Somewhere in the last eighteen months, "we're restructuring for an AI-first future" became the socially acceptable way to say "we cut headcount and we're not totally sure why." Boards approved it. Analysts nodded along. Nobody asked the one question an accountant asks before signing off on any capital project: where's the return.
The data is finally catching up to the storyline, and it is not flattering. Roughly 95 percent of generative AI pilots inside large companies fail to produce measurable returns. Companies abandoned AI initiatives in 2025 at more than double the rate they did the year before. The average AI initiative burns close to seven million dollars and returns under two million. That is not a rounding error. That is a project you kill in the first review meeting, if you are running the numbers instead of running the narrative.
What does the actual AI ROI data show?
It shows a wide split, not a uniform failure. The projects that actually work are returning real money, some reporting ROI north of 180 percent. The projects that don't work are cratering hard enough to drag the average into negative territory. That spread is the whole story. AI is not universally good or universally useless as an investment. It behaves like every other capital allocation decision ever made: some bets pay off, most don't, and the difference is almost always in how rigorously the bet was underwritten before the money moved.
What's new is the confidence with which companies are skipping the underwriting. Only about one in five S&P 500 companies can point to a measurable AI benefit tied to their spend. And yet the spending keeps climbing, and the headcount keeps shrinking, and leadership keeps describing both as if they were the same decision.
Why are companies laying people off before they know if AI works?
Because "we're investing in the future" is a better sentence to say in an earnings call than "we don't have a model for what this actually saves us." Layoffs are legible. Wall Street understands a headcount number going down. Wall Street does not understand, and mostly does not ask, whether the software replacing those people is producing output worth what it costs to run, maintain, and correct.
This is the part that should bother anyone who has ever built a budget. You do not cut a cost center and simultaneously fund an unproven replacement and call that fiscal discipline. You call it a bet, and you size the bet like one. Instead, companies are treating "AI-first" as a mission statement rather than a line item, which means nobody is required to show their work. A mission statement doesn't need a payback period. A capital expenditure does.
The honest version of most 2026 restructuring announcements would read: we reduced fixed labor costs, redirected the savings into a technology whose returns we have not yet modeled, and we are calling this strategy. That sentence would not survive a board meeting. So it gets dressed up instead.
What is the token budget problem, and why does it matter?
Ask any company running serious AI workloads what happened to their usage bill this year and you'll get a wince before you get a number. Firms are blowing through annual AI compute budgets in a matter of months, then treating the overage as a surprise rather than what it actually is: an unmodeled variable cost that scales with usage in ways nobody bothered to forecast. One well-known company reportedly burned through half a billion dollars in AI spend in thirty days before anyone with budget authority noticed.
That is not an AI problem. That is a financial controls problem wearing an AI costume. No CFO would let a manufacturing line run without a unit cost model. No one would greenlight a marketing campaign with no cap on ad spend and no dashboard tracking it in real time. But hand the same organization a chatbot with a usage-based pricing model, and suddenly the guardrails that exist everywhere else in the business quietly disappear. The tool is new. The discipline required to manage it is not supposed to be optional just because the invoice looks different.
Here's the quotable part: what does this actually cost a company?
It costs trust, and trust is the one line item that never shows up on the balance sheet until it's gone. A layoff justified by a return you haven't measured yet isn't a strategy, it's a confession that you didn't want to explain the real one. Employees know the difference. So do the ones who survive the cut and now have to use the tool that replaced their colleagues, often without the training, the workflow redesign, or the realistic expectations that would have made that tool actually useful.
Every consultant currently selling "AI transformation" decks knows this pattern too. Fear of being left behind is a more reliable sales engine than a documented use case, which is exactly why so many of these initiatives get funded before anyone writes down what success looks like. If you can't state, in one sentence, what number moves and by how much, you don't have a business case. You have a vibe with a purchase order attached.
What does a defensible AI investment actually look like?
It looks almost boring compared to the press release version. It starts with a specific, narrow task that currently costs a known amount of money or time. It has a baseline measured before the tool touches anything. It has a usage cap tied to an actual budget owner who checks the number weekly, not annually. It has a kill criterion decided in advance, not negotiated after the sunk cost gets too large to admit to. And critically, it does not require firing anyone before the pilot proves out. If the tool is as good as the pitch, the savings show up on their own, and you make the staffing decision with real data instead of a forecast dressed up as certainty.
This is not a case against building with AI. Plenty of the tools genuinely work, and the companies running the disciplined version of this playbook are the ones pulling ahead with that 180-plus percent return. The case is against skipping the underwriting because the technology is exciting enough to make everyone forget that underwriting is the job.
What should workers and leaders actually take from this?
For workers: the layoff you're watching happen around you is very often not evidence that you were replaced by something better. It's frequently evidence that someone needed a headcount reduction to happen before the AI spend showed up as a line item they'd have to defend on its own. That distinction matters for how you read the moment, even if it doesn't change your immediate options.
For leaders: run the numbers before you run the announcement. If you can't produce a baseline, a unit cost, and a kill criterion, you don't have an AI strategy. You have an expensive hope, and hope has never once shown up favorably on a variance report. The technology is real and some of it is genuinely transformative. But "genuinely transformative" and "financially justified" are two different claims, and only one of them requires evidence. Start asking for both, and the decisions get a lot less embarrassing in hindsight.
