Enterprise AI spend is accelerating faster than the discipline used to govern it. CEOs are pushing to scale AI savings and growth inside 18-month horizons, and AI is now claiming a rising share of IT budgets year on year. Yet by most credible estimates only around a quarter of AI initiatives have delivered the return that was promised when they were approved.
That gap is usually explained as a technology problem, or a talent problem, or a data problem. It is none of those. It is a scoring problem. The way most boards are asked to evaluate an AI investment produces a number that looks decisive, travels well in a slide, and does not survive contact with reality. Three years later, when an audit committee asks why a given investment was approved, the number on the original board pack will not have an answer.
Here are the four structural failures that produce it.
Capability is scored in isolation
The dominant question in most AI business cases is can the technology do this? — and the scoring stops there. Models are demonstrably capable, so the answer is almost always "yes, mostly," and the case advances on that strength.
But technical feasibility is only one of the conditions an activity has to satisfy. Whether the workforce can actually adopt the output, and whether the activity's risk profile permits automation at all, are scored loosely if they are scored at all. A 90% feasibility score for invoice classification is irrelevant if the team that has to act on the output trusts it at 45%. A high feasibility score for any decision that carries material liability is irrelevant if a single error is unacceptable. Capability scored alone always flatters the case, because capability alone is the one dimension AI reliably wins.
The composite is averaged, hiding the binding constraint
Once an activity is scored on several dimensions, the scores are typically averaged into a single composite. Averaging is the quiet killer.
Consider an activity that scores 0.90 on technical feasibility, 0.45 on workforce readiness, and 0.70 on risk. Averaged, that reads as roughly 0.68 — comfortably inside "proceed" territory. But these dimensions are not independent contributors that compensate for one another; they are dependencies that gate one another. If the workforce cannot adopt the output, the feasibility of the model is moot. The binding constraint is the 0.45, and the averaged composite is engineered to hide exactly that. The board sees a number that says go. The constraint that will determine the outcome has been averaged out of view.
Scoring happens at the role level, not the activity level
Most methodologies assess AI readiness at the level of a role or a function — "the finance function is 60% automatable," "claims handling is a strong candidate." This is the wrong unit of analysis.
A role is a bundle of very different activities. Within a single finance role, invoice matching, exception investigation, and judgement-based vendor negotiation have completely different feasibility, readiness, and risk profiles. Scoring the role produces an average of averages, and decisions get made on a blended number that describes none of the actual work. Investment, adoption, and risk all live at the activity level. A score that does not resolve to that level cannot be defended when the activity it was supposed to describe behaves nothing like the role it was rolled up into.
The cost basis assumes a pricing model that is disappearing
Most AI business cases are built on a per-seat, SaaS-style cost assumption: a fixed licence per user, predictable and easy to model. The economics of enterprise AI are moving away from that. Increasingly, cost is metered — driven by consumption, by the orchestration pattern, by the depth of reasoning a task demands, and by the human oversight required to keep the output safe.
A case built on a flat per-seat figure can understate true running cost by a wide margin once an activity scales, because none of the real cost drivers are per-seat. The ROI line that looked clean at approval inverts in production — and unit economics at projected scale, the dimension that would have caught it, was never on the page.
Why this surfaces in the audit, not the approval
Each of these failures is invisible at the moment of decision and unavoidable in hindsight. Capability scored alone looks rigorous. An averaged composite looks balanced. A role-level score looks strategic. A per-seat cost line looks conservative. Every one of them reads as competent governance in the boardroom — and every one of them is the reason an audit committee, three years on, will not be able to reconstruct why the investment was approved.
The fix is not more sophistication. It is a different shape of number. A board pack should:
- Show the binding constraint explicitly rather than averaging it away
- Score at the activity level where investment and risk actually sit
- Treat the dimensions as dependencies that gate one another, not inputs that average out
- Carry a consumption-based cost basis that survives scale
A score built that way is not just more accurate — it is auditable, which is the property the current number lacks.