By Thomas | financial enthusiast
My AI diary: special entry.
Dataiku just released their Global AI Confessions report. 900 CEOs. Eight countries. Harris Poll. Real numbers, not vendor fluff.
The headline stat: 78% of CEOs say a failed AI strategy could cost them their job and their company's future. Up from 74% who said the same thing last year looking at the same timeframe.
I read that twice.
The people supposed to replace workers are scared for their own jobs.
That's the irony nobody's saying out loud. We've spent three years talking about AI eliminating roles, cutting headcount, replacing functions. And here we are in 2026 with 62% of boards actively pressuring their CEOs to deliver measurable AI outcomes — or else.
Eighty-three percent of CEOs plan to deploy AI agents in full production this year. But confidence in doing that at scale dropped from 41% last year to 31% this year. That's not growth. That's ten points of nerves in twelve months.
And 65% are more worried about over-investing in the wrong AI vendors than under-investing. Damned. That tells me something about where the real anxiety lives.
Now let's talk about the money.
This is the part I find genuinely confusing — and I've been thinking about it for weeks.
The AI-versus-employee cost comparison sounds clean on a slide. An AI agent handles a customer interaction for under $0.50. A human agent costs $4 to $8 for the same interaction. Fully-loaded employee cost — salary, benefits, desk, training — runs around $110,000 a year. The math screams "replace everything."
Except.
McKinsey surveyed nearly 2,000 organizations across 105 countries in 2025. Only 39% report any level of EBIT impact from AI at all. Only 6% qualify as high performers — meaning 5% or more EBIT improvement. Nearly two-thirds haven't even started scaling AI across the enterprise yet.
That's a long way from the promise on the slide.
What's eating the savings?
Here's what I've pieced together from multiple sources.
Goldman Sachs projects that agentic AI systems will increase token consumption by 24 times by 2030. Not because tokens are expensive. Because AI agents running tasks burn far more tokens per job than a simple chatbot query. The infrastructure bill scales with ambition.
One Nvidia executive put it bluntly in Fortune this April: "For my team, the cost of compute is far beyond the costs of the employees." Uber burned through their entire 2026 AI coding budget in four months. Microsoft cancelled most of their Claude Code licences after six months — unsustainable token costs.
Then there's the part nobody budgets for upfront. Data preparation. Compliance monitoring in regulated industries (add 10–20% of project budget). Hallucination correction. Gartner puts the global business cost of AI hallucinations at $67.4 billion in 2024 alone. Eighty-five percent of organisations misestimate AI project costs by more than 10%. A significant portion miss by over 50%.
And 96% of CEOs in the Dataiku survey believe their employees are already using generative AI without approval. Forty-two percent think more than half their workforce is doing it. That's unsanctioned spending baked into every department, completely invisible to the finance team. (haha)
The governance gap.
Only 33% of CEOs report high confidence in their AI governance frameworks. Only 25% of CIOs can monitor all of their AI agents in real time. Only 34% of data leaders think their agents could pass a basic decision audit.
And yet 94% of CEOs say AI could offer better counsel than a human board member on some issues. Eighty percent of the same CEOs actively question or challenge AI outputs.
They trust it enough to stake their job on it. They don't trust it enough to let it run unsupervised. That's not a contradiction — that's actually the right posture. But it means the "replace the employee, save the cost" math requires an entire layer of human oversight that never appears in the vendor ROI deck.
My honest read on where this lands.
AI is cheaper than a human for specific, well-defined, high-volume tasks. Customer interactions with clear scripts. Document processing. Code completion. Narrow things.
The moment the task requires judgment, context, accountability, or someone to blame when it goes wrong — the cost gap closes fast. And most business-critical decisions still require all four.
The MIT study puts it bluntly: AI can technically replace work equivalent to 11.7% of the US workforce. But "technically capable" doesn't mean cheap, practical, or risk-free to deploy today.
The CEOs in the Dataiku report aren't scared because AI doesn't work. They're scared because it half-works, the bill is bigger than expected, the board wants proof, and nobody's figured out governance at scale yet.
That feels exactly right to me.
So what actually wins?
The 6% of companies McKinsey calls high performers. They didn't bet everything on one AI vendor. They didn't try to replace all the humans at once. They found the specific narrow places where AI is clearly cheaper and better, ran real pilots, measured the actual EBIT, and scaled from there.
That's slower. Less dramatic. Harder to announce in a press release.
But 94% of the others are still waiting for the ROI to show up.
Is your company in the 6% — or still working out why the AI bill doesn't match the cost savings you were promised?
