The CEO who fired 10,000 people because of ChatGPT.

Last week, the New York Times ran a piece about SAP.

Christian Klein, the CEO, told the reporter he was not sure "someone in two or three years will still code software." The article notes SAP cut nearly 10,000 jobs in 2024, and the company said some of those cuts were "as a result of AI." Klein went on to imply that if his company did not aggressively reinvent workflows, entire European sectors were about to lose jobs to AI at scale.

Very confident claims…casually made, and printed on the front page of the business section.

Two paragraphs later, the same article describes what SAP is actually doing with AI inside its own walls…

Cleaning up patent applications. Handling some customer support tickets. Helping developers add features to existing programs “faster”.

… that’s the entire list.

My mind boggled. How can those “limited” use cases drive 10,000 job losses in a “demanding” company like SAP?

The reality is…

 

The gap between what executives say AI is doing and what AI is actually doing inside their own companies is now wide enough to drive a truck through.

 

The spring of 2024 problem

Let me put SAP's 2024 layoff claim in context, because the timeline is the part nobody bothers to check.

The layoffs happened in early 2024. The state of the art in generative AI in early 2024 was GPT-4o. GPT-4o's headline feature was multimodal input. You could paste in an audio clip. There was no serious coding agent yet. Devin had just been released in an experimental preview, restricted to a handful of testers. Cursor was a niche IDE fork. Claude Code did not exist. Copilot was still an autocomplete plugin most engineers turned off because it slowed the IDE.

There was no version of the technology in the spring of 2024 that could sit in a room and replace a mid-level SAP engineer. Not even close.

And yet, two years later, an executive at one of Europe's largest software companies stands in front of a reporter and says, on the record, that some of those 10,000 cuts were "as a result of AI."

Jensen Huang, who runs the company whose chips are actually training the models, called this pattern ridiculous. His words, in a keynote earlier this year: "It was just a way for executives to sound smart, and I really hate that. I think they're scaring people, and that's irresponsible."

Sam Altman, who runs OpenAI, told a conference audience in May: "I'm delighted to be wrong about this. I thought there would have been more impact on entry-level white collar jobs being eliminated by now than has actually happened."

The people who build the models are publicly saying that the models did not do what the executives said they did.

The executives are saying it anyway.

(…and from my experience as a university professor, those claims, made “to sound smart”, are VERY toxic, especially for young people.)

 

Two conference rooms

Let me walk you down the hall of the company these executives run.

Conference room A. Fourth floor. The Q3 board meeting. The CFO clicks to slide 17. There is a bullet that reads "AI-enabled headcount optimization: 12 to 15 percent flat over 24 months." Nobody in the room asks what AI-enabled means. Everybody nods. The slide travels to investor relations. It becomes a line in the earnings call. It becomes a Bloomberg headline the next morning. Analysts revise the model. The stock ticks up seventy basis points.

Conference room B. Same floor. Twenty meters down. The AI ops working group. Six people around a table, one of them a manager I'll call Thomas. Thomas is Head of AI Ops at a large European software company. Not SAP, but close enough. He runs the meeting.

Item one on Thomas's agenda: a support-ticket triage bot that classifies incoming tickets into fourteen categories. Accuracy at 78 percent. Human agent still confirms EVERY classification before the ticket routes. Deployed in production for eleven months. Saves maybe six minutes per ticket (none in most cases).

Item two: a coding assistant helping the engineering org write CRUD endpoints slightly faster. Adoption at 43 percent of eligible engineers. Most senior engineers turned it off after two months because they spent more time reviewing than typing.

Our coding assistant is actually used, but no or very little real productivity gains are noticed, as “verification” and “ownership” time increase and people become lazier…

Item three: an internal knowledge-base chatbot that answers benefits questions. Legal made them add a disclaimer. HR routes every conversation to a human anyway.

That is what AI does at Thomas's company.

The distance between conference room A and conference room B is twenty meters. The distance between what conference room A says AI is doing and what conference room B is actually shipping is measured in the market cap the CFO defended on the earnings call.

Thomas has stopped attending conference room A. He was asked once, six months ago, whether the coding assistant could support a 20 percent reduction in engineering headcount. He said no, honestly and specifically, and cited the numbers. He was thanked for his input. The slide went to the board the following week with the same 20 percent reduction on it.

The reduction happened. It was attributed to AI.

The coding assistant had nothing to do with it.

These are real stories, I witnessed them multiple times, not my imagination.

 

The Elodie test

Elodie runs procurement at a mid-sized European industrial. She has bought exactly one AI product this year. A summarization tool for supplier contracts. It works. It is boring. It saves her legal team about forty minutes per contract review, which adds up over three hundred contracts a year.

Elodie's CEO told the annual town hall in June that the company was going to be "AI-first" by the end of 2027. He announced this after taking a three-day executive AI course at a European business school. Elodie was in the front row. She counted eight uses of the phrase "paradigm shift" in twenty minutes.

She went back to her office and asked her CEO's chief of staff, in writing, what "AI-first" meant operationally. What headcount targets. What systems. What budget line. What owner.

The chief of staff did not reply for three weeks. When the reply came, it was a three-paragraph message about "cultural transformation" and a link to the McKinsey generative AI report.

Elodie now uses a private phrase for this pattern. She calls it the executive AI vibes gap. Confidence at the top. Contract summarization at the bottom. Nothing in between except a Bloomberg quote and a McKinsey report.

I asked her once what she does when the CEO says something at the town hall that does not match what any of her actual AI systems can do. She said, "I nod. Then I go do procurement."

That is roughly the state of enterprise AI in mid-2026.

 

The uncomfortable truth

Confidence about AI is inversely correlated with operational contact with AI.

I know how that sounds. But this is EXACTLY what I see in every enterprise conversation I have. The people closest to the actual deployment (the Thomases, the Nadias, the Elodies) speak about AI in modest, specific, load-bearing terms. Percentages. Failure modes. What the bot cannot do. What the human still has to check.

The people furthest from the actual deployment (the CEO on the earnings call, the analyst on Bloomberg, the consultant in the deck) speak about AI in sweeping, confident, unfalsifiable terms. Sectors. Transformations. Percentages of the workforce.

The distance between the two registers is not a communication problem. It is not that the operators need to speak up more, or that the executives need to visit the ops room more. It is a structural fact about how the two roles get rewarded.

The CEO gets rewarded for sounding AI-forward on the earnings call. The stock moves or investors, associates are happy. The board is happy. There is no penalty for being wrong about what AI can do, because "wrong" would require someone in the room to have the technical standing to say so, and that person is usually Thomas, who stopped attending (not invited anymore, “too technical for the meeting”).

The Thomas gets rewarded for shipping working systems. Working systems are boring. Working systems do not move the stock.

The market has decided that confident AI-forward posturing is a public good, and that operational AI honesty is a private one.

This is how you get a CEO attributing 2024 layoffs to a technology that could not have done them, in an interview with a national newspaper, and nobody at the paper thinks to check the timeline.

This is organizational folklore. And it is now traveling in newspapers.

 

What to do Monday morning

Pick one. Not three. Run it for ninety days.

One. If you are an operator, write the one-page reality memo. One page. What your AI systems actually do this quarter. What they do not do. What accuracy. The reality of verification and ownership, the fact that AI can’t be held accountable and what people are actually doing now and why “reducing headcount” is not exactly what people think is gonna happen. What headcount they have made more productive (a specific number, or the honest word "unclear"). Send it to your CEO once a quarter. Do not editorialize. Do not push back on the strategy. Just publish the number. If the CEO's public claims start drifting from the one-pager, that drift is now on record. That is your only defense against being the person in the room when the model catches up with the folklore.

Two. If you are an executive, do the walk-down-the-hall test before every public claim. Before you say anything about AI in an earnings call, a town hall, or a newspaper interview, walk down the hall to the person actually running the AI system you are about to reference. Not their manager. Not the deck. The person whose Slack is on fire when the system breaks. Ask them, in their words, what the system does today. If your public claim survives that conversation, say it. If it does not, do not say it. Not because you might be caught. Because you should be caught, and the person running the system deserves better than being made a liar for your quote.

Three. If you are a journalist or an analyst, ask the timeline question. When a company attributes past layoffs to AI, ask which AI. Ask what model. Ask what internal system. Ask what percentage of the workflow it replaced. The answer will be either specific, in which case you have a story worth writing, or vague, in which case you have a different story worth writing. Do not print the claim without the timeline. The timeline is where the folklore breaks.

The reality of “AI communication” in 2026 is this: most people don’t know what they’re talking about…

 

Back to the two conference rooms

I keep thinking about Thomas sitting in the AI ops meeting, twenty meters and a stock price away from the room where the CEO is telling the board that AI is eliminating headcount at 15 percent flat.

Thomas is right. The CEO is confident. Confidence wins the earnings call. Being right wins the outage in eighteen months, when a customer notices that the "AI-enabled" tier of the product does not do the thing the CEO said it did.

The reckoning is not going to come from a regulator. It is going to come from a customer who tries to use one of these systems for real, and finds out that "AI-first" was a slide, not a stack.

Jensen Huang and Sam Altman are giving the game away in public, and the executives are still doing it anyway. That is the tell. When the vendors of a technology are telling you their customers are overclaiming, and their customers do not stop, you are not looking at a technology story. You are looking at a status story. AI-forward posturing has become a professional signaling requirement. The people who refuse to do it (Thomas, Elodie, the ones actually shipping) get called cautious, or slow, or "not visionary enough."

They are not slow. They are in the room where the systems have to work.

AI-first is earned by shipping, not by announcing. Every quarter, the gap between the shipping companies and the announcing companies gets wider. In 2028, one of those groups will be quietly compounding real capability. The other will still be issuing paradigm-shift press releases about a workforce it already claimed to have optimized.

You already know which one your company is.

AI is only as good as the human operating it.

Have a great weekend.

Stay sharp.

— Charafeddine (CM)