Everyone Said AI Would Replace You. The Data Says the Opposite.

The AI job replacement experiment just failed publicly. But the lesson hiding inside that failure is the most useful thing an analyst can know right now.

I’ve spent the last week reading through a set of numbers I can’t stop thinking about.

Not because they surprised me — I half expected something like this to emerge. But looked at together, the picture is sharper and more damning than I thought it would be.

Many organisations are quietly rehiring staff after laying them off for AI. Most of the layoff announcements were loud and public. The rehiring has been far less transparent. Careerminds surveyed 600 HR professionals and found that 30.9% of companies that made AI-driven cuts spent more money rehiring than they had saved from the original cuts. One in three companies lost money, net. Forrester Research puts the overall regret rate at 55%. Gartner, in a February 2026 forecast, predicted that roughly half of all organisations that cited AI as a reason for reducing headcount will rehire for essentially the same functions by 2027 — often under a different job title, because corporate face-saving has always had its own creative vocabulary. Of the companies that have already started rehiring, 52% did so within six months. The value of those roles, as one researcher put it, became undeniable the moment they disappeared.

Photo by Proxyclick Visitor Management System on Unsplash

The names you’d recognise are all in this data. Amazon cut 14,000 corporate roles in October 2025 citing AI investments. Block eliminated nearly half its workforce for the same stated reason. Pinterest slashed 15%. IBM automated 8,000 HR roles into a system called AskHR — then discovered the system couldn’t provide answers based on judgment or empathy, and began quietly hiring humans back to oversee what the AI couldn’t handle. Klarna reduced its customer service headcount from 2,300 to 1,600 using AI, then reached break-even for the first time in May 2026 partly by rebuilding the human oversight capacity it had dismantled. Challenger, Gray & Christmas estimated 55,000 US employees were laid off in 2025 solely due to AI implementation — 4.5% of all US layoffs that year.

The experiment ran at scale. The data came back. And what it’s telling us — if you’re willing to read it honestly rather than through whichever narrative you arrived with — is something nobody in the C-suite who ordered the original cuts wants to say out loud.

They confused tasks with jobs.

I. The $2 trillion mistake hiding in plain sight

To understand why this happened at all, you have to understand the pressure that created it.

According to Reuters, investors and companies poured roughly $2 trillion into AI through the end of 2025. That number is probably low. JP Morgan alone holds approximately $1.5 trillion in AI-related debt. At present, AI represents 15–20% of corporate bond indices. The total burden, when added together, likely exceeds $3 trillion — greater than the debt held across the entire banking sector in some measures.

Now think about the kind of AI that would need to exist for that level of debt to be serviceable.

Not useful AI. Not AI that saves some money here and there. AI that generates hundreds of billions in annual profit — at scale, reliably, replacing human labour in volume. Even Google, one of the most profitable businesses ever built, doesn’t produce enough annual revenue to service that debt on its own. The math only works in one scenario: AI actually replaces people at the scale and speed the pitch decks described.

So the pressure wasn’t just about looking forward-thinking to investors. It was existential. Companies that had taken on AI-related debt needed the replacement story to be true. They needed to fire people to prove the technology was working. The story had to become real — which meant manufacturing evidence that it was real, even when the evidence wasn’t there yet.

Oxford Economics reported in January 2026 that many layoffs CEOs framed as AI-driven were more accurately corrections for pandemic-era overhiring — with AI as the more forward-looking explanation. IBM’s own survey of 2,000 CEOs found that less than 25% of AI projects delivered returns on investment as expected. Just 16% were scaled to achieve broad enterprise-wide adoption.

Then came April 2026, when Oracle, Meta, and Snap all announced significant cuts within the same three weeks — not because they all independently reached the same conclusion simultaneously, but because a corporate playbook had become self-reinforcing. One CEO frames cuts as AI-driven, markets reward it, and every subsequent CEO calculates whether not doing the same makes them look behind the curve.

What actually happened, stripped of the language: companies fired workers for a story their own CFOs didn’t fully believe, rehired when the story broke down, and called the whole sequence innovation.

That is not a technology story. It is an incentives story — and understanding that distinction is the first step toward reading the rest of the data clearly.

II. Why the experiment failed — and what it reveals about value

The companies that tried to replace workers with AI didn’t fail because AI is bad. They failed because they misunderstood what their workers were actually doing.

Think of any knowledge job as two layers sitting on top of each other.

The first is the execution layer. This is the visible work — pulling the data, drafting the email, building the dashboard, responding to the customer query, writing the first version of the report. It’s what most job descriptions describe when they list required skills. And it’s the layer that AI is genuinely, undeniably good at — fast, cheap, tireless, and improving every month.

The second is the judgment layer. This is the invisible work — deciding what to do in the first place, understanding what the person asking actually needed as opposed to what they said they wanted, knowing which number matters and which just looks interesting, reading a room, navigating politics, building trust over time. It doesn’t show up in job descriptions because it’s hard to articulate. It doesn’t show up in performance reviews because it’s hard to measure. But it’s the layer that determines whether the execution layer produces something useful or just produces output.

Here is what companies discovered when they eliminated their workers: they got the execution layer at machine speed. But they lost the judgment layer entirely. And in most real business situations — in customer service, in data work, in HR, in marketing — the judgment layer is what the customer is actually paying for.

IBM’s AskHR handled the repetitive HR queries just fine. It completely failed on anything that required reading someone’s distress, navigating an emotionally complicated situation, or making a judgment call that didn’t fit the script. Klarna’s AI handled 70% of customer service queries. The 30% it couldn’t handle were the situations where a customer needed help most — billing disputes, complex product issues, emotionally charged interactions. The AI boomerang hits because the 70% was the easy part. The 30% was the actual job.

A 2025 Harvard Business Review survey of more than 1,000 executives found that companies achieving strong AI outcomes share one pattern: they’re using AI to augment human workers, not replace them. Industries most exposed to AI — where workers actively use the tools rather than being replaced by them — are seeing 3x higher revenue growth per employee, according to PwC’s 2025 Global AI Jobs Barometer, which analysed close to one billion job postings across six continents.

The same report found wages in AI-exposed industries rising twice as fast as in non-exposed industries. Workers with AI skills command a 56% wage premium — more than double the 25% premium from just a year ago.

That’s not the story the layoff narrative was telling. That’s almost the opposite story.

III. The psychology of why smart people believed a bad theory

Here’s the part that takes intellectual honesty to sit with.

The AI replacement story wasn’t just a corporate lie. Plenty of intelligent, well-meaning people genuinely believed it. And understanding why they believed it — what cognitive error made it seem plausible — is more useful than simply writing them off as wrong.

Daniel Kahneman’s research on what he called “What You See Is All There Is” describes a cognitive pattern where we make decisions based only on the information that’s immediately visible, ignoring what we can’t see. The execution layer of most knowledge jobs is highly visible — you can watch someone pull a report, see the email being drafted, observe the customer query being resolved. The judgment layer is almost entirely invisible — you can’t watch someone notice that a stakeholder’s question reveals a flawed underlying assumption, or observe someone reading the political dynamics of a meeting and adjusting their approach accordingly.

When you can only see half of what someone is doing, you naturally underestimate what they’re worth. Leaders who watched employees do visible, repetitive execution tasks all day saw those tasks get automated and thought: the job is being done. They were wrong — but they were wrong in a predictable way. They couldn’t see the judgment work they were eliminating because they never could see it.

The economist Michael Polanyi described this as the difference between explicit knowledge — things you can write down, transfer, teach in a course — and tacit knowledge — things you can only demonstrate through doing, the kind that lives in experience, judgment, and accumulated context. The execution layer is mostly explicit knowledge. AI is extraordinarily good at explicit knowledge. The judgment layer is mostly tacit knowledge. AI cannot replicate tacit knowledge by definition, because tacit knowledge only exists in the act of applying it in context.

The 25% of AI projects that do deliver on their promised ROI share a common structure: they automate the explicit, repetitive, rules-based tasks while keeping humans in the judgment layer. The 75% that don’t deliver have tried to push AI further — into territory that requires tacit knowledge — and discovered the hard way that there’s a wall.

That wall is not a temporary limitation that better models will eventually eliminate. It’s a structural feature of what knowledge work actually is.

IV. What the boomerang tells you about your career — specifically

Here is the important thing about the AI boomerang data that most people miss.

The rehiring isn’t random. Companies are primarily rehiring experienced employees — people with 3 to 7 years of domain experience, people who carry institutional knowledge, people who have built the judgment layer over time. They’re not rehiring the entry-level roles they eliminated. Those positions, in many cases, are gone — replaced by a combination of AI execution and mid-level human judgment.

Forrester explicitly warns that laid-off domestic workers may not be the ones getting rehired. Many of those roles are coming back offshore or at significantly reduced salaries. The OrgVue study found that 34% of companies also saw existing employees quit as a direct result of AI implementation — compounding the talent loss. Customer satisfaction scores declined. Projects stalled. The institutional knowledge that was casually discarded proved impossible to replace quickly — not because AI couldn’t do the tasks, but because the judgment accumulated over years of exposure to specific clients, specific systems, specific organisational politics had no replacement.

What this means for your career, in plain terms:

The execution layer is getting cheaper and faster every month. If your career value is concentrated in the execution layer — in being the person who pulls the queries, builds the reports, runs the dashboards — your leverage is declining regardless of whether your specific job is at risk right now. The AI boomerang doesn’t save you if your value proposition is identical to what AI can now do for a fraction of the cost.

But if your career value is concentrated in the judgment layer — in being the person who knows which question is worth answering, who can read a business situation and translate it into an analytical problem, who carries the context about how this specific organisation makes decisions — your leverage is increasing. Because that’s exactly what companies are now paying to replace, at great expense, after discovering they can’t easily replace it at all.

PwC is specific on this: job growth is strongest among workers using AI tools to enhance their judgment — not workers positioned as AI alternatives. The skills commanding the highest wage premiums are domain expertise combined with AI fluency — not AI fluency alone.

V. The four things worth building — and why they compound

Abstract enough. Here’s what this looks like in practice.

1. Build institutional knowledge deliberately, not accidentally

Most analysts accumulate institutional knowledge as a side effect of showing up — over time, they learn how their organisation makes decisions, what leadership actually cares about, which stakeholders trust which sources, what questions lead to action and which lead to nothing.

Deliberate means: before every piece of analysis, write one sentence — “This will help [specific person] make [specific decision].” If you can’t complete that sentence, you don’t have enough information yet. Get it before you start. Every time.

The time you waste on inefficient reporting workflows is the time you could spend building this. Fix the workflows so you can invest the hours somewhere that actually compounds.

2. Use AI to compress execution — then reinvest the gap

The companies winning with AI aren’t using it to replace workers. They’re using it to make workers more productive — compressing time spent in the execution layer so workers can spend more time in the judgment layer.

This is the version that delivers the 3x revenue growth PwC documented. Not AI instead of people. AI plus people, with the ratio of execution to judgment shifting dramatically in favour of judgment.

Practically: let AI pull the data faster, build the first version, handle the reporting. Then use the time those things used to take to do what AI demonstrably cannot — reading the business context, sharpening the question, building the stakeholder relationship, developing the narrative that makes someone act on what you found.

Automating your data workflows frees hours. The question is whether you reinvest those hours into something that compounds.

3. Make your tacit knowledge explicit

Take the judgment calls you make every week — the ones so automatic you barely notice them — and write them down. Not the outputs. The reasoning. Why did you frame the analysis this way? Why did you push back on that stakeholder request? What made you sense that the model’s output was technically correct but strategically wrong?

That documentation is valuable for two reasons. First, it makes your implicit expertise visible — to yourself, to your organisation, to people making decisions about your career. Second, it forces you to articulate judgment processes that are currently tacit, making them teachable, reproducible, and demonstrable.

The companies now scrambling to replace institutional knowledge they lost are discovering that most of it was never written down. Your Power BI skills are not your career moat. The documented judgment behind how you use them is.

4. Build signal in public — before you need it

The companies most burned by the AI boomerang made one structural error: they had no way to see the judgment layer before they eliminated it. They could only see the execution outputs, which looked automatable. By the time they understood what else was happening, it was gone.

Making your judgment layer visible — before anyone asks for it, before a layoff forces the question — is now one of the highest-leverage career investments you can make.

This means writing. Not a portfolio of dashboards. Not a list of skills. Writing that shows how you think — what you noticed that others missed, what question you asked that changed the direction of a project, what you understand about your domain that isn’t obvious to people with less context.

This is what high-value analysts do differently from the ones who stay stuck producing output nobody acts on. Not harder work. Better questions, made visible.

The close — what actually protects you

The AI boomerang doesn’t mean you’re safe.

It means the thing that makes you safe is now clearer than it’s ever been — because the experiment that was supposed to disprove it just proved it instead.

The companies that fired people and are now rehiring them at higher cost learned an expensive lesson: the execution layer they automated was the cheap part. The judgment layer they lost was the expensive part. They just couldn’t see that until it was gone.

You don’t have to learn that lesson the expensive way.

Three things. Not ten.

Step 1 — This week, before every analysis, write the sentence: “This will help [person] decide [specific decision].” If you can’t finish it, stop and get the information. That one habit is the beginning of deliberately building the judgment layer.

Step 2 — Find one meeting you weren’t invited to and ask if you can observe. One hour. No agenda except to understand how decisions actually get made in your organisation. The context you absorb from that hour is exactly what the companies now paying to rebuild institutional knowledge would pay a premium for.

Step 3 — Write one honest observation publicly this week. Not a tutorial. Something you noticed, something you figured out, something you understand about your domain that took real effort to learn. That’s the beginning of a visible record of your judgment.

The AI replacement story was wrong. The AI transformation story is not.

What’s transforming is what creates value — and the data from 55,000 layoffs, 600 HR professionals, and billions of dollars of restaffing costs is now very clear about what that is.

It was always the judgment. We just needed a $3 trillion experiment to prove it.

I write about data, AI, and careers in plain language. Follow for more.