Most companies are investing in AI. Very few can show business results. The answer isn’t the technology — it’s what those few asked about themselves before the technology ever arrived. The case, made in the open, from primary sources.
Start where the evidence actually is, not where either camp wants it to be.
Enterprise AI spending is at record levels — there is hardly a board agenda in the world without it. And the measured result, from the research that has looked hardest: MIT found 95% of enterprise AI initiatives producing no measurable impact on the P&L. It isn’t one study — RAND puts AI project failure near 80%, four times the rate of conventional IT, and abandonment is rising, not falling, even as the models improve.
The failure isn’t timidity, and it isn’t budget. Some of the largest AI spends on earth sit inside that 95%. Uber burned its entire 2026 AI-coding budget — about $3.4bn — in four months, its CTO saying plainly, “the budget I thought I would need is blown away already.” Microsoft wound down most of its internal coding-AI licences as costs outran budget within months. When the most technical organisations in the world overspend with little to show, the missing ingredient is clearly not the technology.
It would be easy to read all this as a verdict on AI. That’s the mistake half the commentary makes — and it’s wrong, because the same period produced the opposite result in plain sight:
Both pictures are real at once. A handful of institutions are getting unmistakable value; almost everyone else is spending heavily and showing little — they can show the spend; they cannot show the result. And the gap isn’t closing as the technology improves — which is the tell. If better models were the answer, the 95% would be shrinking. It isn’t. So the difference lies somewhere other than the technology.
If 95% are failing and a handful are winning a billion dollars at a time, then “does AI work?” cannot be the question. It demonstrably works. And “is the technology ready?” cannot be it either — because the institutions that fail and the institutions that succeed are using substantially the same technology.
So the defining management question of this era is not whether AI works. It is:
That question has an answer. It is visible in the winners’ own record, it is consistent from case to case, and — this is the part that matters for everyone currently in the 95% — it is structural, not technical. Which means it is available to any institution willing to ask a harder question first.
Here is the answer, stated plainly, and it is the argument this whole page prosecutes:
The difference is not a matter of degree. It is two different activities that happen to use the same tools.
Select the tool, install it, run a pilot, scale what works. It is a sensible-sounding plan, and it produces an enormous amount of activity — pilots, demos, proofs-of-concept, board decks full of promise. What it rarely produces is value, because it is adding intelligence to a business that was designed for a world without it. At its best, it automates yesterday’s answers faster.
It starts not with the tool but with the work — what the data must support, where decisions should sit, what jobs become, what happens to the people doing them — and then lets the technology do what only it can. It is slower to show motion and far more likely to show results, because the institution has been rebuilt to receive them.
The deepest version of the distinction is the one in the standfirst: the few asked something about themselves before the technology ever arrived. They asked whether the problems they solve for customers today will still be the problems customers have tomorrow — and they rebuilt around the answer. That question has nothing to do with AI. It is the question AI makes unavoidable.
The argument is only worth as much as the cases that test it. Here are five, each from the institution’s own record — including where it went wrong. Every figure below is read from a primary source; none is an analyst’s paraphrase.
And if you ran a pilot that never scaled, read these as an ally, not an indictment: that wasn’t a failure of effort or intelligence — it was the predictable result of the deployment frame, which is exactly why it is addressable.
The most complete success in the set, and the clearest illustration of architecture before tools. Singapore’s DBS attributes over S$1bn in value to AI in its own published results. But the figure isn’t the lesson — the sequence is. DBS spent years rebuilding its data foundations around what banking was becoming, not what it had been, before any of the results that made headlines. That sequence — foundations first, value second — wasn’t caution. It was the strategy. Most institutions discover the data-readiness problem in month nine of a project; DBS treated it as the precondition.
JPMorgan’s own Investor Day deck (not the press coverage of it) discloses the numbers and the method: +35% and +65% value uplifts on measures it defines in its own footnote — value meaning revenue gained, cost lowered, or cost avoided, measured against prior analytical techniques.
But the detail the summaries lose is the one that matters most. JPMorgan’s AI value slide splits the benefit into two columns — revenue generation and cost & risk efficiency — under an explicit banner: shrinking expenses is good, but growing profits is better. The most AI-advanced bank in the world organises its AI value around revenue ahead of cost, and can tell you which is which, because it decided which should be which. (And it built on a data foundation laid first: more than 90% of its analytical data was in the cloud before the value was claimed.)
Where DBS and JPMorgan are top-down builds of proprietary platforms, BBVA reached the same destination bottom-up, on a third-party tool — which makes it the more instructive case, not the redundant one. It put ChatGPT Enterprise in the hands of the people in front of the problems, and those employees built more than 20,000 custom GPTs themselves. It scaled on demand, not mandate — 3,300 licences, then 11,000, then a rollout toward all 120,000 employees — with measured results: hours saved per user per week, a large majority using it daily.
The lesson: the differentiator was never the tool. It was the operating model around it — senior leaders trained hands-on first, security and compliance made partners from day one, a champion community across the bank. Two opposite routes, the same structural markers. Convergence from opposite directions is stronger proof than a third identical case would be.
The case that proves the gap runs both ways. Klarna cut deep into human service on an AI-efficiency thesis — its AI assistant doing the work of an estimated 700 agents, headcount falling from roughly 5,500 to 3,400. The pre-IPO story was built on it. Then the same CEO publicly reversed: cost had become, in his words, too predominant a factor; quality had fallen; the company began rehiring people, insisting there would always be a human for the customer who wanted one. Read correctly, the reversal isn’t the failure — it’s the system working. The failure mode is the institution that gets the same signal and buries it in a success deck. Klarna measured, learned what the technology was actually for, and said so in public. That second kind of honesty is rarer than the technology — and worth more.
The proof that the gap is not a verdict. Upstart was an AI-lending story that broke hard when the regime changed — and then did the harder thing: it rebuilt, rather than retreating from the technology or pretending the fall hadn’t happened. The crossing goes both ways. An institution in the 95% today is not sentenced to stay there.
There is one thing present in every durable case and missing from most of the commentary: deliberate work on the people side — and it begins with the leaders, not the workforce.
The reset tells a leader that what made them successful over thirty years must now be unlearned. That is an identity-level demand, not a skills update. And AI sharpens it — compressing the organisation into fewer layers, where authority comes from present capability rather than tenure, and a leader owns a whole rather than a part.
A leader who has not faced that shift in themselves cannot carry anyone else through it. The work starts at the top, or it does not hold.
Then the workforce — where the felt reality decides whether people help you or quietly resist you. An AI transformation asks people to help build the thing that may make their own role redundant; quiet resistance is the rational response. Underneath it sit three losses, real and compounding:
All three land at once — which is why the industry comfort, “we’ll reskill everyone, no one left behind,” fails. It isn’t always true, and people know it. Not everyone will cross to the other side, and saying so is the beginning of respect. Comfort an institution can’t guarantee breaks trust exactly when it is needed most. People move toward a hard future they understand; they freeze before one hidden behind reassurances they don’t believe.
So the people side is architecture, not sentiment. Where decisions sit, what work becomes, who is carried and who leaves, how those who leave are treated — these are structural choices that decide the P&L as surely as the data model does. No architecture survives the people it ignores.
Klarna proved it. It cut deep into human service on an efficiency thesis — its assistant doing the work of an estimated 700 agents — then reversed in public, its CEO conceding the firm had “underappreciated the human aspects of service delivery,” and began rehiring. The institutions that endured did the opposite from the start: they carried their leaders and their people through the change, rather than discovering too late what they had cut.
The gap is widest where the noise is loudest — and few markets are louder than India. But the deeper reason is structural. India built a services economy on labour-cost arbitrage: doing the world’s back-office, IT and BPO work more cheaply — around six million jobs, ~7% of GDP. The cruel symmetry of this moment is that the traits that sent that work to India — repetitive, documented, scriptable — are exactly the traits that make it the first work AI removes. The arbitrage that was the strength is now the exposure.
This is not a forecast; it is in the sector’s own numbers. TCS’s revenue rose 3.7% in the same quarter its headcount fell by nearly 20,000 — the steepest quarterly drop in its history. Revenue rising while people fall is the “more revenue needs more heads” model — the model the whole services economy was built on — breaking in the open.
The honest nuance, also in the data: Infosys added people over the same period. The sector is splitting — commodity tier exposed, capability tier still growing — not uniformly collapsing.
The harder truth underneath: the services era never became an ownership era. For two decades the region ran other people’s processes, on other people’s platforms, on other people’s IP — earning a margin on labour, not a return on what it owned. AI reprices labour toward zero, and an economy built on arbitrage rather than ownership is the most exposed to that repricing.
The markets are already saying so: foreign holdings in Indian IT have fallen by nearly half — through capital exiting and being repriced down — as the AI trade pulls global capital toward the hardware economies India’s index cannot offer.
That is the stake, and the question this page asks is no abstraction here — it is the difference between climbing and being repriced. The same architecture-not-deployment choice that decides whether one AI programme pays is, at national scale, the choice between moving up to capability and ownership, or defending an arbitrage that intelligence-in-abundance is erasing.
This argument isn’t built to sell a conclusion. It is built to equip a question — one that every leadership team should be able to answer precisely, and most cannot:
Underneath it sit the two axes the whole argument turns on:
Has someone senior actually asked what the business should become — not just which tools to deploy — and does the organisation agree on the answer?
If the question were answered well, is the institution — its data, its decisions, its operating model, its people — actually built to execute it?
Most AI programmes are running hard on the second axis while no one owns the first. That is the shape of the 95%.
The argument turns on two questions: whether your institution owns the right question, and whether it could deliver on the answer. The Assessment gives you an honest position on both — not a score to publish, but a reading to think with. It takes about six minutes. It’s yours to run, yours to keep, and built to be taken into the room that decides.
For boards and executive teams: a working session on what is actually happening with AI, how other companies are embracing it — what is working and what is not — and the priorities your organisation can set. The session is complimentary.
Built to be carried into your institution. Forward the argument; table the questions; take the reading to the room that decides.
No intake process, no gatekeeper. A note reaches me directly, and I reply myself — usually within a day or two.
It comes straight to my inbox, and I reply personally — usually within a day or two.
Your note is with me. I’ll reply personally, usually within a day or two, with possible timings.