The AI productivity paradox: what to tell the board when AI spend is up and the numbers aren't

I read a piece from The Economy on July 1 that put a figure on something a lot of executives are feeling and not saying out loud. Across the US companies it surveyed, AI added roughly 0.6% to labor-productivity growth in 2025. Over the same stretch the headline productivity rate went the other way, from 2.1% for the full year to a 0.3% annualized rate in the first quarter of 2026. Adoption climbed the whole time.

So here is the belief worth retiring: that rising AI spend is proof of rising productivity. It feels true. The invoices are real, the demos land, the dashboards light up. The output line just has not agreed yet.

The task-level wins that make the spend feel earned

Give the myth its due, because it is built on something real. At the level of a single task, AI works. The Economy’s own numbers show professional writers finishing about 40% faster with a chatbot, and customer-support agents completing roughly 15% more per hour. Goldman Sachs found a clean 30% productivity boost in two narrow areas, coding and customer service, when it dug into the data earlier this year.

Every leader has watched a demo that genuinely worked. That memory is persuasive. When the license renewal shows up, it lands next to a real experience of a real person getting real work done quicker.

The problem is what happens to that saved time. A marketer who drafts 40% faster rarely hands back the afternoon. They take on more drafts. The efficiency gets absorbed into higher output expectations, and it quietly evaporates before it ever reaches a margin. Individual speed is not the same thing as organizational productivity, and the gap between the two is where the paradox lives.

The productivity line that slipped from 2.1% to 0.3%

Zoom out from the task to the enterprise and the picture changes hard. A widely cited NBER survey of roughly 6,000 executives found that more than 80% of firms reported no measurable productivity gain from AI over the past three years. Not a small gain. None they could measure.

Deloitte’s 2026 State of AI in the Enterprise report tells the same story from a different angle. Most companies feel faster; far fewer can point to money.

The counterintuitive detail is my favorite one, because it stops the “just adopt harder” reflex cold. The same July data noted that open-source developers, the group everyone assumes AI helps most, actually ran about 19% slower on certain tasks. Faster is not automatic. It depends on the work, the context, and whether the tool fits the workflow or just sits next to it.

And the effect is real enough to show up somewhere, just not in the productivity print. It shows up in who gets hired.

Where the AI money actually went

Here is the line from that July piece that should reframe the whole conversation. Over half of the 2026 respondents said their firms had already invested in AI, but, as The Economy put it, “the bulk of the expenditure was directed towards subscriptions, services and learning and development.”

Sit with that. A large share of what gets reported as “AI investment” is commitment, not transformation. It is seats, contracts, and training days. Those are reasonable things to buy. They are also the easiest things in the world to mistake for progress, because they generate invoices, adoption charts, and a comfortable feeling that the company is moving.

Spend books as progress the moment the purchase order clears. Value only books when a number the business already cared about moves. Those two events can be a year or more apart, and the gap between them is exactly what your board is starting to ask about.

Why this is a measurement gap, not a capability gap

Economists have a name for the AI productivity paradox, and it is worth using at the board table because it signals you have context, not panic. The older version is the Solow paradox, coined when computers were everywhere except in the productivity statistics. Fortune resurfaced it in May, noting that US labor productivity grew about 2.7% in 2025, near double the prior decade’s average, yet the gains stubbornly refuse to show up economy-wide.

There is a genuine, honest disagreement here, and I would not paper over it. One camp, led by Stanford’s Erik Brynjolfsson, argues we are early on a J-curve: heavy investment and reorganization come first, the harvest comes later, and the productivity liftoff has already begun. The other camp, holding the NBER survey data, says most firms simply are not seeing it. Both are looking at real numbers. The World Economic Forum’s own chief economists split the difference in May by pushing their expected timelines for meaningful gains further out across almost every industry.

What both camps agree on is the useful part: the bottleneck is not model capability. The models are extraordinary. The gap is measurement and organizational design. It is whether a company redrew a workflow, retired the old way of doing the task, and instrumented the before-and-after, or whether it bought a powerful tool and bolted it onto a process nobody changed.

The models are not the constraint. The constraint is whether you changed a workflow and measured it, or bought a tool and hoped.

Three moves before your next board update

If the spend line and the productivity line disagree, the worst answer is a bigger spend number. Here is the calmer, more credible path.

Bring one before-and-after number, not a budget total. Pick a single workflow, show what it cost in time or money before AI and after, and own the honest figure even if it is modest. One real moved number outranks a slide full of adoption rates.

Say where the freed time went. If AI saved your support team hours, name what those hours became: more tickets closed, faster resolution, a role redeployed. Time saved that cannot be traced to an outcome is time that quietly refilled with more of the same work.

Name the timeline honestly. If you believe in the J-curve, say so, and say the harvest takes redesign and a few quarters, not another license. Boards forgive a patient, measured plan. They lose confidence in a spend chart pretending to be a results chart.

The reassuring part of this whole paradox is that the 5% of companies pulling real value are not doing anything exotic. They picked fewer things, changed the actual work around them, and measured what happened. That is not a technology secret. It is just good management, holding a genuinely powerful tool, refusing to confuse the receipt for the result.