The AI Investment Frontier

The cost of model capability is collapsing while the cost of deploying it badly is not. Procurement that asks “what can this model do?” produces a portfolio organized around novelty; the question that produces returns is “where does an autonomous system change a number we already manage, reliably enough to trust?” The two questions select almost disjoint sets of projects. This paper is a framework for asking the second one rigorously.

Value density is economic value per task multiplied by task volume — a high-stakes monthly task and a low-stakes million-times-a-day task can score alike, and both beat a medium task done rarely.

The reliability ceiling is the maximum dependable per-step success a task admits given its tolerance for error and the verifiability of its outputs. A task whose correctness cannot be checked has a low ceiling regardless of model strength — and, as the companion engineering paper shows, end-to-end feasibility decays sharply with unverified steps. The ceiling, not the model, is usually the binding constraint.

Defensibility asks whether the resulting capability compounds — through proprietary process data, accumulated evaluations, and operating IP — or commoditizes into something a competitor buys off the shelf next quarter. Value that does not compound is rented, not owned.

The three axes resolve into four actions. High value, high ceiling → build and operate: the durable core, where ownership of the operating discipline is the moat. Low value, high ceiling → buy: it is automatable but commodity; rent it. High value, low ceiling → assist, do not automate: keep a human accountable and use AI to augment, not to act unattended. Low value, low ceiling → decline: the most valuable output of a strategy review, because it redirects budget from theatre to return.

A portfolio is sequenced, not chosen all at once. The first build should be high on the frontier and instrumented to produce the process data and evaluations that raise the reliability ceiling of the next candidate. Early wins are selected as much for what they teach the operation as for their standalone return — the flywheel between operating data and future feasibility is the real compounding asset, and it is why the second project is cheaper and safer than the first.

For the board and CFO, this framework converts “we should use AI” into a defensible capital-allocation decision with an explicit thesis per line item. The discipline it enforces is unglamorous and decisive: fund where value density meets a reachable reliability ceiling, buy the commodity, keep a human on the brittle-but-valuable, and decline the rest out loud. The competitor who funds the frontier and declines the theatre will compound; the one who funds capability will have a museum of pilots.

The frontier is a decision aid, not a calculator; the three axes require judgment to score and that judgment improves with operating data the organization may not yet have. Reliability ceilings are not static — better verification raises them — so a declined candidate may re-qualify later. The framework also assumes value can be attributed to a managed number; genuinely novel, non-attributable bets sit outside it and should be funded, if at all, as research rather than operations.

Strategy is allocation, and AI strategy is the discipline of allocating against value and feasibility rather than capability and fashion. AI strategy ends where the bill begins — and it begins with the honesty to name where agentic AI will pay, and where it will not.