EXECUTIVE PERSPECTIVES  ·  MANIFESTO

Enterprise AI Debt: the liability no one booked.

Julian Morris, Managing Partner, LeanCxOs  ·  October 2026

Last year, 42 percent of companies surveyed had abandoned most of their AI initiatives, more than double the year before. Almost no one asked what happened to the things they walked away from. The answer is a liability with a size, a carrying cost, and a compounding rate. It is expensed everywhere and booked nowhere. It is time to book it.

Download the manifesto (PDF)

The AI disappointment statistics are familiar furniture by now: abandonment more than doubling in a year, nearly half of proofs-of-concept scrapped before production. They are usually quoted as a story about failure. They are actually a story about accounting, because almost nobody asks the follow-up question. What happened to everything that was abandoned?

In most organizations, the answer is that nothing happened to it. The licenses still bill. The cloud instances still run. The model a departed data-science lead fine-tuned is still wired into a workflow no one wants to touch. The integration a vendor built in a hurry still holds credentials to production systems. The dataset copied out for a pilot still sits, ungoverned, where the pilot left it. The promises made on slides two budget cycles ago are still remembered by the board, if not by anyone accountable for them.

What was abandoned did not disappear. It became debt.

A word on the term. The financial press has lately used AI debt for the bonds funding the data-center build-out. This essay is about the other AI debt, the one inside your enterprise, created not by borrowing but by not deciding. The bonds, at least, appear on someone's balance sheet.

A definition

A controller will object, correctly, that every one of those costs is already expensed: a license line here, a cloud line there, spread across a dozen cost centers. That is exactly the problem. The costs are booked everywhere and the liability is owned nowhere. No one aggregates it, no one is answerable for it, and no one brings it to a decision. It is a liability in every sense except the accounting one, which is why no one manages it.

Not technical debt. Leadership debt.

Engineers deserve credit for the ancestry. Ward Cunningham coined “technical debt” three decades ago to describe deliberate borrowing: shortcuts taken knowingly to ship and learn, priced, and paid down on purpose. A landmark 2015 paper, “Hidden Technical Debt in Machine Learning Systems,” warned that most machine-learning debt lives at the system level, not in the code. The engineering conversation named its piece of the problem. No one has named the whole, or given a board a way to book it.

The debt that now matters most was not created in code. It was created in conference rooms. A pilot that was never formally closed is not an engineering artifact; it is an unmade decision. A model with no owner is an unmade decision. An agent running with production access and no review cadence, a dataset with no steward, a vendor claim never validated against results: each is a decision no one made, still accruing interest. That is why the fix is not an engineering program. No refactor retires an unmade decision.

Like Cunningham's original, some of this debt is rational. Borrowing speed against discipline can be a sound trade when the borrowing is underwritten: taken deliberately, priced honestly, and scheduled for retirement. Debt is not the failure. Unbooked debt is.

The six classes of AI debt

In leadership-team work, AI debt shows up in six recognizable forms:

1. Pilot debt. The stalled and zombie initiatives that were never funded forward and never shut down, still consuming budget, attention, and political capital. Most organizations cannot produce the list. The missing list is the finding.
2. Model and prompt debt. Models, fine-tunes, prompt libraries, and agents in production or in limbo, without owners, documentation, monitoring, or retirement criteria. Ask who owns a given model's behavior and watch the question bounce between functions.
3. Integration debt. The point-to-point connections and workarounds that bind AI components to core systems, built quickly and outside architecture governance, now load-bearing. Nothing in an enterprise is more permanent than a temporary integration that works.
4. Data debt. Datasets copied, transformed, and exposed for AI use without lineage, quality ownership, or retention discipline. Every ungoverned copy is a liability with no line item.
5. Governance debt. AI already deployed that will eventually need decision rights, thresholds, review cadence, and regulatory readiness it does not have. That bill arrives all at once, on the regulator's schedule rather than yours.
6. Workforce debt. The commitments, skills gaps, shadow usage, and quiet cynicism created when AI adoption outruns the redesign of roles and accountability. The hardest class to reverse, because it compounds in trust.

Assemble the register at a typical mid-market enterprise and the external record predicts the shape. Expect dozens of initiatives where leadership counted a handful. Expect a meaningful minority with no owner anyone can name. Expect carrying costs approved one line at a time and never once read as a total. The sizes will vary. The silence around them will not.

Why it compounds

Financial debt compounds by arithmetic. AI debt compounds by neglect, along three curves at once.

The cost curve: consumption grows with adoption, agents multiply faster than anyone inventories them, and every orphaned asset keeps drawing budget through a dozen small line items no one reads as one. Gartner predicts more than 40 percent of agentic AI projects will be canceled by the end of 2027, and each cancellation leaves exactly this residue. The exposure curve: each unowned model, ungoverned dataset, and undocumented integration is surface area for security incidents, for compliance findings, for the audit that eventually asks a question no one can answer. And the organizational curve, the steepest of the three: every quarter in which AI activity visibly escapes decision discipline teaches the organization that AI is something that happens to it rather than something it governs. People stop asking permission, or stop trying at all. The first builds shadow debt. The second forfeits the upside all this spending was meant to buy.

There is also a clock. Boards are asking harder questions each quarter, while two-thirds of board members and executives, by Deloitte's count, say their boards have limited or no knowledge of AI. Regulatory obligations, including the EU AI Act's high-risk requirements for stand-alone systems, now deferred to December 2027, will demand exactly the inventory and ownership discipline most organizations lack. “We're still figuring it out” is an answer with an expiry date, and regulators have already set it.

Book it. Value it. Decide.

No board would tolerate an unbooked liability of unknown size anywhere else in the enterprise. AI is currently the exception. Ending the exception takes the same discipline that governs every other liability: book it, value it, and decide.

Booking it means a census: every AI initiative, whether active, stalled, abandoned, or shadow, in one register with its annualized carrying cost, its exposure, and, above all, its owner.

Valuing it means numbers leadership can defend in front of its own controller. The census records the whole estate; Total AI Debt books only part of it: the annualized carrying cost of every item carrying one of the six defects, an unmade decision, reported beside a rated exposure profile across security, compliance, and operations. Healthy, governed, value-producing systems are operating cost, not debt. Then one ratio above the rest, the Debt-to-Value Ratio: what your AI estate quietly costs, divided by what it demonstrably returns. Most companies have never computed it; a ratio no one has read is exactly how the exception survives.

Deciding means the step almost every AI review skips: a session in which leadership disposes of every material item, with a name and a date attached to each. Fund it, govern it, scale it, measure it, or stop it.

Whether or not you ever engage an outside firm, your organization already has a debt register. The only question is whether it is written down, or running silently in your infrastructure, your risk surface, and your people's patience.

AI strategy defaults to the future tense: what to build, what to buy, what to pilot next. The debt conversation is the discipline of the present tense. Leadership teams that master it are not slowing AI down. They are the only ones who get to scale from a position of authority.

Decide before you scale.