The AI productivity gap: why 80% feel it and only 37% can prove it
McKinsey's 2026 State of AI survey, 1,719 respondents, found 80% of individual users report feeling more productive with AI. Only 37% of organizations report any EBIT, earnings before interest and taxes, impact from it, and just 6% qualify as "high performers," attributing 5% or more of EBIT to AI with a significant effect. Both numbers are flat against last year despite continued spending. The gap is a sign that most companies are measuring the wrong layer, not that AI doesn't work: individual feeling is a fact about a person's chat window, EBIT impact is a fact about whether the output of that chat window ever reached a system the business runs on. This post is a way to tell which side of the gap your own company is on, and what closes it.
What is the 80/37 gap, exactly?
McKinsey's 2026 State of AI survey asked two different questions and got two different answers. Asked whether AI made them personally more productive, 80% of respondents said yes. Asked whether their organization could point to any EBIT impact from AI, only 37% said yes, and both figures are essentially unchanged from the year before despite companies spending more on AI in the meantime. Six percent of respondents go further and qualify as "high performers," meaning they attribute 5% or more of their organization's EBIT to AI with a real, stated effect. The shape of the gap: a large, honestly reported personal experience, and a much smaller organizational one, holding steady while the money going in keeps rising.
Why doesn't "everyone feels more productive" show up as EBIT impact?
Because the two questions measure different things. "I feel more productive" is a claim about a person's own hour: a draft came faster, a question got answered without a search. "The organization saw an EBIT impact" is a claim about the business's numbers, which only move when something changes in a system the business actually runs on: a ticket closed without a human step, for instance. A chat window can produce the first kind of change constantly and the second kind never, because a chat window doesn't write anything back anywhere. It answers. It doesn't finish.
Is this just a measurement problem, or a real gap in what AI is doing?
Both, and they're related. Part of the gap is a measurement lag: a faster draft is real value even if no dashboard captures it cleanly. Part of it is not a measurement problem at all. If the record update itself never happens, there's nothing for a dashboard to eventually catch up to. The distinction matters because the fix is different in each case. A measurement lag gets solved by better attribution. A missing mechanism gets solved by giving the AI a way to act on the systems the work lives in, which no amount of better measurement will substitute for.
What's the tell that a function is talking about AI rather than completing work with it?
Ask what changed in a system of record, not what changed in a chat log. A support team that says "we use AI to draft replies" is describing a chat window. A support team that says "our AI closes a ticket end to end for the categories it's cleared to handle, and a person only reviews the exceptions" is describing something that touches EBIT, because tickets closed without a person is a number that shows up in cost per ticket, not just in how the team feels about their week. The same test applies in ops, in sales, in finance: did the AI's output become the new state of a system, or did it become a paragraph someone still had to act on manually?
How does a chief information officer or head of AI diagnose which side of the gap their own company is on?
Walk the functions one at a time and ask two questions of each: where is AI currently used, and what changed in a system of record as a direct result, without a person carrying the output there by hand. A function that can only answer the first question is contributing to the 80%. A function that can answer both is contributing to the 37%, and probably knows roughly how. This is a more useful audit than counting seats or licenses, because seat count measures adoption of the chat window, not whether anything downstream of it changed.
Does spending more on AI close the gap by itself?
McKinsey's own numbers say no: both the 80% and the 37% held flat year over year while spending kept climbing. More licenses produce more people chatting, which is exactly the layer that was already large. Spending closes the gap only when it goes toward the part that's missing: connecting the AI already in use to the systems it needs to write to, and toward the harder, less glamorous work of redesigning the workflow around that connection, not just toward more seats of the same chat interface.
What does the research say closes the gap, once spending doesn't?
The Stanford Digital Economy Lab's own study of 51 successful AI deployments found that the technology itself was consistently the easiest part of getting a deployment to move the numbers. The hard part, cited in 77% of the toughest challenges, was intangible: change management, data quality, and redesigning the workflow itself around what the AI could now do. Different survey, same mechanism: the gap doesn't close because a better model shipped. It closes because someone did the unglamorous work of deciding which steps of a workflow the AI should now own outright, and rebuilt the process around that decision.
Isn't this just an argument for hiring more engineers to wire up integrations?
For the parts of a company with engineers already building automations, that path already exists and often works. The gap this post is about lives with the much larger group of functions, support, sales, ops, finance, HR, that don't have an engineer on standby to build a custom integration every time a workflow needs one. For that group, the fix is giving the AI client people already use a governed connection to the company's actual systems, provisioned by role rather than requested one integration at a time, and a shared way to hand it a specific workflow someone already worked out how to do correctly, not a bigger engineering backlog.
What does this look like as a decision, not a tooling purchase?
It looks like picking one workflow per function, the one everyone already agrees is repetitive and low-judgment, and deciding explicitly that the AI now owns the end-to-end version of it: the record update, not just the draft. The rollout decision belongs to whoever owns the function, not whoever owns the AI budget. The technology to do this, one connection between an AI client and the systems a company runs on, provisioned by role through the identity system already in place, with the workflow itself captured as a shared, versioned skill rather than reinvented per person, already exists. What's usually missing is someone deciding which workflow goes first.
What should a CIO or head of AI ask in the next AI review meeting?
Not "how many people are using AI," which is already answered and already high. Ask which functions can point to a system of record that changed because of it, and for the functions that can't yet, ask what's stopping the AI already in use from reaching the systems that function runs on. This is the question that sorts a company into the 37% instead of the 80%, and it's a question about access and workflow ownership, not about which model anyone is running.
FAQ
Is the 80/37 split from McKinsey's own survey, or an estimate?
It's McKinsey's own reported figures from its 2026 State of AI survey: 80% of 1,719 respondents report personal productivity gains, 37% of organizations report EBIT impact, and 6% qualify as high performers attributing 5% or more of EBIT to AI.
Does McKinsey explain what the highest-performing 6% do differently?
Not in the breakdown available from this survey. The closest adjacent evidence comes from a separate study, the Stanford Digital Economy Lab's deployment research, which found that successful deployments were distinguished by workflow redesign and change management, not by the underlying technology.
Is this gap a sign that AI adoption has stalled?
No, adoption is high and rising, which is exactly why the 80% figure is large. The gap is about whether that adoption reaches systems of record, not about whether people are using AI at all.
Does this mean chat-based AI tools have no value?
No. They produce real, honestly reported personal productivity gains. The gap is specifically about EBIT impact, which requires the AI's output to change a system of record, something a chat window alone doesn't do regardless of how useful the drafting is.
Is the fix just to buy more AI licenses?
McKinsey's own data argues against this: spending increased while both the 80% and the 37% held flat. The fix is connecting the AI already in use to the systems it needs to act on, and redesigning the workflow around that connection.