is neither a pitch nor an analysis. It is a bequest, a document that will be executed by people you have not met, in conditions you cannot forecast, at a moment when your own enthusiasm is unavailable as evidence. Everything in it that depends on your presence in the room is decoration. What remains — a problem the business already admits it has, a person whose life is worse if it stays unsolved, and an honest account of why it has resisted every previous attempt — is the only part that will still be doing work three readings from now, long after everyone who argued for it has moved on.