The P&L Can Be Right — and the AI Decision Can Still Be Wrong

At Gartner's Data & Analytics Summit in London this May, Frances Karamouzis said something I think will age better than most of what was said about AI this year: "2026 has begun with a mortgage on talent due to massive 2025 layoffs."²

A mortgage is a precise word, and I don't think she chose it casually. It means the money has been spent, the asset is already in use, and the payments have not started yet. In the same session she gave the number that makes the metaphor uncomfortable: by Gartner's estimate, less than 1% of layoffs are directly attributable to AI, and headcount-led AI business cases have become, in her word, irreconcilable with how AI actually creates value.²

Sit with those two statements together. Through 2025, a large share of workforce reductions were explained — to boards, shareholders, CEOs, the market, other decision-makers and the people being let go — with some version of "AI is making us more productive." If less than 1% of those cuts were actually caused by AI, then either the explanation was a pretext, or the business cases were sincere but wrong about something. I think it was mostly the second, and I think the "something" is worth naming precisely, because two Gartner papers published since May describe it in detail.

1. The four payments on the mortgage

The Hidden Workforce Costs of AI (Bansch and Coyle, June 2026)¹ starts from a fact that sounds like good news: 88% of organizations plan to increase AI spending.¹ Its argument is that a large share of what that spending will actually cost is not in the business cases that justified it — because AI does not so much reduce workforce cost as move it to places the original model didn't look. The authors identify four mechanisms. I want to walk through each, because the pattern across them is the whole point of this note.

Now look at the four together and ask a simple question: which of them was caused by the technology?

None. The model did what it was supposed to do. Each of the four costs was caused by a decision — how aggressively to automate, whom to cut, whether to keep the juniors, whether to redesign compensation — that somebody made about the technology, using a business case that contained the savings and not the payments. That is the pattern, and it is the reason I think Gartner's evidence points one step further back than Gartner itself takes it.

Four costs, four decisions. The AI worked. The mortgage came from leadership decisions.

2. Two chains, one deployment

The chain in the business case is familiar:

AI → Productivity ↑ → Headcount ↓ → Cost ↓ → P&L ↑

There is nothing false in it. Every link is real and measurable, which is precisely why it is persuasive in a budget meeting. The problem is that "headcount ↓" is not an end state. It is the first link of a second chain — the one Bansch and Coyle's four mechanisms describe:

Headcount ↓ → Pipeline ↓ → Tacit knowledge ↓ → Capability ↓ → External hiring at 3–4× → Rehire up to 30% by 2029 → Future cost ↑¹

Two things distinguish the second chain from the first. It is slow: the payments arrive in eighteen months to three years, not in the next quarter. And it is anonymous: when they arrive, they arrive as a recruiting line, a consulting contract, a retention bonus or a stalled second wave of AI, and none of those entries carries the name of the decision that produced them. So the organization ends up with two curves. One is booked every quarter. The other is booked, if at all, at the moment somebody notices the organization can no longer do something it used to do. That asymmetry — not the technology — is what makes the decision dangerous.

The P&L can be right — and the AI decision can still be wrong.

3. Whose debt is it?

In August, Gartner gave the aggregate of those consequences a name. The Workforce Impact of AI: Acting on Changes to Work, Jobs and Talent (Vaziri, Mullery, Mesaglio et al., 4 August 2026)³ states in its abstract that "AI compounds organizational characteristics and contributes to the accrual of talent debt across enterprises," and tells CIOs to manage that debt across three dimensions: work, jobs and the redesign of the talent pipeline.

The verb matters. Compounds. AI does not neutralize a thin pipeline or undocumented knowledge; it amplifies whatever was already there. That is a useful and, I think, correct observation. But notice where Gartner stops. It describes the debt at the level of the workforce, where it shows up. It does not say who incurred it. The step back to the decisions that produced it is mine, and the argument should be judged on that step.

Talent debt borrows its logic from technical debt: a compromise made today that generates cost tomorrow. Borrow the logic all the way and something follows. Technical debt is never really the code's fault — somebody chose to ship the shortcut. Talent debt, by the same reasoning, is not the talent's fault. The talent did not decide to leave. Someone decided how much to automate and how fast. Someone decided which roles were "just cost." Someone decided how much data architecture to build first, how much governance was enough for now, whether the savings would go to margin or be reinvested, and whether the operating model would change or simply shrink. None of those are AI decisions. They are leadership decisions about AI, and they are the common cause behind every form of debt the transformation leaves behind.

AI Leadership Debt. The accumulated future organizational cost created when leadership decisions about AI adoption, investment and transformation exceed the organization's ability to redesign the capabilities required to sustain value creation.

Talent debt is the form Gartner has measured, and it shows up first because people are the easiest asset to cut. But the same decisions leave at least five other kinds: architecture debt, governance debt, capability debt, operating-model debt and measurement debt — the last of which needs its own paragraph, because Gartner's newest paper is about nothing else.

Six forms, one origin. Talent debt is the form with an external name and a measurable signal. Measurement debt is the newest — Gartner's August 31 paper finds 84% of CFOs cannot measure AI ROI while spend heads toward $5.6 trillion.⁴ Each form has a different executive owner, which is precisely why the total has none.

A fair objection

Isn't this just a taxonomy? Six known problems under a new label. I take the objection seriously, and my answer is practical rather than conceptual. Each of those six debts has a different owner — CHRO, CIO, board, COO, CEO, CFO — which in most organizations means nobody owns the total. A liability with six partial owners and no aggregate is a liability that never reaches the agenda. Naming the aggregate is what puts it there.

4. The variable the business case leaves out

Here is where I think the economics change, not just the vocabulary. A conventional AI business case is static: investment, productivity, savings, ROI, evaluated over a period. But the four mechanisms above are all dynamic — a decision at time t changes what the organization can do at t+1, and therefore what the next AI initiative can return. That means two different objects are hiding behind the phrase "AI value":

Take Gartner's own numbers and build the concrete case. An AI program generates $10 million a year by eliminating a hundred roles. The full $10 million is booked as cost reduction. Now apply the 30%: by 2029, thirty of those roles are rehired, at a premium, plus recruiting. Apply the pipeline effect: the seniors those roles would have produced are bought externally at three to four times. Apply the skill half-life: the AI specialists hired to build the system need replacing in two to five years.¹

Was $10 million created? Or was part of 2029's organizational capability monetized in 2025 and booked as if it were value? A single-period ROI cannot tell those two cases apart. That is not a flaw in ROI; it is a limitation of the horizon.

Which is why the conventional equation needs the term it currently lacks:

The obvious objection: you cannot subtract what you cannot measure. And here Gartner's fourth paper lands. The Real Threat to AI ROI Is Unchecked AI Spend (Clougherty Jones, 31 August 2026)⁴ reports that 84% of CFOs struggle to measure AI ROI — while AI spend heads toward $5.6 trillion by 2030, more than double today's level. That is measurement debt in one sentence: the organization is doubling a spend it cannot evaluate. But "hard to measure" and "should be ignored" are different things, and finance handles the first case routinely — through provisions, depreciation schedules, the discount rate. Nobody provisions for AI leadership debt today. That is a choice, not a law of accounting.

Where this sits in AVRT.⁶ In the AVRT measurement system, debt accumulation is the upstream constraint captured by ALDI — the AI Leadership Debt Index — which throttles every conversion rate downstream (DCR, ACR, VCR, AIRR). Gartner's AI-Labor Cost Ratio⁵ fits as a contextual variable alongside them, not as a replacement.

5. When extraction is right — and what the satisfied organizations do differently

I want to be careful not to overstate this. There are situations where capturing the whole gain as margin is the correct decision: a business in managed decline, a commodity process whose capability has no future value, a balance sheet that needs the cash now. Extraction is not a mistake in itself. It is a choice with a ceiling.

Suppose AI creates $100 of value. Leadership can take it to margin, or it can allocate part of it to reskilling, data capabilities, architecture, governance, workflow redesign, the juniors it was going to cut, the next initiative. Two trajectories.

Until a few weeks ago the compounding model was an argument by construction. Gartner's newest paper adds something closer to evidence: organizations highly satisfied with their AI outcomes spend 30% more on data management, governance and talent than the least satisfied, and organizations reporting high AI value spend 5.6 times more on AI and its foundations than those reporting low value.⁴

I would read this carefully — satisfaction is self-reported, and a correlation is not a mechanism. But the direction is the one the compounding model predicts and the extraction model does not: the organizations that got the value are the ones that paid for the capabilities underneath it.

Extraction model	Compounding model
Optimizes a single period	Optimizes Σ value over time
Question: "how many people can we eliminate?"	Question: "which capabilities do we eliminate, preserve, redesign, build?"
100% of AI value → margin	Explicit AI reinvestment rate (AIRR)
Loop closed; next wave starts from a weaker base	Loop open; each wave funds the next
Right when the capability has no future value	Consistent with the +30% / 5.6× spending pattern

Which leads me to suspect that AI capital allocation — the explicit decision about what fraction of AI value is reinvested, and into which capabilities — will matter more over the next five years than AI adoption. Adoption is now table stakes; 88% are increasing spend.¹ Allocation is where the trajectories diverge.

Which capabilities should AI allow us to eliminate, which should we preserve, which should we redesign — and which new ones should we build?

Karamouzis's prescription at the summit was a "paradigm shift from employee empowerment to process redesign and work integration."² I would put it as a change in the question. How many people can we eliminate? is not a wrong question. It is an incomplete one, and its incompleteness is exactly the shape of the debt. A business case that can answer the harder question above is a transformation. One that can only answer the first is a cost program with an AI label — and the 1% figure suggests we have been running a lot of those.

The question for any AI business case in 2026: what will we have to rebuild later, not just what are we eliminating today? If the answer is nothing, the savings are real. So is the mortgage.

Gartner has now measured the first form of that debt and given it a name. The rest of the balance sheet — and the mortgage — is a leadership job.

References

1. Gartner (2026). The Hidden Workforce Costs of AI. Jan Bansch & Joe Coyle, 1 June 2026. gartner.com/en/articles/ai-workforce-costs
2. Gartner (2026). Beyond the Hype: AI's Impact on Headcount and Business Value. Frances Karamouzis, Data & Analytics Summit, London, 13 May 2026 (Day 3 highlights). gartner.com/en/newsroom
3. Gartner (2026). The Workforce Impact of AI: Acting on Changes to Work, Jobs and Talent. Kabeh Vaziri, Alicia Mullery, Mary Mesaglio, Kristin Moyer, Meghna Joshi, Lily Mok & Caroline Hewings, 4 August 2026. gartner.com/en/documents/8226561
4. Gartner (2026). The Real Threat to AI ROI Is Unchecked AI Spend. Lydia Clougherty Jones, 31 August 2026. gartner.com/en/articles/rein-in-ai-spend
5. Gartner (2026). The AI-Labor Cost Ratio: A New Metric to Guide Workforce Change. Jonathan Jackson, Grace Myers & Vatsala Syed, 21 July 2026. gartner.com/en/documents/8155129
6. Barrientos, C. (2026). AI Value Realization Theory (AVRT): Framework and Measurement System. Working paper and LinkedIn series. AVRT executive brief