The price of deploying Generative AI without a container for human judgment — measured in capital, reputation, and governance exposure.
In Anatomy of Human Judgment, we argued that every economic era resolves the same way: when the dominant container for human judgment breaks, institutions that cling to it pay in lost trust, lost capital, and lost relevance.
Today the broken container is the billable hour and the corporate hierarchy, and the force breaking it is Generative AI.
We call the current condition the competence trap: organizations can now produce strategy, code, and analysis at near-zero marginal cost — but the lineage of each decision is unverified.
Output is infinite.
Verifiable judgment is not.
This article is the invoice. What follows is not speculation. Every figure below is drawn from named, dated, primary or S-grade sources, because the institutions we are addressing — boards, foundations, grant partners — fund on evidence, not on narrative.
Ledger Entry 1: Wasted Capital — The Pilot Graveyard
The most visible cost of the competence trap is capital deployed into GenAI programs that never return it.
- MIT Project NANDA — in The GenAI Divide: State of AI in Business 2025, based on 150 executive interviews, a survey of 350+ employees, and analysis of 300 public AI deployments — found that 95% of enterprise GenAI pilots fail to deliver measurable P&L impact. The report’s own language: the 95% failure rate “represents the clearest manifestation of the GenAI Divide.”
- Gartner predicted in mid-2024 that at least 30% of GenAI projects would be abandoned after proof of concept by end of 2025 — and its January 2026 retrospective confirmed that at least 50% were, citing poor data quality, inadequate risk controls, escalating costs, or unclear business value. Gartner also prices a single business-model-innovation GenAI deployment at $5 million to $20 million. Multiply that by a 50% abandonment rate, and the scale of institutional capital quietly evaporating becomes visible.
- Gartner’s forward view is worse, not better: through 2026, organizations will abandon 60% of AI projects unsupported by AI-ready data; and over 40% of agentic AI projects will be canceled by end of 2027, driven by escalating costs, unclear business value, and inadequate risk controls.
A CFO reading these figures should ask one question: where did the judgment go? The answer, consistently, is that the models performed as advertised — but the organization had no container for certifying that a human evaluated, weighed, and stood behind the output before it entered a decision.
The pilots died of unverifiable judgment, not bad technology. MIT’s lead author said it plainly: the core issue is “not the quality of the AI models, but the ‘learning gap'” — tools that don’t learn workflows, and organizations that don’t learn to integrate judgment into AI output.
Ledger Entry 2: Reputation Liability — The Mata Precedent
The competence trap’s second cost arrives the moment unverified output touches a system that demands accountability. The canonical case is now three years old, and its lessons compound annually.
In Mata v. Avianca, Inc. (S.D.N.Y., 2023), attorneys submitted a brief opposing a motion to dismiss, citing six federal cases. None of the six existed. They were fabricated by ChatGPT — with fake judges, fake docket numbers, fake internal citations, and plausible-sounding reasoning. When opposing counsel could not locate the cases, one of the attorneys asked ChatGPT whether they were real. ChatGPT confirmed they were. (It was hallucinating again.)
Judge P. Kevin Castel sanctioned both attorneys and their firm $5,000, finding “subjective bad faith,” and ordered the attorneys to send the sanctions ruling to the client and to each judge whose name had been falsely attached to the fabricated opinions. The $5,000 figure was not the real cost. The real cost was that a mid-career attorney’s name became the global reference case for AI recklessness — cited in ethics opinions, CLE programs, and standing court orders requiring disclosure of AI use.
Since Mata, the legal trade press describes the fallout for firms caught filing even a single hallucinated citation as “catastrophic, almost like a data breach incident.” Courts now issue standing orders requiring disclosure of generative AI in pleadings. State bars have published ethics opinions. The liability environment has formalized around a single principle: the human who signs is accountable for the machine’s output — and is presumed negligent if they cannot show how they verified it.
Ledger Entry 3: Adoption Friction — When Staff Won’t Trust the Tool
The third cost is quieter: programs that survive their pilots still stall because the humans inside the institution decline to rely on outputs they cannot trace.
In a Dataiku poll of enterprise customers, the top reasons GenAI efforts lost momentum after the pilot phase were poor fit with existing systems (40%) and — tellingly — users not trusting outputs (33%). This is the competence trap operating from the inside: the tool generates fluently, the workflow receives the output, and the operator, unable to see who weighed what, quietly routes around the system and re-does the work manually.
This is the precise mechanism MIT observed from the other direction: generic tools stall in enterprises “since they don’t learn from or adapt to workflows.” Both findings describe the same missing component from opposite sides — management can’t verify the judgment inside the output, and staff can’t trust it. The result is shadow re-work: AI output that must be fully re-verified by hand, capturing none of the promised productivity, while still carrying all of the tool’s costs.
Any board evaluating AI ROI should treat this as the hidden line item. If every AI-assisted deliverable requires 100% manual re-verification, the institution has not adopted AI. It has adopted a very expensive draft generator — and it is paying twice.
Ledger Entry 4: Governance Exposure — The Regulatory Clock Is Already Running
The fourth cost is the one foundations and compliance-minded boards should weigh most heavily: the gap between what institutions do with GenAI and what regulators increasingly require them to demonstrate.
The direction of regulation is unambiguous — accountability must be attributable. The EU AI Act’s risk-based framework (in phased application from February 2025 onward) requires documentation and human oversight obligations proportionate to risk for systems that inform consequential decisions. In the United States, courts have already constructed the common-law version of this requirement through Mata-style Rule 11 sanctions and a growing body of standing orders: if AI contributed to a filing, the human signer must be able to show the verification chain.
Gartner’s own failure taxonomy now lists “inadequate risk controls” and “unclear business value” among the top abandonment causes — an implicit acknowledgment that governance is not an overlay on the AI strategy; it is the AI strategy’s survival condition.
The institutional position most exposed in 2026 is not the firm that banned GenAI. It is the firm that deployed it broadly, produced volumes of consequential output, and holds no auditable record of the human judgment inside any of it. When the audit comes — a court order, a grant review, a regulator, an insurance carrier after an AI-assisted error — “our staff reviewed it” is not a chain of custody. It is an assertion. Assertions are what Mata’s attorneys made.
The Total: A Composite Picture for Funders
| Cost class | Evidence | Who pays |
|---|---|---|
| Capital waste | 95% of enterprise GenAI pilots show no measurable P&L impact (MIT NANDA, 2025); ≥50% of GenAI projects abandoned post-POC (Gartner, confirmed 2026); 5M–5M–20M per business-model deployment (Gartner) | Boards, investors, grant-makers |
| Reputation liability | Mata v. Avianca sanctions; trade press: fallout “catastrophic, almost like a data breach” | Firms, professionals, signatories |
| Adoption friction | 33% of enterprise users distrust AI outputs (Dataiku poll); shadow re-work consumes promised productivity | Operating management |
| Governance exposure | Rule 11 precedent, standing court orders, EU AI Act phased obligations; “risk controls” a top-3 abandonment cause (Gartner) | Compliance officers, boards, insurers |
The Inverse Is Also True
Here is the asymmetry that matters for our thesis. The same evidence shows what the 5% who succeed did differently. MIT found the successful pilots share “tight integration between AI solutions and the business processes they are meant to improve.” Gartner’s survivors turn pilots into production at twice the rate of others — by fixing fundamentals first. Trullion’s analysis of the same MIT data finds that specialized, domain-specific, compliance-aware deployments succeed at roughly 67%, versus ~33% for internal generic builds.
Read those findings through the lens of The Anatomy of Human Judgment, the survivors built a container — a domain envelope, a verification step, an accountable workflow — around the model’s output. The failures shipped raw output and hoped trust would follow. It never does. Every era in our historical surgery series produced the same result: trust follows the container, not the capability.
The thesis, restated for funders: Institutions do not lose money on GenAI because the models are weak. They lose money because they have no standardized way to certify, trace, and price the human judgment that must stand between the model and the decision. That missing container is what GreenDevex is building — the Certified Judgment Unit (CJU) standard — and the cost data above is why we believe the institutions that adopt it first will be the ones writing the case studies the 95% read in 2027.
References & Verification
- Fortune, “MIT report: 95% of generative AI pilots at companies are failing” (Aug 18, 2025): https://fortune.com/2025/08/18/mit-report-95-percent-generative-ai-pilots-at-companies-failing-cfo/
- Gartner press release, 30% abandonment prediction (Jul 29, 2024): https://www.gartner.com/en/newsroom/press-releases/2024-07-29-gartner-predicts-30-percent-of-generative-ai-projects-will-be-abandoned-after-proof-of-concept-by-end-of-2025
- Gartner, “Why 50% of GenAI Projects Fail — And How to Beat the Odds” (Jan 26, 2026): https://www.gartner.com/en/articles/genai-project-failure
- Gartner press release, 60% AI-ready-data abandonment (Feb 26, 2025): https://www.gartner.com/en/newsroom/press-releases/2025-02-26-lack-of-ai-ready-data-puts-ai-projects-at-risk
- Gartner press release, 40% agentic AI cancellation (Jun 25, 2025): https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027
- Mata v. Avianca, Inc., 678 F. Supp. 3d 443 (S.D.N.Y. 2023): https://www.law.berkeley.edu/wp-content/uploads/archive/2025/12/Mata-v-Avianca-Inc.pdf
- JD Supra, “Federal Court Turns Up the Heat on Attorneys Using ChatGPT for Research” (Aug 13, 2025): https://www.jdsupra.com/legalnews/federal-court-turns-up-the-heat-on-1849454/
- Dataiku, “MIT says 95% of GenAI pilots fail” (Oct 2025): https://www.dataiku.com/blog/moving-past-genai-pilots
- Trullion, MIT data analysis (Sep 8, 2025): https://trullion.com/blog/why-95-of-genai-projects-fail-and-why-the-5-that-survive-matter/
Next Article in the GreenDevex Judgment Economy Research Series:
“The Anatomy of Human Judgment“
