The AI productivity boom nobody has measured yet

What this gives you

Four studies, in the next ten minutes. None of them published by anyone who sells AI. They cover what AI has actually done to productivity and employment inside ordinary firms, as opposed to what executives expect it to do next.

Then the same question asked about the people already on your payroll, with the arithmetic set out plainly enough that you can run it on your own numbers before lunch.

Every figure names its publisher, its title and its release date. Every figure also carries the strongest objection to it, written down. Where an assumption is mine rather than a researcher’s, it says so.

The gist

The mass-layoff forecast rests on a productivity gain that nobody has measured yet.

That is the whole of it. The forecasts are real, they come from serious people, and they are forecasts. The measured effect underneath them, across tens of thousands of firms, currently rounds to nothing.

What firms report AI has actually done

Between November 2025 and January 2026, four central-bank and university survey panels asked the same questions of more than 5,000 CFOs, CEOs and senior executives across the United States, the United Kingdom, Germany and Australia. The panels are run by the Federal Reserve Bank of Atlanta, the Bank of England, the Deutsche Bundesbank and Macquarie University.

About 69% of those firms were already using some AI technology. Adoption is not the issue.

What they reported about results:

• 89% said AI had made no difference to their productivity over the past three years.
• More than 90% said it had made no difference to their employment.
• Averaged across every firm, the measured productivity effect was 0.29% over three years.

What the same executives then forecast for the next three years:

• A productivity gain of about 1.4%, roughly 0.5% a year.
• An employment reduction of about 0.7%, roughly 0.23% a year.

Read those two lists together. That is the layoff story, in full. A measured result of roughly zero, and an expectation.

One more detail worth holding. The researchers also surveyed employees. Employees expect AI to increase employment at their firms by about 0.5%. Executives expect it to cut employment by 0.7%. The people who see the work up close and the people who see the budget are looking at the same technology and reaching opposite conclusions, and neither group has a measured result to point at.

Source: Yotzov, Barrero, Bloom, Bunn, Davis, Foster, Jalca, Meyer, Mizen, Navarrete, Smietanka, Thwaites and Wang, Firms predict an AI productivity boom is coming, CEPR / VoxEU column, 12 March 2026. Underlying paper: Firm Data on AI, NBER Working Paper 34836.

Where the gains that do exist are coming from

A second survey, of 734 senior financial executives at typical US firms, ran through The CFO Survey with panels from FEI, Nasdaq and Duke University. It asked executives to attribute changes in revenue, employment, productivity and capital to AI directly, rather than inferring the effect from outcomes that move for a dozen other reasons.

The implied productivity gain from AI was about 0.6% in 2025, expected to reach about 1.8% in 2026. In manufacturing and construction, executives reported 1.1% for 2025 while the implied figure, derived from their own revenue and employment answers, came in at 0.4%.

Three findings in that paper matter more to an owner than the headline numbers.

The gains run through demand, not through cutting

The researchers found productivity gains most strongly associated with developing or improving products and services, and with reaching and serving customers more effectively. Cost reduction and labour-saving applications happen, mostly at large firms, and they were not the dominant short-run channel in the data. The firms getting anything out of AI are getting it by selling and serving better.

Employment change is small and it is not a cliff

Firm-size and sector-weighted employment is expected to fall by less than 0.4% due to AI in 2026, concentrated in large companies. Smaller firms in the sample anticipate modest AI-driven employment growth.

What firms are buying is a subscription, not a machine

Capital deepening accounts for only about 15% of near-term AI-attributed productivity growth in this data. Most spending is subscriptions, services and training, expensed rather than capitalised. The description the researchers use is renting intangible capital from upstream providers. That is a different economic act from installing a line, and it produces a different kind of return.

Source: Baslandze, Edwards, Graham, McClure, Sparks, Meyer, Waddell and Weitz, AI, productivity, and work: Evidence from US firms, CEPR / VoxEU column, 27 June 2026. Paper forthcoming; see CEPR Discussion Paper 21313.

The Canadian number, and the thing that made it disappear

Statistics Canada linked multiple waves of the Survey of Digital Technology and Internet Use to administrative business microdata, then asked whether Canadian firms that adopted AI were more productive.

The first answer looked strong. AI adopters showed 16.8% higher labour productivity than non-adopters, controlling for industry, size, age, capital intensity, ownership and trade activity.

Then the researchers controlled for how productive those firms already were before they adopted anything. The premium fell to 10.2%. Part of the advantage was simply that productive firms adopt sooner.

Then they added the capabilities the firms already had in place: research and development spending, cloud computing, data analytics, advanced robotics, and information and communications technology training for employees. The premium fell to 5.1% and stopped being statistically significant.

The authors’ own conclusion, in their words as published: AI adoption in isolation is likely insufficient to deliver transformative productivity gains, and the impact depends on a supportive ecosystem of infrastructure, workforce skills and research rather than the technology alone.

Hold that middle item. Workforce skills. A federal statistical agency, with no product to sell, ran the numbers and found that the productivity premium attributed to AI was largely the capability of the company that bought it, including how it had trained its own people.

Source: Li and Liu, Artificial intelligence adoption and productivity in Canadian firms, Statistics Canada, Economic and Social Reports, catalogue 36-28-0001, released 22 April 2026. Underlying study forthcoming in Canadian Public Policy. Note the date on the underlying data: adoption is measured in the 2019 and 2021 survey waves, which predates the generative AI wave. The finding about complementary capabilities is the durable part.

The pilot number

MIT’s Project NANDA found that roughly 95% of enterprise AI pilots produced no measurable impact on profit.

That figure travels further than any of the others, so it needs its limits stated before someone else states them. The report calls itself preliminary findings. The base is 52 structured interviews, 153 survey responses and a review of 300 publicly disclosed initiatives. The authors say plainly that the deployment figures are directionally accurate from interviews rather than from company reporting. It has not been peer reviewed. And it contradicts itself once: the executive summary puts sales and marketing near 70% of AI budget allocation, while two later sections say 50%.

It is the weakest of the four sources cited here, and it is the one you have already seen quoted. That is worth noticing on its own.

Source: Challapally, Pease, Raskar and Chari, The GenAI Divide: State of AI in Business 2025, MIT NANDA, July 2025.

The lever with 736 studies behind it

Now the other side of the ledger.

Gallup’s State of the Global Workplace 2026, released April 2026 on 2025 World Poll data, puts global engagement at 20%. Eight in ten working people are not engaged. North America is the strongest region in the world at 31% engaged, which still leaves 69% not engaged. Manager engagement sits at 22%, down nine points since 2022.

The evidence behind the link between that number and money is not a survey. It is a meta-analysis: 736 studies, 183,806 business units, 347 organisations, 53 industries, 90 countries, 3,354,784 employees.

And one finding that tells you where to aim: managers account for roughly 70% of the variance in team engagement.

Sources: Gallup, State of the Global Workplace 2026, April 2026. Gallup Q12 meta-analysis, 11th edition, May 2024.

The arithmetic, on your numbers

Take a $30M business at a 10% margin. That is $3M of profit.

Run it on your own revenue and your own margin. The point is not my numbers.

The four conditions

Two of the sources above point at the same thing from opposite directions, and the joint between them is worth making explicit, because it is where the money actually sits.

Statistics Canada found the productivity premium credited to AI was largely the capability the buying firm already had, including how it had trained its own people. Capability is not a stock sitting on a shelf. It is built by people doing work under someone who can tell them when they are wrong. Gallup measured that someone: managers account for roughly 70% of the variance in team engagement.

One study says the capability is what pays. The other says who builds it. Neither is a claim about software.

So here are the four conditions, in the order they usually fail.

1. The work is written down. Somebody outside the two or three people who do it can follow it. Not a policy binder. The actual sequence, the actual decision points, the actual exceptions.
2. The numbers are trusted. The data the tool reads is the data the business runs on. If your estimator keeps a private spreadsheet because the system is wrong, the system is wrong, and anything built on top of it inherits that.
3. The managers can tell good work from work that looks good. This is the condition the technology makes harder rather than easier. A polished output now arrives faster than the judgment required to check it.
4. The people are willing to change how they work. Adoption is a behaviour, not a licence count. At 31% engaged in North America, buying the licence is the easy part.

An assumption of mine, not a researcher’s, and it is the one I would most like to be wrong about. The work being automated first is often the work through which people learned the job. Building the estimate from scratch. Chasing a discrepancy to the bottom. Sitting with a complaint until the pattern shows itself. Remove that and the third condition degrades quietly, because the person who should be checking the output is the person who never had to build one.

How long this takes

Nobody publishes the number you actually want, which is how fast engagement moves when a company decides to move it. Gallup does not publish it. Nobody else cited here publishes it either. Any figure I gave you would be invented, so I am not going to give you one.

Here is what the sources do fix in time.

The 0.29% is a measured result across three years. The 1.4% is a forecast for year three. Neither is a next-quarter number, in either direction.

And Gallup’s manager engagement figure sits at 22%, down nine points since 2022. That is measured movement of roughly two points a year, from a named publisher, across a population of millions. It is the only rate of change in this piece, and it is worth reading twice for two reasons. It proves the number is not fixed. And it fell, which is the direction things move when nobody is holding them.

My assumption, flagged as mine: falling is easier than rising, so treat two points a year as a ceiling on speed rather than as an expectation.

What moves faster than the engagement number is the four conditions themselves. In my experience the operating conditions move on a quarter. Writing down how a quote actually gets built is weeks of work, not years. Whether the engagement number follows, and how far, is decided by the managers in condition three, and that is not something I will promise on a page.

Where this argument is weakest

Four places. Each of them is the first thing a sceptic will reach for, so here they are first.

The two sets of numbers are not the same kind of number

Gallup’s 23% is a spread between the top and bottom quarter of business units. The AI figures are averages across all firms in a sample. Comparing them directly, as though 23% and 1.4% were rival estimates of the same thing, would be dishonest. What they share is that each one sizes a lever available to an owner. That is the only comparison being made here.

Gallup sells engagement products

They benefit commercially if you believe the finding. The study is correlational: engaged units are more profitable, which is not the same as engagement causing the profit. A fair sceptic will point out that profitable units can afford better managers. The strongest available answer is that the comparison runs between business units inside the same companies, same industry, same customers, same balance sheet, which narrows the alternatives without eliminating them.

The forecast may be right and simply early

This is the honest case for the other side, and it is a good one. Brynjolfsson, Rock and Syverson’s work on the productivity J-curve, cited in the US study above, holds that general-purpose technologies deliver their gains only after learning, organisational change and complementary investment. The Solow paradox is the same observation from the 1980s. The gain may be sitting in front of us, waiting on the organisational work.

Notice what that concession actually says. The complementary investment is the process, the documentation and the training. Statistics Canada found the same thing from the other direction. If the J-curve explanation is correct, the sequence is not in dispute: the operating work comes first, and the software pays only after it.

There is a version of that argument which is worse than early, and it is mine rather than a researcher’s. If the complementary investment is training and judgment, and the work being automated is the work through which training and judgment happened, then the curve does not lift late. It flattens. That is not a finding and I would not present it as one. It is the reason the four conditions above are worth checking before the next licence renewal rather than after it.

The composition of work is genuinely changing

The US study expects the share of employees doing routine clerical work to fall by more than two percentage points by 2028, mostly at large firms, with partially offsetting increases in skilled technical roles, mostly at smaller ones. Office and administrative support roles carry the most negative exposure. That is real, and an owner should plan for it. It is also a different claim from mass unemployment, and it is the claim the data supports.

The gist, again

Every number that exists says the machine has not yet moved the needle.

Every number that exists also says the other lever has been measured for twenty years, across three million employees, and most buildings are not pulling it.

One of those is a forecast about technology. The other is a description of your Monday.

What you now have

Four non-vendor findings you can cite in your own boardroom, with their release dates and their weak points. A number for what AI has actually returned to firms like yours. A separate number for what the people already on your payroll are worth if the conditions around them change. Four conditions to check the building against. And an arithmetic you can run without asking anyone’s permission.

When your margin last moved, was it something you bought, or something your people started doing differently?