AI productivity paradox: everyone is faster, nothing moved

Start here

Three questions about your own company.

Answer these before you read the research, because the research will make all three feel obvious afterwards. There are three places where the value leaks and each one has a different fix. These questions tell you which one you have.

Is individual output up over the past year? People finishing things that used to take longer.
Has anything in your top-line numbers moved because of it? Revenue, throughput, delivery time, margin. Something you would show a board.
Can you say where the time saved went? One sentence, per team, that somebody could check.

1

The time was never pointed anywhere

Output is up, nothing has moved, and nobody can say where the hours went. Two thirds of people get no guidance on what to do with the time AI frees up.4 This is the most common one and the cheapest to fix.

2

The gain is real and the queue is somewhere else

People are faster at their own step and the work still waits, because the constraint sits between teams rather than inside one. Speeding up a step that was never the bottleneck changes nothing you can measure.

3

You are not measuring in a place where it could show

Adoption is being counted, licences are being counted, and the thing the business actually runs on is not. Then the honest answer to the board is that you do not know, which sounds the same as nothing happened.

Out in the wild

Which of these was your last leadership meeting?

Four moments. If you recognise two, this page is about your company.

Somebody demonstrated how much faster a task goes now, everybody nodded, and nobody asked what happened to the time that was freed up.
The AI spend has a line in the budget. The return has no line anywhere, and the last person who asked got an example rather than a number.
The answer to “what has it delivered” took four minutes and contained no figure that could be checked.
You decided to invest more next year, on the same reasoning you used last year, because stopping felt riskier than continuing.

That last one is the one I would watch. Deciding to continue is fine. Deciding to continue without being able to say what the previous round bought you means the next round will end the same way, and by then there is a lot more of it to explain.

The measured part

Three studies, three methods, one answer.

This is not something happening to you in particular. Between late 2025 and the middle of 2026, three separate research efforts went looking for the impact and came back with roughly the same finding.

Different samples, different questions, same shape. The 6% has not moved in a year, while investment and conviction kept climbing.1 2 3

The NBER work is the one I would read if you only read one. Research teams at the Federal Reserve Bank of Atlanta, the Bank of England, the Deutsche Bundesbank and Macquarie University put identical questions to nearly six thousand CEOs, CFOs and senior finance people across four countries between November and January.1 Sixty-nine per cent of those firms actively use AI. Over ninety per cent report no effect on employment in three years, and eighty-nine per cent report none on labour productivity.

Three years of no measurable effect, and the same people expect a gain starting next year. That is hope with a budget attached. Paul Musters

Because that is the part of the NBER data almost nobody quotes. The same executives who saw nothing over three years expect labour productivity up 1.4% and employment down 0.7% over the next three.1 Something has to change for that to happen, and in most companies nothing is scheduled to.

Which level is the leak on?

Six questions, three minutes, no name attached. Your level, what it is costing you, and what one step up changes.

Do the Culture Level scan

Why it happens

The gain is individual. The result has to be organisational.

Somebody saves four hours on a Tuesday. For that to reach a company number, three things have to happen after that Tuesday, and in most companies none of them do.

Time saved Real, and already yours.

Given a direction Somebody says what the hour is for, or it fills with more of the same work. 66% get nothing here4

Aimed at the queue It lands on the step that actually holds delivery, rather than one that was already fast. almost nobody checks this

Your numbers The quarter where it finally shows up.

Where the value leaks between one person's Tuesday and the company's quarter. The first gap is documented.4 The second one I run into in almost every company I walk into.

Nobody said what the time was for

BCG found that 42% of regular users save at least a full working day a week, while 66% get limited or no guidance on what to do with it.4 An hour with no destination goes into more of the same work. It feels productive on the day and it is invisible by the quarter. This is the leak I would fix first, because saying what the time is for costs one meeting.

The bottleneck was never the typing

Most AI gains land on the step where somebody produces something. The thing that actually holds your delivery is usually a handover, an approval, or one person everything routes through. Make the drafting twice as fast and the queue in front of the approver just gets longer. You have bought speed at a place that was never the constraint, which is why the customer notices nothing.

You are counting the wrong thing

Licences, logins, weekly active users. All easy to count and none of them tell you anything about the business. MIT's read on this is that the barrier is organisational rather than technical, and they call it a learning gap.2 That matches what I see: the companies with results are not running better models, they changed how the work moves.

A leadership team looking at their own numbers on screen

How to see it

Two columns, twelve months apart.

No project, no consultant, one page. Put your top-line business metrics from twelve months ago next to today, and put your AI adoption data next to both.

What to put on the page.

The three or four numbers your business actually runs on. Revenue per head, delivery time, throughput, margin. The ones you would defend in a board meeting.
The same numbers from twelve months ago.
Your adoption figures over the same period: who uses what, how often.
One line per team on where the time saved has gone. If you cannot write that line, that is the finding.

If individual productivity is up and the organisational numbers are flat, the constraint is structural. Speed is not what is missing, so buying more speed will not close it.

One warning about this exercise. It is tempting to fill the fourth line with something plausible. Resist that. An honest blank there is worth more than a confident sentence nobody could check, and the blank is what makes the next conversation possible.

Where it sits

Individual gains become company results at Level 3.

01Campfire60%

02Wild West25%

03Blueprint10%Gains become results here

04Engine4%

05Ecosystem1%

Share of companies per level. The five AI Culture Levels we use with clients, from everyone doing their own thing to AI being part of how the company runs.5

In the five AI Culture Levels we use with clients, Level 2 is daily individual use with no shared standard. Everybody found something that works and nobody works the same way. Everything on this page is what Level 2 looks like in a P&L, which is to say it looks like nothing at all.

Level 3 is where the way of working gets written down and the good method becomes the normal method. That is the first level where one person's Tuesday can reach a company number, because the gain stops being personal. About one in ten organisations is there, which lines up uncomfortably well with the 6% that McKinsey counts as high performers.3

What a measurement adds to a spreadsheet

An Operating Profile in use. Personality type and AI level in one profile, with the agents that fit it.

The two-column exercise tells you that the gain is not arriving. It does not tell you where in the company it stops, and that is the part you need before you can spend another euro sensibly. A total is useless here. What you want is the distribution: which team is three levels ahead, which one has not started, and where the handover between them sits.

We measure per person: how somebody thinks and works, and how far along they are with AI. Then the leak has an address, and usually it turns out to be one handover between two teams rather than a company-wide problem. That is a much cheaper thing to fix than the transformation programme somebody is about to propose.

About emaho

emaho measures one Operating Profile per person: personality type and AI level in a single profile. On that we build a personal set of AI agents that fit how that person works, inside the tools they already use. Fifteen minutes to complete, first profile free, built for companies between 20 and 500 people.

See your own profile, free Find your AI Culture Level

Fifteen minutes per person. No credit card, no strings.

A team working out where their work actually gets held up

What works

Four steps, and none of them is a bigger licence.

Say what the time is for. Out loud, per team, this quarter. “The extra day goes into client conversations.” “The extra day goes into cutting our delivery time.” Any clear answer beats the current situation, because right now two thirds of your people have been given nothing and are filling the space with more of the same work.4

Then find the queue. Walk one piece of work from request to delivered and write down where it waits. In every company I have done this with, the waiting is longer than the working, and it waits in a handover rather than inside a task. That is where speed would actually show up, and it is almost never where the AI budget went.

Pick one business number and hold it still for two quarters. Delivery time, revenue per head, whatever your business runs on. One number, two quarters, no changing the definition halfway. Most AI reporting fails because the metric moved more often than the result did.

And measure the distribution, per person and per team, so the next investment goes where the gap is. That is the difference between a plan and a purchase.

Done the two columns? Send me both

The business numbers and the adoption numbers, twelve months apart. I will tell you which of the three leaks it points to and what I would do first. You get a straight answer and nothing after it.

Message me on WhatsApp

If you do nothing

Three years of spend, and one uncomfortable question.

The money is the obvious part and the smallest. Another year of licences, another pilot, another round approved on the same reasoning as the last one. Painful in the budget, survivable.

What costs more is the credibility. The second time a leadership team promises that this year AI will show up in the numbers, the room stops listening, and the people who were genuinely making things better get grouped in with the hype. After that it takes years to propose anything with the word AI in it and be taken seriously.

And there is the competitive one, which arrives slowly and then all at once. Six per cent of companies are getting real earnings impact and that number has not moved in a year.3 They are not running better models than you. They rebuilt how work moves through the company, which takes about eighteen months, so the gap you are looking at today was decided a year and a half ago.

Asked and answered

What people ask when the numbers refuse to move

Why does AI adoption not show up in company results?

Because the gain is individual and the result has to be organisational. One person saving four hours only reaches a company number if that time is pointed at something, if it lands on the step that actually holds delivery, and if somebody is measuring in a place where it could show. BCG found 66% of people get limited or no guidance on what to do with the time AI frees up.

How many companies actually see a productivity gain from AI?

Very few so far. An NBER study of nearly 6,000 executives across the United States, United Kingdom, Germany and Australia, fielded between November 2025 and January 2026, found 89% reporting no impact on labour productivity over the past three years and more than 90% reporting no effect on employment. This is despite 69% of those firms actively using AI.

What is the AI productivity paradox?

It is the gap between visible individual speed and invisible organisational output. Everyone can point to a task that got faster, and nothing in revenue, delivery time or margin has moved. The paradox resolves once you look at where work waits: the gains land on tasks, while the delays live in handovers between teams.

How do you find out where your company loses the AI gain?

Look at the level the organisation works on rather than at the tools. emaho reads organisations on five AI Culture Levels, and the jump from individual speed to company results happens at Level 4, where about 4% sit. The scan is six questions, three minutes and anonymous, and it names what it is costing you at your level.

We invested in AI and got nothing back. What went wrong?

Usually one of three things, and it is worth knowing which. Either the time saved was never given a direction, or the speed landed on a step that was not the bottleneck, or nothing is being measured in a place where the result could appear. MIT described the barrier as organisational rather than technical, calling it a learning gap.

How do I measure AI ROI properly?

Pick one business number your company already runs on, hold its definition still for two quarters, and put it next to your adoption data over the same period. Licences, logins and weekly active users are easy to count and tell you nothing about the business. If individual productivity is up and the business number is flat, the constraint is structural.

Should we invest more in AI if we have seen no results yet?

Investing more is defensible. Investing more on the same reasoning as last year is how a second year ends the way the first one did. Before the next round, be able to say what the previous round bought, where the time saved went, and which step in your delivery it was aimed at. If those three answers are blank, more licences will not fill them.

Where does the time saved by AI actually go?

Into more of the same work, in most companies. BCG found 42% of regular frontline users save at least a full working day per week while two thirds get no guidance on what to do with it, and more than half do not redirect it into strategic work. Time with no destination gets absorbed by the day and shows up nowhere by the quarter.

Where can I read about the other AI challenges companies run into?

This is one of 25 AI challenges emaho documents, each with the research behind it and a test you can run this week. This one travels with the coordination tax and the measurement blindspot.

What separates the companies that do get results from AI?

They changed how work moves rather than how fast people type. McKinsey counted 6% of nearly 1,700 organisations as high performers with real earnings impact, a share that has not moved in a year while investment kept climbing. BCG measured the same thing from the other side: a clear strategy lifts measurable impact by around 25 percentage points, better tools by about five.

Is it a technology problem or an organisation problem?

Organisation, on the current evidence. MIT concluded the barrier is a learning gap rather than model capability, and the NBER data shows firms with high adoption and no measurable gain sitting side by side with firms that have both. The models are the same in both companies.

How does this relate to AI maturity levels?

Everything on this page is what Level 2 of the five emaho AI Culture Levels looks like in a P&L: daily individual use, no shared standard, nothing that adds up. Level 3, where the way of working is written down and the good method becomes the normal method, is the first level at which one person's saved hour can reach a company number. About one in ten organisations is there.

Nearby

Three places the gain gets stuck.

Each of these is one of the leaks in the diagram above, worked out on its own page.

06

The usage gap

Three quarters of your people already use AI weekly and two thirds were told nothing about what to do with the time. The first leak, with the numbers behind it.

Read this one In the full set 11

Coordination cost

Everyone ships more and the work moves through the system slower than before. The second leak, and the one that eats the gain the fastest.

Read this one In the full set 21

The measurement blind spot

You count licences and logins because they are countable. The third leak: proving what AI delivered when nobody agreed what would count as delivered.

Read this one

All 25 AI challenges, and what to do about each

Twenty-five things that break inside a company once people start using AI, each with the research behind it and the level where it starts to bite. This one starts to bite at Blueprint, and 7 of the others start there too.

Open the full set

Getting going

Trace the leak before you fund another year of it.

The two columns tell you the gain is not arriving. A profile per person tells you where in the company it stops, which is the thing you need before the next budget round.

Take one workflow and follow the gainHours saved at the desk are easy to find. Where they stop is the number you actually need.
Locate the floor it stops onOnce everybody has a profile you can see where the company absorbs the gain instead of passing it on.
Decide what the saved hour is forIf nobody has said, it quietly becomes more of the same work and the gain never shows up anywhere you can see it.

Start with your own profile, free Ask me anything first Plan a 30-minute call

First profile free · no credit card · built for companies of 20 to 500 · you decide what your team gets to see

Rather start at company level? The Culture Level scan is six questions, three minutes, and asks nothing of you.

Paul MustersFounder of emaho

Paul Musters

Fifteen years of leadership and team development in Dutch scale-ups. That practice now sits in software: one Operating Profile per person, with agents that actually fit. He writes these pages from what he runs into with clients, not from a research summary.

LinkedIn · paul@emaho.world · WhatsApp

Sources and numbers used on this page

National Bureau of Economic Research, working paper Firm Data on AI, February 2026. Identical survey questions fielded by research teams at the Federal Reserve Bank of Atlanta, the Bank of England, the Deutsche Bundesbank and Macquarie University between November 2025 and January 2026, answered by nearly 6,000 CEOs, CFOs and senior finance managers in the United States, United Kingdom, Germany and Australia. Source of: 69% of firms actively using AI, more than 90% reporting no effect on employment over three years, 89% reporting no impact on labour productivity, and the forward expectation of 1.4% higher labour productivity and 0.7% lower employment over the next three years.
MIT, 2025. Despite 30 to 40 billion dollars of enterprise AI investment, 95% of organisations report no return, with the barrier described as organisational rather than technical and named a learning gap.
McKinsey, August 2026. Across nearly 1,700 executives worldwide, 6% qualify as AI high performers with real earnings impact, unchanged from a year earlier, while investment and conviction continue to climb.
Boston Consulting Group, AI at Work, fourth annual edition, 3 June 2026. Global survey of 11,749 workers across 14 markets. Source of: 42% of regular frontline users saving at least a full working day per week, and 66% receiving limited or no guidance on what to do with that time.
emaho AI Culture Levels. Level 2 is daily individual use with no shared standard and holds about a quarter of organisations. Level 3, where the way of working is documented, holds about one in ten. Calibrated against BCG 2025 and McKinsey 2025.

Numbers are quoted as published. The NBER figures were checked at the source. The MIT and McKinsey figures come from emaho's own research file on the 23 challenges.