If you work in tech, AI feels like it's everywhere. It writes code, reviews pull requests, generates images, edits video, and answers support tickets. Every week there's a new model, a new tool, a new demo that makes last month's look old.
So it’s a little jarring to read what Torsten Slok, the chief economist at Apollo, published this week: outside the tech sector, AI has delivered essentially no financial payoff to corporate America.
Not a small payoff. Not a slower one. According to his charts, basically nothing yet.
Two economies, one technology
Slok splits the S&P 500 into two groups. On one side are the "Magnificent 7", the giant tech companies at the center of the AI boom. On the other side are the remaining 493 companies: banks, retailers, manufacturers, insurers, healthcare firms, the ordinary businesses that make up most of the economy.
For the big tech companies, revenue per worker is rising and profit margins are going up. For the other 493, both lines are flat. And across the wider economy, there's still no visible sign of AI lifting productivity growth.
This isn't the only data pointing in that direction. Forrester recently projected that AI will make up a fifth of corporate technology spending in 2027, while more than half of IT leaders are expected to miss their return-on-investment targets. In their survey, only about one in five AI decision-makers said AI had improved their company's operating profit over the past year.
So the money is flowing. The results, outside tech, aren't.
And according to Slok, companies are starting to notice. Instead of seeing AI spending as a path to growth, many are beginning to treat it as a cost that eats into margins.
Why AI shines in some places and stalls in others
The split makes more sense when you look at where AI works well.
Coding, images, and video have something in common. The work is digital from start to finish, and the output is easy to check. Code either passes the tests or it doesn't. An image either looks right or it doesn't. You can generate ten versions, throw away nine, and keep the one that works. Mistakes are cheap and fast to catch.
Most business work isn't like that.
Deciding whether to approve a loan, how to price a contract, which supplier to trust, or how to handle an upset customer depends on context, relationships, history, and judgment. There's often no test suite. You may not know if a decision was good until months later. And when it goes wrong, someone has to take responsibility.
Slok makes a similar point: AI may be a great calculator, but in business, intelligence is mostly judgment. And things like trust, culture, accountability, and fairness are what ultimately drive revenue, yet they're exactly what AI doesn't understand.
In other words, AI has mastered the work that's easy to check. Business mostly runs on the work that isn't.
So is this a bubble?
This is where opinions split hard. Both sides have real arguments, so here they are as fairly as I can put them.
The case for a collapse
The spending is enormous, and it's concentrated in a small number of companies. Much of the AI economy is circular: AI startups spend investor money on tokens from AI labs, which spend on chips and cloud capacity from big tech, which reports the growth. If the companies at the end of that chain, the ordinary businesses, aren't seeing returns, the whole structure rests on expectations rather than results.
There's also a pattern here. Companies under pressure to show returns quickly may cut AI budgets once the novelty wears off. Forrester found that nearly half of AI decision-makers expect an investment to pay off within a year. If that doesn't happen, patience could run out fast.
And history has examples of exactly this. The dot-com boom saw massive investment in a genuinely transformative technology, and it still ended in a crash that wiped out most of those companies.
The case for a slow payoff
The other side points to history too, but a different part of it.
In 1987, the economist Robert Solow famously observed that computers were showing up everywhere except in the productivity statistics. It took more than a decade before they clearly did. The reason wasn't that computers were useless. It was that companies first had to rebuild how they worked around them. New processes, new roles, new ways of organizing information.
Forrester's analysis points to the same issue today: many companies bought AI tools without changing their workflows, so the gains stay trapped in isolated pockets.
And even the dot-com crash didn't kill the internet. The bubble burst, but the technology went on to reshape almost every industry, just later and through different companies than investors expected.
On top of that, AI models keep getting cheaper. Slok himself raises the question of whether ordinary businesses can turn increasingly cheap intelligence into real growth. If costs keep falling, the math could change for companies that can't justify the spending today.
What the data can't tell us yet
It's worth being honest about the limits here. Slok's note is a short piece with a handful of charts, not a full study, and Apollo is an investment firm with its own view of the market. Productivity is notoriously hard to measure, and early gains can hide in places that don't show up in revenue per worker or profit margins.
At the same time, "it's just early" is an argument that can be stretched forever. At some point, a technology this expensive needs to show results outside the companies selling it.
What the data does show clearly is the gap. AI is booming in tech, and it's flat almost everywhere else.
The question that matters
Slok ends his note with a line that I think frames this better than any hype or doom headline: the AI story will be decided in the 493 ordinary companies, not the 7 giants.
Whether that story ends in a crash, a slow transformation, or something in between depends on whether businesses outside tech can find a way to turn AI into real value, not just more spending.
I don't think anyone knows the answer yet. So I'll put it to you: when you look at your own company, or the businesses around you, are you seeing real returns from AI, or mostly real costs?
Top comments (19)
I opened both sources, because the paragraph leans on three numbers at once.
The Apollo half holds. Slok's own wording: "Revenue per worker is rising for the Mag 7 but is flat for the S&P 493. Profit margins are rising for the Mag 7 but are flat for the S&P 493. And there are still no signs of AI boosting productivity growth." Nothing in this post overstates him.
The Forrester paragraph is where it drifts. Computerworld's brief credits "more than half of IT executives are expected to miss their return-on-investment targets" to The Register, and that is a forecast, not a survey finding. The 22% figure is a survey finding, and those respondents are "AI decision-makers", not IT executives. So one sentence joins a 2027 spending projection to a 2026 survey answer, across two different populations. The numbers are right; "in their survey" is carrying weight it cannot hold. Worth one clause, too: the chain is Forrester to The Register to Computerworld to you.
One method note, in the direction your post already argues: a flat aggregate line bounds the visible payoff, it does not measure the technology's usefulness. "Helped some, hurt others, netted out" draws the same flat line. Slok's "so far" is the honest part of it.
Thanks for checking the sources. The Apollo numbers hold, as you said, and they carry the argument. The Forrester line already says the ROI figure is expected, and the survey number stands on its own. Both point the same way. And a flat line isn’t a verdict, which is why the post ends without one.
Fair on the wording. "Expected" is your sentence, so that half of my read does not stand.
The half that does: those two numbers are not two sources. Same survey population, AI decision-makers, meaning the people already running the tools they are reporting on. Agreement between two questions asked of one group is not corroboration. And 22% operating profit from that group is a self-report about their own organisation, not a company-wide measure. One population, quoted twice.
Apollo's line still carries the post, as you said. That is the part I would keep.
Fair, that’s a genuinely sharper point. Two answers from the same group, about their own tools, aren’t independent evidence, and self-reported profit impact has obvious limits. That’s why Apollo’s line matters more: it’s measured from company results, not opinions. Forrester adds colour, Apollo carries the weight. Thanks for pushing it.
This is a fantastic teardown of statistical sleight of hand. Tracing the chain (Forrester -> The Register -> Computerworld -> Article) and catching that the 2027 forecast was mixed with a 2026 survey of a totally different population ("IT execs" vs "AI decision-makers") is exactly how you debunk hype-driven reporting.
The note about the flat aggregate line is also the smartest part of the whole debate—a net-zero productivity gain doesn't mean AI is useless, it just means the gains and losses are canceling each other out right now.
A fair point about wording, not sleight of hand. Both figures point the same direction, and Apollo’s flat line holds. Whether gains and losses are canceling is the open question.
Here is a short, balanced reply you can post to James Anderson:
Fair enough. The "canceling out" theory makes the most sense. For every company genuinely streamlining a workflow, another is just burning budget on tools that don't fit their processes. The flat aggregate line isn't hiding a failure; it's just hiding the winners.
Canceling out is one possible reading, but the flat line can’t tell us that. It looks the same whether winners and losers balance each other or whether almost nothing is happening. Until someone shows the winners in company results, “hiding” is a guess, not a finding.
You're absolutely right, "hiding" is speculative. The flat aggregate line is a black box—it can't distinguish between zero sum (winners canceling losers) and zero activity.
The only way to know is to look at the variance inside the 493. If we see a bimodal distribution (a few big winners, many flat), it's the Solow paradox playing out. If it's just a single flat line across the board, then AI just isn't translating to operational value yet. The macro metric is just too blunt to tell the difference.
Agreed, variance is the real test. A few clear winners would suggest the gains are real but uneven, while a uniformly flat spread would point to no payoff yet. Sector breakdowns would help here too. If anyone has seen dispersion data for the 493, I’d genuinely love to read it.
Sector breakdowns would definitely paint the clearest picture. I suspect the aggregate flat line is masking a massive divide: a few legacy companies are successfully restructuring their workflows around AI, while the vast majority are just buying licenses and bolting them onto old, broken processes. Without that dispersion data, we're all just guessing. If you find a good source breaking down the 493, definitely share it!
That divide sounds plausible, and it matches Forrester’s point about bolting AI onto old processes. Until the dispersion data shows it, it’s a hypothesis. I’ll share anything solid I find.
Sounds like a plan. The 'bolting AI onto old processes' problem is definitely the core issue—it's just expensive duct tape until the workflow itself is redesigned. Looking forward to seeing the data if you find it!
Expensive duct tape is the perfect phrase for it, and I’ll share the data the moment I find something solid.
Looking forward to it! It's a fascinating topic, and real data on it is surprisingly hard to find. Cheers!
The "easy to check vs. judgment" distinction is exactly why AI feels like magic in a dev environment but a money pit in an ops department.
great insights
😊
Thanks! The gap between shipping code and making real business judgment calls is definitely the biggest hurdle right now.