When AI Fails Through Unbounded Cost

The largest AI cost failures of the current cycle were not failures of capability. The models worked. The systems shipped. What failed was discipline. Across five documented cases spanning retail, ride-hailing, enterprise software, and venture-backed automation, the same breakdown recurs. No one bounded the cost against the value, no one bounded the time-to-value against a deadline, and no one set a stopping rule. Organizations optimized capability while leaving cost unbounded, and the stopping rule was supplied, eventually, by an exhausted budget, a balance sheet, a fatality, or an insolvency rather than by management.

This paper makes that argument visible. It reads each failure through Herbert Simon's bounded rationality, the idea, settled in 1947 and recognized with the Nobel Prize in 1978, that rational decision makers do not maximize. They satisfice. They search until they find an option that is good enough against a criterion set in advance, and then they stop. The five cases are organizations that behaved as if they were the perfectly rational economic man who could afford an unbounded search. They could not, and the bill came due.

The economic frame Simon settled in 1947

The paper rests on two of Simon's ideas, and both are about cost. The first is the stopping rule. Any rational search must specify, in advance, the condition under which it ends. A search with no stopping rule is not thorough. It is runaway. The second is the economics of search itself. Satisficing is not a concession to human weakness. It is the correct response to the fact that the marginal cost of continued optimization eventually exceeds its marginal value. Past that point, the maximizer is not more rational than the satisficer. He is less rational, because he is spending more to gain less. There is a particular irony here. Simon was not only the economist who gave us bounded rationality. He co-created the Logic Theorist with Allen Newell in 1956 and helped invent the field. The man who taught us to bound our spending helped build the technology on which we have most spectacularly failed to bound it.

The five cases

1. Starbucks NomadGo, an AI inventory tool scaled to more than 11,000 stores on a claim of 99 percent accuracy and eightfold speed, retired nine months later because verifying every output doubled the task and erased the efficiency it promised.
2. Uber, two eras, a firm that failed to bound AI cost twice a decade apart. It carried more than 10 billion dollars of exposure to autonomous vehicles before selling the unit, and in 2026 it burned its entire annual AI budget on coding tools in four months.
3. Microsoft, vendor and buyer, which priced GitHub Copilot below its marginal cost, losing an average of 20 dollars and up to 80 dollars per user per month on a 10 dollar fee, and rebuilt the pricing to token-based billing to impose the bound the flat fee lacked.
4. Builder.ai, valued at roughly 1.3 billion dollars, which collapsed into insolvency in 2025 after its AI was revealed to be largely human engineers, the vendor-side proof that an AI claim concealing a human cost base never had unit economics that could close.
5. Amazon Just Walk Out, withdrawn from full-size grocery stores after the autonomous checkout carried a hidden human-review cost base that never reached the structure required to replace the cashier.

These are not anomalies. They are the visible peaks of a systemic cost-to-value gap.

Cost failure has a home address

The paper uses the same Gen AI Value Architecture as the first two papers in the series, so the three can be read as a single diagnosis. The pattern is clean and it is new. In the governance paper the failure markers clustered in governance and evaluation. In the bias paper they clustered in data and evaluation. Here they cluster in Value and Adoption, in the Input layer's Use Case Definition, and in the runtime cost controls of Runtime Operations. The technical model layers were rarely the primary origin. Cost failure does not originate where the model is built. It originates where the organization approves the use case, measures value, and controls runtime spend. Five components account for most of the damage: Cost Management, ROI Measurement, and Use Case Definition, each implicated in all five cases, followed by Human-in-the-Loop and Rate Limits and Timeouts.

Five cost-failure mechanisms

Across the cases, the same mechanisms recur. Unbounded unit cost, where consumption is celebrated as adoption while the bill compounds. A hidden human cost base, where concealed human labor breaks the unit economics. Net value never computed, where gross productivity is mistaken for net value. Indefinite time-to-value, where an open-ended payback horizon lets an unproven program run until an external shock stops it. And no stopping rule, where the balance sheet, a fatality, or budget exhaustion supplies the stop instead of management.

The analyst forecasts

Two research organizations converge with the case evidence. Gartner reports that generative AI initiatives can cost 5 million to 20 million dollars each, that at least 30 percent of generative AI projects will be abandoned after proof of concept by the end of 2025, and that over 40 percent of agentic AI projects will be canceled by the end of 2027, in part over escalating cost and unclear value. MIT's Project NANDA reports that despite an estimated 30 to 40 billion dollars in enterprise generative AI investment, 95 percent of pilots produced no measurable return. Gartner counts the projects that die on cost. MIT counts the far larger set that survive technically and still return nothing measurable. Both are the satisficing failure at scale.

What the paper provides

The paper provides the economic frame, the architectural reference model, five forensic case diagnoses, the analyst-confirmed convergence, the cross-cutting patterns, a practical framework organized around cost-bound, time-to-value, net-value, and stopping-rule principles, and a single-page Cost-to-Value Gate of five questions a CEO, CFO, COO, or board chair should be able to answer for any production AI spend. It is written for executives, product leaders, finance partners, risk officers, and board members responsible for AI strategy and AI spend.

Companion to When AI Fails and When AI Fails Through Bias

This is the third paper in a series. The first showed that AI failure is, with rare exceptions, a governance and design failure. The second showed that AI failure through bias is a leadership and design failure now carrying a statutory standard of care. This paper shows that AI failure through unbounded cost is a discipline failure, a failure to satisfice. The same architecture diagnoses all three. The markers just move.