AI's "zero payoff" isn't a model problem. It's how companies work and count

If AI were really failing to deliver, would banks be reporting double-digit gains while an Apollo economist reports nothing? Both things are circulating right now, and the contradiction says more about how companies work than about how good the models are.

The trigger is Torsten Slok, Apollo's chief economist, whose "zero financial payoff" warning is making the rounds. As Seeking Alpha summarizes it, AI has not improved productivity or profit margins for most companies outside the tech sector. PCMag notes that revenue per employee at the "Magnificent 7" has risen by around 30% over four years. Set that against the rest of the economy, where he sees no such payoff, and the gap is the story.

The tool was bought, the job wasn't changed

Our reading, and it is an interpretation, not Slok's stated diagnosis: a company that buys a chatbot license and leaves the workflow untouched has bought a faster typewriter. The savings land in a few minutes per employee, dissolve into the day, and never reach a margin. Big tech sees returns because its work is already software, already instrumented, and easy to rebuild around the new tooling. A mid-size insurer or retailer has to redesign approvals, handoffs and responsibilities first.

There is some evidence for this in the documentation, though from a vendor-commissioned study. Protegrity cited research from Enterprise Management Associates, and SiliconANGLE reports that 83% of leaders saw project delays and 82% pushed agents into production in a diminished state, stripped of the tools and data that would make them useful. Protegrity's own Diego Loureda put the governance issue well: companies approve projects, but "they have to approve patterns." Treat the numbers with the caution any sponsor-adjacent survey deserves. Still, the picture is a familiar one: the agent isn't dumb, it's been handcuffed by a process built for a slower era.

You can't see a payoff you never set up to measure

The second half of the problem is accounting. Apollo's own research page concedes that academic papers are inconclusive on AI's productivity impact, while arguing we're still early in a labor-productivity boom, with clear impacts in specific sectors. That sounds contradictory until you notice the unit of analysis. Aggregate profit margins are a blunt instrument. Gains inside one team's turnaround time can vanish before they reach the income statement, especially if nobody decided in advance whether the freed hours become lower costs, more volume or new products.

That is probably why a headline like "banks report 10-50% gains in productivity, turnaround time and costs" can coexist with "zero payoff." We only have the headline from Economic Times, not the methodology, so we won't lean on it. But a gain in cycle time inside a function is measurable. A gain in company-wide profit requires someone to bank it. Most firms haven't built that muscle.

Even the vendors hedge. Adobe, which reports AI-first annualized recurring revenue growing more than 150% year over year, also warns in its filing that customers may not adopt or pay for its AI offerings as expected. When the seller lists adoption as a risk, the buyer's ROI question is not paranoia.

Why the "bubble" reading is too simple

It's tempting to take Slok's warning as proof that the models are overrated. We think that's the weaker conclusion. Nothing in what we reviewed shows the models failing at tasks; it shows organizations failing to absorb them. That distinction matters for investors, too. A live-trading newsletter tallies Amazon, Alphabet, Microsoft and Meta's 2026 capital spending plans at roughly $720 billion to $740 billion, a figure we'd treat as that outlet's own tally. If the bottleneck is process rather than capability, then more capacity doesn't fix the payoff gap by itself. Apple's CEO John Ternus said he doesn't want to ship AI "for AI's sake," and that line reads better as a management principle than as a product slogan.

What to do with this

If you run a team, don't ask whether the model is good enough. Ask which workflow you'd redesign if the tool were free, who owns the saved hours, and what number on which report should move. Pick one process, rebuild it, and measure it before and after. If you're reading the "zero payoff" headline as an investor, read it as a measurement-and-adoption warning, not a verdict on the technology.

Our verdict: the ROI problem is real, but it is mostly ours. Companies that treat AI as a purchase will keep seeing zero. Those that treat it as a reason to rewrite how work flows, and count the result honestly, are the ones that will make Slok's chart look dated.