AI isn't a bubble. Most companies are just skipping the homework.

A few weeks ago the CEO of GitLab sent an open letter to customers that read, between the lines, like a warning shot.

The cost of running their platform had gone from tens of dollars per seat to hundreds, and he expected it to hit thousands. Nobody was gouging anyone. The agents just open merge requests in parallel now, kick off pipelines through the night, and push commits faster than any human team ever has.

GitLab used to sell you a seat. Now they sell you a task. Sit with that for a second, because it tells you exactly where this is going.

And it's not just them. As I write this, GitHub Copilot is switching to usage-based billing tomorrow — June 1. The flat premium-request model is out; every plan now comes with a monthly allotment of token-metered "AI Credits," and anything past that gets billed by actual token consumption. The base seat prices don't move — Business stays at $19, Enterprise at $39, each bundling that much credit back. Code completions stay unlimited. What changes is the expensive stuff: chat, code review, and especially agentic sessions, the workflows that chew through tokens. GitHub said the quiet part plainly — they're aligning Copilot with the metered pricing both Anthropic and OpenAI already moved to for enterprise. The token bill was always there. It was just hidden behind a subscription, and now it isn't.

If your team runs agents hard, this is the month the gym-membership pricing ends and the meter starts. (One detail that'll bite if you miss it: reviewing a PR with Copilot now also burns GitHub Actions minutes, on top of the credits.)

That letter is the backdrop for the question I get in pretty much every leadership channel and half my one-on-ones lately: is this a bubble? People want a yes or a no. The honest answer is more annoying than either, which is exactly why I wanted to write it down.

The case for "yes, obviously"

If you only read headlines, "yes" looks like a strong bet.

Writer's 2026 enterprise survey found that only 29% of companies see meaningful ROI from generative AI, and 23% from agents. Mercer's Global Talent Trends 2026 — about 12,000 people polled — found just 27% of CEOs said the return met or beat expectations, down from 38% a year earlier. And MIT's NANDA report is the brutal one: 95% of enterprise AI pilots showed no measurable impact on the P&L.

Three organisations that agree on almost nothing, all pointing at the same hole. Feels like a slam dunk.

It isn't, and here's the catch. Look at what those numbers actually count before you screenshot "95% fail" for your next all-hands. Writer's 29% is perceived ROI — executives reporting how they feel about the return. MIT's 5% is measured impact on the bottom line, a much harder ruler. They're not the same finding, and stapling them together is the kind of thing a sharp reader catches in about four seconds.

The careful version is less dramatic and more useful: by the soft measure, most companies aren't happy. By the hard measure, almost none can prove a dollar of return. Both true. Both about the same root cause.

So the cost is real and climbing — that part's settled. The interesting question is the other half. Where's the return going?

The money leaks somewhere between the dev and the business

Here's the bit that gets buried under all the doom: the individual gains are real. Measured, not vibes.

Google's DORA team has spent a decade studying what makes engineering orgs fast, and their one-line read on AI is the cleanest thing I've seen on the subject: AI doesn't fix a team, it amplifies whatever is already there. Strong teams get faster. Weak teams get faster at making a mess.

And honestly? I feel both sides of that depending on which machine I'm sitting at.

On my own side projects, where I control the whole setup, AI is a force multiplier and it isn't close. Good test coverage, clear standards, a tidy context for the model to work in — and it'll apply the same convention across 200 files without getting bored or sloppy. It writes more tests than I would on a tired Friday afternoon. The leverage is obvious because the foundation under it is solid.

At work, modernising a big legacy platform? Messier. And that's the whole point. The same tool that flies on a clean repo will happily hand me confident, plausible, subtly wrong code on top of fifteen years of decisions nobody wrote down. The model didn't get dumber between Saturday and Monday. The ground it's standing on did.

That gap is the entire story. The benefit lands on the person, and it lands hard — Writer's own data has super-users saving the better part of a day every week. But it doesn't climb the ladder and show up as a number the CFO can point at. The money leaks somewhere between the developer who feels ten times faster and the business that can't find the win on its spreadsheet.

The bottleneck is organisational, not technical

MIT calls that leak a "learning gap," and I think they nailed it.

The constraint isn't model quality. Everyone's working with roughly the same frontier models. The constraint is everything around the model — whether your processes were built for this, whether anyone actually redesigned the workflow instead of bolting a chatbot onto the old one, whether there's any governance for what agents can touch and what they cost.

This is the uncomfortable part for leaders, because you can't fix it with a credit card. Buying licences is the easy 5%. The expensive 95% is the work nobody enjoys: redesigning how work flows, deciding which judgement stays with humans, paying down the technical debt that makes AI dangerous on a legacy codebase, teaching people to use the tools well instead of handing them out and waiting for magic.

Most companies want the fruit without planting the tree. Then they act surprised when the tree didn't grow.

The trap inside doing it right

Here's a subtler one, and it's been on my mind because I've been close to a pilot lately. Not the licence-and-pray kind — a real one. Cross-functional teams, a fixed timebox, and an explicit brief to figure out how we work in the age of AI as we go. Genuinely the right instinct. Most orgs never get this far.

But there's a catch hiding in how these things usually get scoped, and it's worth naming because it's easy to walk straight into: the pilot runs on greenfield work. New projects, clean slates, no fifteen years of history pushing back. And of course it does — greenfield is where everyone wants to start, because the barrier to adoption is so much lower. The AI has nothing to fight. It flies.

You see the problem. The day-to-day for most engineering teams isn't greenfield. It's brownfield — the legacy platform, the service nobody fully understands anymore, the codebase where every change has three invisible dependencies. That's where the real work lives, and it's exactly where AI gets hard. So if your experiment only ever runs on the easy ground, you risk learning a flattering lesson that doesn't survive contact with the actual job. "AI works great" — right up until you point it at the thing you actually spend your week on.

I don't think this sinks a pilot. But I'd want at least one team wrestling AI into a brownfield mess from day one, precisely because that's where the lesson is. The clean-slate result is the one you can already predict. The messy one is the one you're actually paying to find out.

The most expensive mistake is leading with layoffs

If there's one move I'd talk any leader out of, it's treating AI as a headcount story before it's an output story.

The data here has gotten hard to wave away. Forrester found 55% of employers who made AI-attributed cuts in 2025 already regret them, and expects about half those roles to get quietly rehired — often at a premium. The reason never makes the press release: institutional knowledge doesn't transfer to a model through a deployment. Let go of the person who understood the edge cases and the why we do it this way, and that context is gone. It's not in any training set. You can't fine-tune your way back to it.

It gets blunter. At the India AI Impact Summit this February, even Sam Altman admitted what a lot of us already suspected — a good chunk of the AI layoff wave is "AI washing," companies pinning cuts they were going to make anyway on the technology because it sounds better. He didn't coin the term; it's borrowed from greenwashing, and it fits like a glove.

Which gets at something worth saying plainly: cutting people before you've built a way to make money from the tools isn't an AI strategy. It's a cost cut wearing an AI costume. If you want to find the actual froth in all this, it's not the technology — it's that.

What it looks like when it actually works

The companies in that 5% aren't running better models than you. They did the boring work, and there's a clear shape to it.

They reinvested the time instead of cutting it. DORA's explicit advice is to reclaim the capacity AI frees up and point it at the rework and the backlog — not shrink the team and book the saving. They budgeted for the dip, too. DORA describes a "J-curve": a stretch where things get slower and buggier before they get better, because people are learning and your seniors are spending real hours reviewing AI output. Don't plan for that tuition cost and you'll panic and pull the plug right before the payoff lands. And they built the harness first — tests, standards, good context, sane architecture. The exact same things that make AI fly on my side projects are what turn it from a liability into a multiplier on a real codebase.

The payoff is worth it, but it is patience. In DORA's modelling a serious rollout takes around eight months to pay back and lands somewhere near a 39% first-year return — and they're careful to call that a high-uncertainty estimate, not a promise. Eight months of investment in people and process before the curve turns up.

That's not a quarter. That's a strategy.

Why I'm still not calling it a bubble

Strip away the noise and the base case is intact.

The individual value is real and it's measured. The market for AI coding tools more than doubled in two years — roughly $5 billion to nearly $13 billion — which is not how a popped bubble behaves. People pay for these tools because they get something back. I use them every day, I think they're expensive, and I still wouldn't give them up. (You? Be honest.)

There's history on this side too. When a technology gets cheaper and more efficient, demand tends to grow and swallow the savings rather than shrink. We've watched it happen before. And the loudest doom predictions are already softening — over the last month both Dario Amodei and Sam Altman walked back their earlier warnings about AI gutting white-collar work, with Amodei now reframing automation as something that expands what people do rather than deletes it. Worth a pinch of salt given both their companies are eyeing IPOs — but the direction is telling.

So no. I don't think the bubble's about to pop.

I think we're on an adoption curve, most people are climbing it badly, and the gap between the 5% and everyone else is only going to widen. The teams that treat AI as a redesign of how their people work will keep pulling away. The ones who treated it as a licence to skip the hard part will keep paying the bill and wondering where the return went — and at some point they'll announce the whole thing never worked, when what actually happened is they never did the homework.

That's the version of this story I wish more people in leadership were telling.

If you're living it, I'm genuinely curious which side of the gap your team is on right now — is AI amplifying your work, or just amplifying your mess? Tell me in the comments.

Sources

• MIT Project NANDA, The GenAI Divide: State of AI in Business 2025 (July 2025)
• Writer, Enterprise AI adoption in 2026 (2026)
• Mercer, Global Talent Trends 2026
• GitLab, "GitLab Act 2" open letter from CEO Bill Staples (2026); reporting via CIO / InfoWorld
• Forrester, Predictions 2026
• Google Cloud DORA, State of AI-assisted Software Development (2025) and ROI of AI-assisted Software Development (2026)
• Fortune, reporting on Sam Altman ("AI washing," Feb 2026) and the Altman/Amodei walkbacks (May 2026)