The leaders who call AI a support tool are reading the ceiling right

The verdict is usually baked into the slide before you get to speak: you’re behind. You planned for AI as an assistant, something that drafts and suggests while people stay in charge, and that now reads as the timid bet. The future everyone points at is agentic AI, software that takes on a whole job and runs it start to finish with no person in the loop. Calibrate for the assistant and you’re building your org around a world that’s already being superseded.

That framing has a real source. This summer the consulting firm McKinsey published its State of Organizations 2026, a survey of more than 10,000 executives across dozens of countries and industries. One finding traveled fast: 53% of leaders expect AI to work mainly as a support tool, and only 25% expect agentic AI to take on autonomous roles. McKinsey’s read on that gap was blunt. Teams are designing their governance and operating models for “a paradigm already being superseded,” the mismatch is structural, and it amounts to a planning calibration failure.

Read the verdict slowly and something in it doesn’t hold. The survey measured what leaders expect, then graded those expectations against a forecast of where AI is going, when the thing that decides who’s calibrated is what AI can actually do today. Put those same numbers next to a reliability benchmark instead of a forecast, and the ranking reverses. The leaders planning for a support tool turn out to be the ones reading the capability data right.

Here’s the data the verdict skips. METR, an independent lab that stress-tests frontier AI models, measured how long an agent can work before its success rate falls to a coin flip. The answer, as of this spring, is two to four hours, and only on tasks that are well-specified and low on context, the kind you could hand a competent freelancer on their first day. Push past that window, or into work that needs organizational history and cross-person judgment, and reliability drops off a cliff. METR has a name for the weak spot, context-fragility, and it reads the pattern as durable rather than a bug about to be patched.

Two production benchmarks say the same thing from inside real codebases. The software analytics company PostHog ran agents across its 1.6-million-line codebase and had to build more than a dozen guardrails to keep them safe, because the agents kept breaking things downstream they couldn’t see. The payments company Stripe built eleven realistic test environments and found a clean split: agents can build a working integration, but they struggle to validate it. The failures are the unglamorous, load-bearing ones. Making sure a payment doesn’t get charged twice. Retrying safely after an error. Staying inside the permissions they were granted. Different companies, different domains, and the agents hit the same wall. They produce fluently; they can’t yet be trusted to check their own work or own a process end to end.

Now translate. “AI as a support tool” is a plain-English name for exactly that deployment shape: agents doing bounded pieces of work inside a process a human still runs. “Agentic AI in an autonomous role” is the shape the benchmarks say no current agent can hold at reliability. So the 53% expecting a support tool are describing what the technology actually delivers, and the 25% expecting autonomous roles are describing what it demonstrably can’t. The survey called the first group behind. The reliability data calls them correct.

The strongest objection deserves stating in full, because it nearly rescues the verdict. Expectations aren’t about today, they’re about a planning horizon. And the ceiling is rising fast: METR’s own history shows the reliable-work window roughly doubling every seven months. Plan for the assistant you have now and you’ll be caught flat when the autonomous version lands in eighteen months. On that reading, planning for agentic AI is prudent forward-calibration, and McKinsey is right after all.

That objection is a bet, and it pays to see the bet clearly. It wagers that the doubling curve will carry agents straight through context-fragility inside the planning window. But the durable part of the wall is exactly the judgment layer: validating your own output, recovering from a broken state, staying inside your authority across a long task. That layer is the part the benchmarks say isn’t closing on the same curve. The window where an agent can grind gets longer; the ability to own a role hasn’t moved with it. You can believe the wall dissolves on schedule. You can’t call that belief measured, because nobody measured it. The survey measured expectations and compared them to a forecast. There was never a capability number in the room.

And notice who benefits from the forecast. A firm whose business is selling transformation has a standing reason to frame the cautious expectation as a lag, because “you’re behind the wave” is the finding that books the engagement. That doesn’t make the forecast wrong. It does mean the confident verdict is carrying more certainty than its evidence. The number underneath, 53% and 25%, is solid. The judgment wrapped around it is a house opinion.

So the next time a deck tells you you’re behind for treating AI as a support tool, run one check before you re-plan the roadmap around it. Ask what the person calling you miscalibrated actually measured. If they measured your expectations and scored them against a picture of where they’re sure things are heading, that’s a forecast dressed as a measurement. If they measured what the agents reliably do, listen closely, because that’s the number that settles who’s calibrated. Most of the time it’s the first one. And when it is, the calibration question doesn’t point at you. It points back at the slide.