Your teams’ AI productivity gains are probably real, and probably not reaching your P&L. The best evidence so far says they are stuck at the step after the one AI sped up, which almost nobody measures.
That is a claim worth arguing with, so here is the working.
The gap shows up everywhere
Start with the boardroom view. McKinsey’s 2025 State of AI survey found that only 39% of organisations could attribute any EBIT impact to AI at all, and most of those put it at under 5% of EBIT. The roughly 6% they class as high performers were nearly three times as likely as everyone else to have fundamentally redesigned their workflows.
Then the worker’s view. Economists Anders Humlum and Emilie Vestergaard linked AI chatbot adoption surveys covering 25,000 Danish workers in 11 occupations to official employer records. Average time saved came out at around 3%, and they found no measurable effect on earnings or hours worked. One detail stood out to me: 8.4% of workers reported the chatbots had created new tasks for them, including some who did not use the tools themselves.
So people feel faster, the time saving is real but small, and some of it reappears as new work. What neither study can show you is where the rest of it went. For that you need a field where the work leaves a trail.
Software is where you can watch the work move
Software engineering records almost everything. Code is written in one tool, reviewed in another, and the finished piece of work is closed off in a third. That makes it the one place you can follow a productivity gain from the keyboard to the customer.
Fiona Chen and James Stratton at Harvard did exactly that. Their working paper, updated on 14 November 2026, uses data from an engineering analytics platform covering 445 million work events from 720,542 engineers at 655 firms, from January 2021 to August 2026. They compared firms before and after they adopted AI coding tools, using differences in when each firm adopted.
Once firms adopted AI coding agents, pull requests (a chunk of new code submitted for someone to check) went up by 35.7%. Jira issues resolved (a unit of work actually finished and delivered) went up by 13.7%, and the authors can rule out any rise above 30.7%. Lines of code went up 44%.
Their model puts the underlying gain in coding productivity at around 74%. Finished work rose by less than 14%.
Where it went
It went into the review queue. After agents arrived, the time from a pull request being submitted to being approved rose by 56.1%. Code waited longer before anyone looked at it, spent longer in review once they did, and drew more comments.
The authors explain the waiting with Little’s Law, a piece of queueing maths from 1961: if work arrives faster than it can be cleared, the backlog grows, and the wait is simply the size of the backlog divided by how fast you clear it. Speed up the people writing the code without changing the people checking it, and you have built a queue. Nobody decided to build it. It is just what happens.
Their conclusion, in their words: “As one stage of production speeds up, the bottleneck shifts to another, and labor must be reallocated towards it.” And the line I would pin to the wall of anyone running an AI programme: “intermediate metrics, such as lines of code or commits, overstate the gains from AI adoption.”
A separate NBER study by Demirer, Musolff and Yang found the same shape at a bigger scale. With autonomous agents, commits rose by around 240% but software releases by around 30%.
This was never really about code
The authors say so themselves. The split between doing the work and managing or checking it is, in their words, “nearly ubiquitous across knowledge work settings”, and their model shows that “even if AI adoption increases the speed of implementation, output will be limited by the speed of management or review”.
Think about where review sits in your own organisation. The contract a lawyer has to check before it goes out. The report a manager signs off. The credit decision a committee approves. The clinical note a doctor has to read before it goes on the record. AI has made the first draft of all of these cheaper. It has done far less for the sign-off, and in many cases it has made sign-off harder, because the reviewer now has to check work they did not write and cannot fully trust. Deloitte had to refund part of an A$440,000 fee to the Australian government after a report it produced turned out to contain fabricated references. That is a review step failing under load.
I call this constraint migration: the task got faster, and the knot moved. I am not claiming the idea is new. Eli Goldratt’s theory of constraints said in 1984 that every system has one binding constraint at a time, and that when you break it you go back and find the next one. What AI changes is the speed. It can move the knot across a whole organisation in a quarter, and quietly, because every dashboard upstream of the knot goes green.
You don’t need a 655-firm study to see this in your own organisation, though. Pick one piece of work AI has sped up and follow it to the point where it is finished. Wherever it sits waiting longest is where your knot is now.
What the paper does not say
It does not say AI failed. Finished work did go up, by a modest amount. The authors are careful to say this is a working paper covering the first 12 to 16 months after adoption, mostly at medium to large established firms, and the data comes from clients of a single analytics vendor. Their own projection is that output rises substantially over the next few years, and by far the most if review gets faster at the same rate as coding.
Which is the point. The ROI is not missing. It is waiting for the organisation to catch up with the tool.
So where is the ROI?
It is in the step after the one you sped up, and getting it out is iterative by nature. There is no single redesign that releases it, because the moment you clear one knot the next one becomes the constraint. Goldratt built that into his method as a loop, and economists have a name for the delay: Erik Brynjolfsson and colleagues call it the productivity J-curve, where returns from a general purpose technology lag behind the slow, hard-to-measure work of reorganising around it.
The examples I have of this working are small, and they came round more than once before they held. In practice that loop looks like this:
- Measure finished work, end to end. If your AI dashboard reports prompts, lines, documents drafted or hours saved, it is reporting the stage that was never your constraint.
- Find where the knot moved. In my experience it tends to land in one of three places: verification (checking the output), coordination between teams, or decision authority (who is allowed to say yes). I built Flowscope for this step: you map how the work actually flows, and it shows you where the constraint sits now and where the waiting is. The free version gives you one map, which is enough to find your first knot.
- Redesign that one small part and prove it on real work. Smaller batches of work so review is lighter. More people trusted to review. A clear line between what needs a senior signature and what does not.
- Measure again, and expect the knot to have moved.
I will be straight about the gap. As far as I can find, nobody has yet published a firm-level study showing an organisation redesigning its review step and getting the gain back. The theory is strong and the Harvard model supports it, but the evidence for the fix is still being written, mostly by the organisations willing to go round the loop.
So the question I would ask of any AI programme right now is where the work is waiting, because that is where the return is.
Where has your constraint moved?
Sources: Chen & Stratton, “Artificial Intelligence in the Firm: Bottlenecks in Software Production” (Harvard, version of 14 November 2026); McKinsey, The State of AI in 2025; Humlum & Vestergaard, “Large Language Models, Small Labor Market Effects” (NBER w33777, revised 2026); Demirer, Musolff & Yang, “Writing Code vs. Shipping Code” (NBER w35275); Brynjolfsson, Rock & Syverson, “The Productivity J-Curve” (AEJ: Macroeconomics, 2021); Goldratt, The Goal (1984).