Promise Arbitrage: Why Nobody Prices an AI Promise Until Engineering Pays It

On September 29, 2015, GE promised more than $15 billion in software and solutions revenue by 2020, in its own press release. The promise cost nothing to make. No line item priced the delivery, nobody owned the definition of production-ready, no evaluation evidence was required to say it out loud. On December 13, 2018, another GE press release announced the digital business would become a standalone company, starting with $1.2 billion in annual software revenue against the $15 billion promised. The division's chief departed. In between, successor CEO John Flannery took to telling his teams "no more success theater."

This is not a story about one CEO. It is about what the same promise costs at each table it passes through. An org chart is a market for promises. In most companies, nothing prices them.

Walk the Promise Down the Chart

A board verifies stewardship through the two instruments it can read: the financial statements and the strategic narrative. In Weill and Ross's study of 250 enterprises, only 38 percent of senior managers could describe their own firm's IT governance. So the board asks about guidance and consensus, and its questions define what gets staged below.

The CEO converts ambition into public commitment. BCG's AI Radar, published January 2026, surveyed 2,360 executives and found 72 percent of CEOs now calling themselves the company's main AI decision-maker, double the year before. The biggest technical bet now sits in the seat with the longest clock and no instrument for verifying a technical claim. A promise made here is called strategy.

The CFO prices everything. Almost. Graham, Harvey, and Rajgopal surveyed roughly 400 financial executives, most of them CFOs: 78 percent would give up economic value for smooth earnings; 80 percent would cut R&D, advertising, and maintenance to hit a target. The most financially fluent people in any building. The driver is EPS against consensus, the one number outsiders can verify, and career risk. Never ignorance. So the $15 million AI approval memo carries priced rows for licenses and headcount and none for data readiness, evaluation, integration, or maintenance. No memo format forces those rows to exist.

The COO converts the announcement into a dated operating plan, and dates become the currency: on-time is legible to the whole table, working is not. Boeing's 737 MAX program made this literal with a countdown clock in the program conference room. I will come back to that clock.

The CMO holds the shortest clock at the table: 3.5 years of average tenure in Korn Ferry's study of the 1,000 largest US companies, on 2019 data. The payoff that fits that window is an announcement. MIT's Project NANDA report, "The GenAI Divide: State of AI in Business 2025," widely covered in the press, found 50 to 70 percent of GenAI budgets flowing to sales and marketing for "easier metric attribution, not actual value."

The BU head fights the internal capital market, where the AI ask competes with every other division's ask and inflates to clear budget season. The build-versus-buy call gets made here under time pressure, and the upward report carries pilots launched and adoption counts. Eval scores stay off the page, and so does what the thing will cost to own once integration starts.

And the CTO sits where the promise comes due. CIO and CTO tenure averages 4.6 years against a CEO's 6.9 to 7.4, so a CEO outlasts roughly one and a half technology chiefs, and 42 percent of CIOs do not even report to the CEO. This seat inherits commitments made at tables it never sat at, the data readiness, integration debt, evaluation, and maintenance costs still unpriced, and pays them in nights, attrition, and blame.

Now read the walk back. Every seat optimized the signal its own table can verify. No villain appeared in this section. That is the point, and the problem.

Promise Arbitrage

An AI promise is free at the table where it is made and expensive at the table where it is delivered. Each seat books the credibility gain now and passes the delivery cost one level down. An economist would call that cost an externality: real, unpriced, borne by someone outside the transaction. No seat in the approval chain runs an evaluation function, so a technical promise never gets marked down on its way up, and the accumulated debt lands on the one table that cannot pass it further. Engineering already has a name for deferred cost that compounds: technical debt. Promise debt is the general case, and engineering pays the interest on both.

Sull and Spinosa argued in 2007 that a company runs on its network of promises. I am after something narrower: nobody in the network prices them. What a leadership team has spent its careers doing shapes what it can verify. Hambrick and Mason named this upper echelons theory in 1984, and the meta-analytic evidence since, across 308 studies, backs them. When a table's shared language is financial, a financial claim clears in minutes and a technical claim never gets priced at all.

I want to be straight about the edge of the evidence. No study connects executive-team composition to white-labeling or vaporware; that link is my synthesis over adjacent, verified findings. The thesis is also testable: show me enterprises with no evaluation function anywhere in the approval chain whose AI commitments deflate on the way up, and it dies. I've written about why so few leaders know whether their AI works. This is what that missing answer does to every promise stacked above it.

The Rational Shortcut, and Its Mirror

Buying and wrapping is often the right call. External AI partnerships reach deployment roughly 67 percent of the time against 33 percent for internal builds, per the NANDA report, and a16z's 2025 survey of 100 CIOs found LLM spending out of innovation budgets falling from 25 to 7 percent as buying became routine. The pressure is rational too. Terry showed in Econometrica in 2023 that the short-termist trade pays off for the individual firm even as it burns socially valuable R&D. In that same CFO survey, 55.3 percent would delay a value-positive project just to hit the quarter.

The failure the mechanism produces is misrepresentation: bought reported as built, demo presented as product. The SEC's first AI-washing actions landed on March 18, 2024, fining Delphia $225,000 and Global Predictions $175,000 for advertising AI capabilities they did not have. Builder.ai entered insolvency in mid-2025 after claiming roughly $220 million of 2024 revenue that Bloomberg reported was closer to $55 million, with humans doing "the vast majority" of the AI-marketed work, per Rest of World.

Engineering runs the mirror, and I say this from inside the tribe. Jensen showed in 1986 that managers build empires when free cash flow lets them; engineering's version is the over-built platform. In one study of 591 developers, 82 percent believed trending technologies make them more employable. Resume-driven development is promise arbitrage aimed at a different audience. The background research even runs backwards: marketing careers predict more R&D spending, per Barker and Mueller.

Credentials do not settle it either. Ginni Rometty holds a computer science and electrical engineering degree from Northwestern and presided over Watson's oversell era. Brian Chesky studied fine arts at RISD, Stewart Butterfield philosophy; they built Airbnb and Slack. Tamaseb's study of unicorn founders found technical and non-technical CEOs in roughly equal numbers, and Bhagat and colleagues, across more than 14,500 CEO-years, found no education-performance link. The variable is what the table can verify, not which degrees sit at it.

This Is Older Than AI

Boeing's countdown clock, then. The House Transportation and Infrastructure Committee's final report on the 737 MAX, released in September 2020, records the clock and its mantra: "the value of a day." A November 2016 internal survey found 39 percent of Boeing's own FAA-authorized representatives perceiving undue pressure; 2,000 hours had been cut from avionics regression testing. The promise was priced at the top in schedule and paid at the bottom in test hours.

IBM marketed Watson as a revolution in cancer care. STAT reported in 2017 that the deployed product was "nowhere close," and internal documents STAT obtained in 2018 showed unsafe and incorrect treatment recommendations in test scenarios. IBM sold Watson Health's data and analytics assets to Francisco Partners, announced January 2022.

The pattern predates all of it. Control Data sued IBM over "paper machines" in 1968. The Justice Department filed a formal vaporware memorandum in US v. Microsoft in 1995. Bayus and colleagues found roughly 47 percent of 123 preannounced software products shipped three or more months late. Selling the promise ahead of the product is a litigated, half-century-old practice.

In every case, engineers sat at or near the top, so expertise cannot be what was missing. The missing piece was a pricing mechanism for the promise.

Evaluation Rights: Pricing the Promise

The instinct is to hand this to the board. Wrong altitude, or incomplete. Promises get approved wherever budget authority sits, so the fix attaches there: the CFO's approval memo, the investment committee's term sheet, the BU capital review, the board as backstop. I call the instrument Evaluation Rights. Three clauses.

Clause 1. No next funding tranche without a written, named-owner definition of production-ready. Anyone who ran Stage-Gate product development will recognize the shape. The gate here is evaluative rather than calendar-driven. Your VC already funds you in tranches against milestones; your internal capital market funds announcements.

Clause 2. Evaluation evidence reports outside the judged executive's line. Evidence that reports to the person it judges is testimony, and it gets edited like testimony.

Clause 3. Unscripted reproduction rights before any external claim. Whoever approved the capital picks the inputs and watches the system run before the press release. The demo team does not get to drive.

One calibration figure, consultancy research not peer review, so weigh it accordingly: McKinsey and Oxford examined more than 5,400 IT projects over $15 million; on average they ran 45 percent over budget and delivered 56 percent less value than predicted. That spread is what unpriced promises have always cost. AI raised the stakes and the story value.

The closest prior art is Jonas Schuett's proposal for internal audit functions inside AI labs. Evaluation Rights work the buyer's side: a demandable right attached to capital at the deploying enterprise, where most of the promises now live.

Monday, By Seat

CEO: Announce no capability that has not survived an unscripted reproduction.

CFO: Put Clause 1 in the funding memo, and price data readiness, evaluation, integration, and maintenance before you sign.

COO: Take "launched" off the milestone tracker. Put "passed unscripted evaluation under production conditions" on it.

CMO: Keep a dated register of every external capability claim, and label demos as demos.

BU head: Report bought as bought and built as built. Attach the integration and maintenance price to every pilot before claiming it as delivery.

CTO / VP Engineering: Write the production-ready definition with a named owner and a price. Publish evaluation evidence outside your own reporting line. Demo nothing you would not reproduce unscripted.

For the board, four standing questions that audit whether the chain below has adopted the rights:

1. What does production-ready mean for this system, and who at this table can define it?
2. Who has the authority to tell us this demo isn't real, and when did they last use it?
3. What did we celebrate last quarter: an unveiling or uptime?
4. What happened to our last three pilots after the announcement?

The Chain, Not the Chair

I started this piece convinced the problem was credentials, too few engineers at the top. The evidence refused to cooperate, in both directions, and I have left it mixed on the page. What survives is structural. Composition shapes the shared language of verification. Incentives reward what that language can price. And without an evaluation function somewhere in the chain, the promise stays free until engineering pays it.

Whatever seat you read this from, the promise crosses your table exactly once on its way down. Price it there.