In the 1930s, a middling painter named Han van Meegeren became the top seller of the lost early works of Vermeer.1 Naturally suspicious, collectors called in experts to verify their authenticity, but it checked out. The canvases and stretchers matched dating techniques, the paint displayed characteristics of having dried over a hundred years, and the craquelure and pigments were all period-appropriate for a real Vermeer.

What the experts didn't realize was that van Meegeren was blending old forgery techniques with new technology2 to create these “masterpieces.” And he would have gotten away with it, too, if he hadn't sold Holland's greatest “treasures” to the Germans.

When he was put on trial for being a Nazi collaborator, he faced an impossible task: to prove that he wasn't a traitor, but rather a fraud. Under observation, he painted what would be his last Vermeer, demonstrating his techniques to the jury. The spell now broken, people couldn't believe they'd been fooled by paintings that were (by any honest standard) poorly done.

A story very much like this one is playing out right now, as CEOs come to their earnings calls with a compelling narrative about layoffs driven by “AI adoption.” Like van Meegeren, they're trying to clear up questions about what would otherwise look like poor management practices with a story just credible enough, and just opaque enough, that few (if any) people will demand technical proof.3

I.Questionable ProvenanceA story of convenience.

If you're a CEO looking to justify an aggressive round of cuts, claiming AI as the cause is a slam dunk. It flatters your leadership (“you seem to be ahead of the curve!”), and yet it's impersonal enough to avoid accountability (“this technology is a game changer!”). Nobody needs to own the mistake, or explain why basic forecasting failed. The reduction in headcount is just the consequence of progress. It also draws attention away from softening demand, which might stir panic, and instead implies that there's no real alternative.

In truth, most people don't understand AI well enough to push back, so they accept the story as told. They know it can produce text, code, images, summaries—even analysis. What they don't appreciate, and what we've discussed in previous pieces, is that AI generates the same or more demand for outcomes: correctness, accountability, auditability, and integration. AI produces drafts;4 businesses run on decisions, and at the end of the day, decisions are made by people.5

When AI creates real efficiency, it shows up in one of three ways. First, the same volume of work gets done with fewer human minutes. Second, more volume gets done with the same headcount. Or third, the same volume gets done faster with the same headcount, and that speed converts into revenue or avoided hiring.

Only the first supports the claim that people were laid off because of AI. The other two can support a productivity claim, but productivity has always been hard to pin down in metrics that survive scrutiny. We have seen this movie before. In 1987, the economist Robert Solow quipped that you could see the computer age everywhere except in the productivity statistics, and it took most of a decade (and a great deal of organizational redesign) before the gains from computing finally showed up in the numbers. A company booking AI productivity in a single quarter is claiming something that historically arrived only after years of patient process work. Most companies haven't done that work.6 Buying a tool and calling it a strategy has never been enough.

So when you hear the claim of AI gains, ask the hard questions. Where? With what baseline? And who is paying the verification tax?

II.Authenticating the ClaimShow me the numbers.

Much like carbon dating and spectrograph analysis helped expose artistic frauds, these claims can be validated with four simple numbers: volume, human minutes per unit, AI cost per unit, and quality.

Volume is how many units of work happen per week or per month—tickets, cases, claims, campaigns, invoices, deployments. Human minutes per unit is exactly what it sounds like, before and after, not “productivity improved” but actual minutes. AI cost per unit is the true cost that includes model usage, tool calls, orchestration runtime, retrieval, storage, monitoring, evaluation, and the human review minutes that nobody wants to count but absolutely should. And quality is the rework, escalations, incident rate, repeat contacts, refunds, and SLA misses. If it's internal, use downstream correction metrics. How often does someone have to fix what the system produced?

Without this bridge, executives will cloud the story with anecdotal evidence. You've heard the greatest hits on earnings calls in the past six months: “We're seeing strong adoption”—measured how? Logins? Messages? Task completions? “Employees love it”—unfortunately “love” is not a metric, the same way “hope” is not a strategy.7 “It's saving time”—whose time? Doing what? What happened to quality? And my personal favorite: “We can do more with less”—great, did you actually do more? Or did you just do less and brand it as “focus”?

The second place to look is the expense sheet. AI costs rarely show up as “AI costs.” You'll find them tucked under infrastructure, security, engineering, vendor spend, restructuring, and professional services. If headcount is down and those costs are creeping up, or headcount is down and quality is down, expect a secondary cost wave.

Specifically, look at cost of revenue. If the “AI efficiency” is customer-facing automation, inference often sits in COGS. You can lower SG&A and still compress gross margin.8 You haven't saved money so much as relocated it. Look at cloud spend. AI feature adoption scales with usage, which is the opposite of the SaaS story everyone grew used to, where incremental usage was mostly margin. Metered AI brings variable cost back into parts of the business that have forgotten what that feels like.9 Look at security and compliance headcount.10 And look at evaluation and quality operations, because if you want reliability, you pay for test sets, red-teaming, regression runs, and human adjudication. Skip paying for it now and you pay for it later in incidents.

III.Value AppraisalThe leverage shift.

If you've read other insights I've published, you know this is a recurring theme, but it bears repeating. Consider the electronic spreadsheet. It made financial modeling cheap, and cheap modeling raised the ambition of finance teams. “We can model this” became “we must model this.” The work expanded to fill the new capability. The people stayed, and their leverage changed.

Generative AI behaves the same way. It collapses the cost of producing intermediate artifacts and shifts the constraint to review, decision-making, governance, and integration. Those constraints are staffed by people. So when a company claims “AI gains drove layoffs,” you should probably assume one of two things is true.

Either the company found a workflow that was already close to an assembly line, instrumented it, held quality, and converted minutes into headcount reduction. That's possible. But it's rare enough that it should come with receipts.

Or—and this is the more common pattern—the company did what companies usually do. It used AI to justify a correction it already needed, and it will spend the next year rebuilding some of the capability it cut, just in different departments and with different job titles.11

That second pattern is the natural outcome of pretending output is throughput. Speeding up one stage of a process only helps until the bottleneck moves to the stage you didn't speed up.12 If drafting was a fifth of the work and judgment was the rest, making the draft instant just loads the part that was already the constraint: the human review waiting on the other side. So even when AI makes individuals faster (sometimes dramatically faster), that speed doesn't convert cleanly into fewer people. Organizations are remarkably bad at turning productivity into headcount reduction without harming the business. The easiest conversion is hiring avoidance followed by attrition. Actual layoffs are a blunt instrument with second-order effects: morale, trust, institutional knowledge, and the risk that you cut the people who understood the tricky parts of your operations.

So “AI-driven layoffs” is often a category error. The company got some productivity benefit and chose to turn it into layoffs because the market wanted margin repair now rather than optionality later. The more accurate version of the sentence is “we chose the fastest visible lever.” “The AI made me do it” simply plays better on a call.

That's why the market reaction can look irrational. A stock pops on the announcement13 because investors see immediate OPEX reduction and a confident story. But the market will punish it later when quality degrades, or growth stalls, or the savings prove illusory because costs migrated, just like it always does.

IV.The Appetite for More VermeersJevons' paradox.

There is a detail in the van Meegeren story worth dwelling on. He didn't succeed in a vacuum. He worked a market that was hungry for rediscovered Old Masters, and the easier it became to “find” a lost Vermeer, the more the market wanted to buy. Cheapening the supply of Vermeers did nothing to satisfy the appetite for them. It fed it.

That is a microcosm of a pattern an English economist articulated more than a century ago. Worried that Britain was running out of coal, observers of the day assumed that more efficient steam engines would stretch the remaining supply and cool demand. In 1865, William Stanley Jevons argued the reverse, and he was right. As engines grew more efficient, coal became cheaper to put to use, and cheaper coal got used for more things, so total consumption climbed. “It is a confusion of ideas,” he wrote, “to suppose that the economical use of fuel is equivalent to diminished consumption. The very contrary is the truth.”

AI lowers the cost of producing the intermediate artifacts of knowledge work: the draft, the analysis, the first pass of code, the summary. The layoff narrative assumes that when these get cheaper, a company does the same amount of work with fewer people (and even that is predicated on an assumption that the amount of work is fixed, which it almost never is). When a capability gets cheap, we tend not to do less of it; we do far more. The team that could write five reports writes fifty. The service that shipped quarterly ships weekly. And every one of those artifacts still needs the human-bound work that surrounds it: someone to check it, reconcile it, decide on it, own it, and wire it into everything else. That work scales with the volume of output, and the volume is rising.

A word of caution, because this argument is easy to abuse. Jevons describes a tendency, not a law of nature,14 and AI optimists have lately waved it around as a guarantee of endless new work. It is no such guarantee. The rebound only shows up where demand is elastic, where cheaper supply unlocks uses that weren't worth the cost before. For genuinely commoditized, satiable tasks, cheaper production really can shrink the labor attached to them, and those are the cases where an AI layoff claim might actually hold. But most knowledge work is nowhere near satiated, and the human tasks downstream rarely diminish in proportion to the output above them. A company cutting staff while its AI-assisted output climbs is therefore making a specific, testable bet: that downstream demand is satiable and that quality will hold as volume rises. That is precisely the kind of claim that should arrive with the four numbers attached, not a slogan.

V.Beyond the BrushstrokesProvenance, on demand.

To be clear, none of this is anti-AI. If you've read our other work, you know we're bullish on what this technology can do when it's applied with discipline. Our objection is to handwaving, and right now there is a great deal of handwaving (with and without a paintbrush).

Van Meegeren's forgeries held up because they were credible enough and because the people buying them wanted the impossible. His story survived right up until it was forced to pass a test it had never been built to pass. The experts hadn't gotten smarter in the meantime. The conditions had simply changed.

That's where “AI-driven layoffs” is headed. Right now, executives can borrow credibility from the ambient AI moment. They can point to a few internal tools, sprinkle in terms like “MCP” and “models” and “agents,” and let everyone else fill in the operational details they never provided. That works until someone demands the details. And someone always does. Eventually it's the board asking why cloud costs didn't stop rising after the cuts. Or the regulator asking for traceability and oversight. Or maybe it's your biggest customer escalating a quality failure that the reduced support org can't absorb.

When that moment comes, there are only two outcomes. Either the company can produce provenance (the four numbers, the bridge from volume to minutes to cost to quality) and show that labor minutes actually disappeared, that the AI cost per unit didn't erase the savings, and that quality held. Or it can't, in which case the layoff was narrative-driven all along.

The point of this piece isn't to argue about whether AI is transformative (it absolutely is). The point is to stop letting transformation claims stand in for evidence. An organization should be able to demonstrate its acceleration the way van Meegeren, in the end, had to demonstrate his forgery: in the open, on demand, with everyone watching. Surviving that kind of scrutiny is hard to do alone. At Privlex, we provide the privileged oversight that lets you operate in the open with confidence.