What does The Metric Boards Skip When They Fund AI Projects mean in practice?
When funding AI projects, boards frequently skip the reversal-cost ratio, a metric introduced by Dr. Jonah Tebaa to measure what an initiative costs to unwind relative to its initial capital ask. Calculated alongside projected return on investment, this metric accounts for decommissioning systems, contract renegotiations, and reversing workforce decisions like severance and rehiring. Tracking this ratio enables capital committees to identify high-risk proposals and fund deeply embedded initiatives through staged gates rather than outright commitments.
In a recent capital allocation cycle Dr. Jonah Tebaa uses to illustrate a pattern he sees repeatedly in his advisory work — an illustrative composite, not a named client — a board had $900,000 to split across three AI proposals. Ranked by projected return on investment, the choice looked obvious. Dr. Tebaa argues it was also wrong.
The Deck Says One Thing
The three proposals were a predictive-maintenance system for production lines (a $340,000 ask, 11% projected ROI), an AI customer-service layer replacing 18 contact-centre roles (a $280,000 ask, 24% ROI), and a dynamic-pricing engine embedded across roughly 600 active customer contracts (a $280,000 ask, 32% ROI — the strongest number in the deck). Ranked by ROI alone, the board funds the pricing engine first, the customer-service layer second, and defers predictive maintenance for lack of remaining budget.
In Dr. Tebaa's framing, that is the default failure mode of AI capital committees: optimizing for the size of the return without pricing what it costs to discover the return doesn't materialize. He points to Gartner's estimate that roughly three in ten generative-AI initiatives are abandoned after proof-of-concept as evidence that "does it work" is a live question for every proposal on the table, not a formality already settled by the pilot slide.
A Second Number, Not a Replacement
Rather than discard ROI, Dr. Tebaa's advisory practice adds a second figure alongside it: the reversal-cost ratio. It is calculated from four inputs any sponsor should be able to produce before a funding vote:
- The capital ask itself.
- The full cost to unwind — decommissioning systems, reversing people decisions such as severance and rehiring, exiting or renegotiating contracts, and covering legal and brand exposure.
- How long before the organization has reliable evidence the bet is working, rather than a vendor's projected ramp.
- The ratio of unwind cost to capital ask — with results near or above 100% flagged as initiatives where reversal costs as much as, or more than, building did.
He connects this directly to how regulators are already framing AI risk: the EU AI Act treats risk management as a continuous, iterative process across a system's lifecycle, not a single approval gate, and the NIST AI Risk Management Framework is built on the same premise — that the risk profile of an AI system doesn't freeze at the funding decision. Dr. Tebaa's contribution is turning that lifecycle principle into a single number a capital committee can act on inside one meeting.
The Same Numbers, Inverted
Run the reversal-cost ratio on the same three proposals and the order flips. Predictive maintenance unwinds for roughly $45,000 against its $340,000 ask — a 13% ratio, cheap and fast to correct if wrong. The customer-service layer unwinds for roughly $310,000 against $280,000 — 111%, meaning severance, rehiring, retraining and service remediation would cost more than the project itself. The pricing engine, threaded through 600 live contracts, unwinds for roughly $460,000 against $280,000 — 164%, the worst ratio of the three, attached to the highest ROI of the three.
Dr. Tebaa's read: the proposals with the biggest projected upside are often the ones most deeply embedded in people, contracts and customer relationships — which is exactly what makes them slow and expensive to reverse if the projection misses.
The Practical Shift
His recommendation is not to reject the higher-ROI proposals but to fund them differently: a bounded pilot cohort, a hard evidence checkpoint before further capital releases, and a pre-agreed unwind trigger set before money moves rather than negotiated after results disappoint. Predictive maintenance, with its low ratio, gets funded outright. The customer-service and pricing initiatives get funded behind gates. Same $900,000, same three proposals, a different commitment schedule.
Dr. Tebaa ties the stakes to broader adoption data. Stanford HAI's 2026 AI Index finds that "Organizational adoption reached 88%", so adoption itself is no longer the bottleneck — and the gap he points to, between organizations adopting AI and the smaller share actually capturing enterprise-level value from it, he attributes less to model quality than to governance that rewards the best story on a slide rather than the best-protected bet in the portfolio.
