The 20% Who Actually Make Money on AI

Table of contents

The money is real — and almost nobody has it

Start with the uncomfortable distribution. PwC’s 2026 AI Performance Study interviewed 1,217 senior executives, mostly at large public companies, and found that 74% of AI’s measured economic value is being captured by just 20% of organizations. The companies PwC calls “AI-fit” are posting AI-driven financial performance 7.2 times higher than everyone else in the survey.

Hold that next to the failure data and the shape gets sharper. MIT’s now-famous study, reported by Fortune, found that 95% of generative-AI pilots produced no measurable P&L impact at all. Not a loss, not a scandal — just nothing. Billions spent, needle unmoved.

Put the two studies in the same sentence and you get the real state of play in 2026: AI absolutely produces profit, and the overwhelming majority of companies spending on it are not the ones receiving it. This is not a technology that lifts all boats. It’s a technology that has, so far, lifted a fifth of them and left the rest paying for the dock.

That should change the question every leader is asking. The interesting question was never does AI deliver ROI — the honest answer is “spectacularly, for a few.” The question is what does the winning fifth do that the failing 95% don’t. And here the data is unusually specific.

Trait one: they chase growth, not savings

This is the through-line, and Klarna is the cautionary version of it. The losers treat AI as a cost-reduction program. The winners treat it as a growth program.

PwC’s leaders weren’t simply deploying more tools than their peers. They were 2.6 times as likely to say AI improved their ability to reinvent their business model, and two to three times as likely to use AI to chase new revenue opening up as industries blur into each other. The framing matters more than it sounds. A company that asks “where can AI cut headcount” runs a bounded search — there’s a floor, and you hit it, and Siemiatkowski’s “lower quality” is what’s waiting at the bottom. A company that asks “what can we now sell that we couldn’t before” runs an unbounded one.

MIT found the same asymmetry from the failure side. More than half of generative-AI budgets were pointed at sales and marketing, but the biggest realized ROI came from back-office automation — the unglamorous work of retiring a business-process-outsourcing contract or collapsing an agency relationship. The winners spent where the value was, not where the demo looked best. Cost savings, in other words, are real and worth having — but they’re the floor of a growth strategy, not a strategy by themselves. The 95% bought the floor and called it a building.

Trait two: they rebuild the workflow, not just buy the tool

The single most actionable number I found is this one. Companies that paired their AI deployment with redesigned workflows and explicit KPIs got 2.7 times the ROI of companies that bolted the same tools onto their existing process — an average uplift of 34.6% versus 12.8%.

That ratio is the quiet refutation of the whole “just give everyone a license” theory of adoption. The model is not a faster version of your current process; dropped into your current process, it mostly produces the 95%‘s nothing. The value shows up when the process bends to fit the tool — when someone redraws the steps so the AI is doing work that compounds instead of work that’s merely adjacent to what a human used to do.

This is the problem I spend my days on, so take what follows as the bias it is: Grove is a workflow-orchestration engine built on exactly this premise — expressing AI work as an explicit graph of calls, tools, and data steps, so the process is something you design and version rather than something the model improvises inside a chat box. The bet is that the 2.7× lives in the wiring, and the wiring deserves real infrastructure.

MIT put the same finding in a blunter form: generic tools like ChatGPT “excel for individuals because of their flexibility, but they stall in enterprise use since they don’t learn from or adapt to workflows.” The flexibility that makes the tool feel magical to one person is exactly what makes it inert to an organization that won’t change shape around it. This is the same comprehension-and-integration cost I wrote about in The Speedup Trap — the easy part (generating output) got cheap, and the hard part (fitting it into the specific shape of your system) is where the gains actually live or die. The winners pay that cost on purpose. The losers assume the license fee covered it.

Trait three: they buy from specialists and partner smartly

The most counterintuitive finding, and the one most likely to sting an ambitious engineering org: in MIT’s data, buying AI from specialized vendors succeeded 67% of the time, while building it in-house succeeded only about a third as often. MIT noted this is “particularly relevant in financial services and other highly regulated sectors,” where prestige and caution both push firms toward proprietary builds — and where that instinct is quietly halving their odds.

The startups MIT watched go “from zero to $20 million in a year” did it with a formula that reads like an anti-platform manifesto: “they pick one pain point, execute well, and partner smartly with companies who use their tools.” One pain point. Not a transformation initiative. Not a center of excellence. A single, sharp, well-chosen problem, solved by someone who does only that, integrated by a customer willing to actually wire it in.

I want to be careful not to over-read this — “always buy, never build” is too clean, and plenty of the documented wins are bespoke. General Mills credits an AI-driven supply-chain system with more than $20 million in savings since fiscal 2024; Salesforce’s internal legal-ops agent cut over $5 million in spend; Unilever’s AI hiring platform halved time-to-fill. Those are built or deeply customized systems, and they work. But notice the shape even there: each is one pain point, owned end-to-end, with the workflow rebuilt around it. The “build” that fails is the open-ended platform play. The “build” that works is indistinguishable from buying a specialist, except the specialist is internal.

The cautionary tail is part of the data

Here’s the discipline the success stories demand: you have to read the walk-backs as part of the same dataset, not as a separate genre of “AI fails” content.

Klarna is the obvious one — the savings were real and the strategy was wrong, and both halves are true at once. The Commonwealth Bank of Australia version is starker: it cut 45 customer-service roles to a bot, claimed a reduction of 2,000 calls a week, and then reversed the layoffs within weeks when call volumes went up instead of down. The number that justified the cut evaporated under contact with reality.

What these have in common is not that AI failed. It’s that each company optimized the one variable that’s easy to put in a press release — headcount removed, dollars saved — and discovered the bill for everything that variable was quietly load-bearing: brand, escalation capacity, the complex cases the bot couldn’t hold. They were cost-first deployments, and cost-first is precisely the trait the PwC winners don’t share. The cautionary tail isn’t the opposite of the success stories. It’s the success stories run by someone who measured only the floor.

This is also why I’d treat any single vendor-sourced ROI figure with a raised eyebrow. The honest signal in this space isn’t one company’s heroic number — those get announced at the top of the arc and revised quietly at the bottom. The honest signal is the distribution: a fifth of companies, consistently, doing four or five specific things, pulling 7x ahead. That’s a pattern you can act on. A 853-agents headline is a slide you can get fired by.

What the winners’ profile actually says

Strip it down and the 20% are running a playbook that’s almost the photographic negative of default enterprise adoption:

• They aim AI at new revenue, and treat cost savings as the floor — not the goal. The unbounded question beats the bounded one.
• They rebuild the workflow around the tool and accept that this is the expensive, slow, un-AI-shaped part where the 2.7× actually comes from.
• They solve one sharp pain point with a specialist — bought, or built so narrowly it might as well be — instead of launching a platform.
• They put the gains where the value is (often the boring back office), not where the demo dazzles.
• They measure the whole job, which is why they don’t have to walk anything back six months later.

None of that is about having better models. Everybody has the same models — that’s rather the point of a foundation model. The winners aren’t winning on technology. They’re winning on the decidedly old-fashioned managerial work of pointing a capable tool at the right problem and reshaping the organization to absorb what it produces. The model is necessary and radically insufficient. The 95% proved the necessary part is cheap and the sufficient part is everything.

So when someone asks whether anyone is really making money on AI, the answer is an unambiguous yes — and the more useful follow-up is which of the five traits above your own company actually has. Because the gap between the 20% and the 95% isn’t a gap in tools, budget, or ambition. It’s a gap in everything that happens after you sign the license. The companies on the right side of it figured out the least exciting truth in technology: the breakthrough was never going to do the hard part for you. It just made the hard part the only part left that matters.