Has AI Actually Delivered? A 2026 Productivity Reality-Check

The short version

On July 6, 2026, Apollo Global Management's chief economist Torsten Slok warned that AI has not delivered the productivity gains its valuations assume, and that markets could face a "painful repricing" as a result. His framing was precise: "The key issue is the length of the ROI runway outside the tech sector."

He is not saying AI does nothing. He is saying the payoff is arriving slower and more narrowly than the spending implies. The data from the last year backs him up, and it points to a useful conclusion for anyone actually buying software: the gains are real, but they are concentrated, and they do not show up by default. They show up when you pick the right tool for the right task and measure whether it helped.

The evidence that the hype outran the results

Three findings from 2025 and 2026 tell a consistent story.

Most enterprise AI pilots produce no measurable return. MIT's NANDA initiative published "The GenAI Divide: State of AI in Business 2025" in August 2025. Its headline number: only about 5% of enterprise generative-AI pilots achieve rapid revenue acceleration. The other 95% deliver little to no measurable impact on profit and loss. The study drew on 150 leader interviews, a survey of 350 employees, and analysis of 300 public deployments. The root cause was not weak models. It was a "learning gap": tools that never got integrated into real workflows. Tellingly, vendor-bought tools succeeded around 67% of the time, roughly triple the success rate of internally built systems.

AI can make experienced people slower, even when they feel faster. METR ran a randomized controlled trial from February to June 2025 with 16 experienced open-source developers working on mature repositories they knew well, across 246 real tasks. Allowing AI tools increased completion time by 19%. The developers had predicted AI would make them 24% faster, and even after finishing they estimated they had been 20% faster. Actual result: 19% slower. That is a roughly 39-point gap between belief and reality. METR is careful to scope this to a specific setting (experienced devs, mature codebases, early-2025 tools) and warns against over-generalizing, but the perception gap is the part everyone should remember.

Time saved is not the same as value created. A Boston Consulting Group survey of about 12,000 frontline employees found that 42% saved roughly 8 hours a week from regular AI use. But most got little to no guidance on what to do with the freed time, and half were not reinvesting it in more strategic work. Saving eight hours and spending them on nothing is not a productivity gain. It is a rounding error with a subscription fee.

The market picture matches. From early 2023 to early 2026, the Magnificent Seven's profit margins rose from about 15% to about 25%, while the rest of the S&P 493 stayed near 10%. The gains are real. They are also concentrated in the companies selling the AI, not the broad economy buying it.

Why the gains are real but hidden

None of this means AI is a bust. It means value is conditional. The 5% of pilots that worked, the developers who do get faster, and the teams that turned saved time into output all had something in common: a specific job the tool was genuinely good at, wired into how work already happened, with a way to tell whether it helped.

The failures share a pattern too. A tool bought because a competitor bought one. A model dropped on top of a workflow no one changed. A "productivity" claim no one ever measured. The MIT finding that vendor tools beat internal builds three to one is the same point from another angle: the teams that succeed usually pick a proven tool for a narrow job rather than trying to build a general AI capability from scratch.

What to actually do about it

The reality-check is not "avoid AI." It is "stop buying AI in general and start buying it for something specific."

• Name the task before the tool. The pilots that fail start with "we need an AI strategy." The ones that work start with "drafting first-pass support replies takes three hours a day." Pick a task where competent-but-fast beats slow-and-perfect, and buy for that.
• Prefer proven tools over builds. The three-to-one edge for vendor tools is not an accident. Unless AI is your product, a specialized tool that already solves your job will beat an internal build you have to maintain.
• Measure the before and after. The METR gap exists because people trust the feeling of speed over the fact of it. Time a real task with the tool and without it. If you cannot show the delta, you do not have a productivity gain, you have a vibe.
• Reinvest the time you save. Eight saved hours only count if they go somewhere. Decide in advance what the freed time is for, or it evaporates.

Slok's warning is about markets, but the operational lesson is smaller and more useful. AI pays off when it is aimed, integrated, and measured. That is not a reason to sit out. It is a reason to be specific, which is the whole point of comparing tools on their actual merits before you buy one.