Fil rouge · N° 04 · AI & teams
Why doesn't AI boost teams as much as individuals?
On my own, with an AI agent, I build in one evening what would have taken me a week. A learning app for my daughter, a screen for the whole house: I see the gain every evening.
In a large organization, it’s another story. You hear the same little tune everywhere: AI transforms the work of individuals, much less that of teams. Rumor or reality? I wanted to get to the bottom of it, with recent studies rather than impressions.
It’s not a rumor
Individual gains exist, and they have been measured. In field experiments at Microsoft, Accenture and a Fortune 100 company, developers equipped with an assistant complete 26% more tasks. In customer support, AI adds 15% more issues resolved per hour.
But the further you move along the process, the more the gain melts away.
Commits
Projects
Releases shipped
+240%
+80%
+30%
An NBER study published in September followed more than 500,000 developers. With autonomous coding agents, their commits jump by 240%.
The number of projects, by 80%.
Releases actually shipped, by 30%.
Demirer, Musolff & Yang, “Writing Code vs. Shipping Code,” NBER w35275, September 2026. More than 500,000 developers on GitHub, cumulative effects of autonomous coding agents.
The numbers
| Commits | +240% |
|---|---|
| Projects | +80% |
| Releases shipped | +30% |
Same finding inside companies. A Harvard working paper, updated this month, covers 720,000 people in 655 companies. With agents, pull requests rise by 36%, but resolved tickets by 31% at most.
At the scale of companies and of the economy, the signal becomes very faint. Surveyed in seven countries, thousands of executives put the gain at 0.2 points of productivity per year. US productivity grew 2.1% in 2025, less than in 2024. As for the ECB, it doesn’t see an overall signal yet.
A chain moves at the pace of its weakest link
It’s the best explanation I’ve found, and it comes from the NBER researchers themselves: the weak-link hypothesis.
Understand
Decide
Write
Review
Validate
Ship
×5
WEAKEST LINK
CHAIN OUTPUT
Making software, like almost any teamwork, is a chain: understand the need, decide, write, review, validate, ship.
AI makes only one of these links dramatically faster: writing.
And a chain never goes faster than its slowest link.
A chain’s output flows at the pace of its slowest link (theory of constraints). Weak-link hypothesis: Demirer, Musolff & Yang, NBER, 2026.
Their estimate puts it bluntly: AI and human effort are very poor substitutes for each other. Speeding up one without the other barely changes the outcome.
Why it’s obvious when you’re on your own
On my own, the chain is short. I decide, I write, I review, I ship: it all fits in one head, and no link waits on anyone else.
And the gain feels all the stronger because perceived speed is deceptive.
Expected
Perceived
Measured
−24%
−20%
+19%
≈ 40-point gap
← faster
slower →
In a randomized trial by METR, experienced developers expected to go 24% faster with AI.
Afterwards, they still believed it.
In reality, they had taken 19% longer.
METR, randomized controlled trial, July 2025. 16 experienced open-source developers, 246 tasks, early-2025 tools. METR considers that this figure no longer reflects current tools; the gap between perception and measurement remains the solid lesson.
The numbers
| Expected | −24% |
|---|---|
| Perceived afterwards | −20% |
| Measured | +19% |
The tools have improved a lot since, but the lesson stands: perceived speed is not a measurement.
Where team chains break
Lines of code
Commits
Pull requests
Resolved tickets
Review time
+44%
+36%
+36%
≤ +31%
+56%
1. Review becomes the weak link. In the Harvard study, agents inflate code output: +44% lines, +36% commits and pull requests. But at most +31% resolved tickets.
And the time between submitting a code review and its approval grows by 56%. More code means more code to review, in bigger pull requests, with more changes requested.
Chen & Stratton (Harvard), “Artificial Intelligence in the Firm: Bottlenecks in Software Production,” version of November 1, 2026. 720,542 people in 655 companies. Effects of coding agents, on top of those of assistants.
The numbers
| Lines of code | +44.0% |
|---|---|
| Commits | +36.0% |
| Pull requests | +35.7% |
| Resolved tickets | at most +30.7% |
| Review time | +56.1% |
Outside of code, it’s “workslop”: a document produced fast, which costs time to whoever receives it. According to a Workday survey, about 40% of the time saved goes back into rework. The work doesn’t disappear: it shows up in another form.
2. The coordination links don’t budge. In a six-month trial run by Microsoft with 7,137 employees, Copilot users saved about two hours of email a week. Neither their meetings nor the nature of their tasks changed.
3. Some links are invisible. This summer, I analyzed the development activity of an organization I know well. The finding struck me: the problem wasn’t how people code, but everything that isn’t written down anywhere. The intent behind a change, sharing between teams, verification, the memory of who knows what. AI only sees what is written. It can’t speed up a link it can’t see.
4. The hidden gain. One link remains, almost a taboo: fear. I imagined it pushed teams to drag their feet, to buy themselves a few months. The studies seem to point to something else. The fear is real: depending on how the question is asked, between one in five and two in five employees think AI threatens their job, and those who use it most are the most worried. But it produces little refusal, and a lot of silence. In a KPMG and University of Melbourne survey across 47 countries, 42% of the employees surveyed say they avoid, at least occasionally, disclosing that they used AI. They’re not entirely wrong: in experiments published in PNAS, people who admit to using it are seen as lazier and less competent. So part of the gain isn’t lost in the chain: it never enters it.
As for the reprieve, it exists, but it owes little to foot-dragging. For now, companies are mostly adjusting by hiring fewer beginners, not by letting go of the people already in place. According to Stanford, employment of 22- to 25-year-olds in AI-exposed occupations has fallen 19% behind that of young people in less exposed ones. Among more experienced workers, no gap. Dragging your feet does little to protect the people in place, and nothing at all for the ones who never get hired.
Shorten the chain
The result that struck me most comes from Procter & Gamble.
Alone without AI
Pair without AI
Alone with AI
Pair with AI
0
+0.24
+0.37
+0.39
+0.02
776 professionals worked there on real innovation challenges, alone or in pairs, with or without AI.
One person alone with AI did as well as a team of two without AI.
And the team with AI added almost nothing more.
Dell’Acqua et al., “The Cybernetic Teammate,” NBER w33641, 2025. One-day field experiment at Procter & Gamble; teams pairing one R&D person and one commercial person. Solution quality, in standard deviations above one person alone without AI.
The numbers
| Alone, without AI | 0 (baseline) |
|---|---|
| Pair, without AI | +0.24 |
| Alone, with AI | +0.37 |
| Pair, with AI | +0.39 |
There, AI didn’t speed up the team: it made one link unnecessary, the handoff between two people. An organization built on teams, handoffs and sign-offs doesn’t capture this gain until it redesigns its chain.
I live both extremes every week. On one side, an established company: large teams, processes honed over years. By the time you’ve followed them all, even a small feature takes several weeks, at the very least, to reach production. On the other, an AI-native startup I co-founded: we sometimes push several releases over a single weekend.
On both sides, I work with AI. So the gap doesn’t come from executing individual tasks, but from the length of the chain. The startup was born with AI: its chain was designed around it, with few links and almost no handoffs. The established company, for its part, added AI to a chain designed for another world.
AI-native companies clearly have a head start. In agility at least, and at first glance: those processes aren’t there by accident. Robustness is another question, and probably another thread to pull.
What I take away
We delegated writing, the easiest link to delegate. Without really deciding to, we kept everything else as it was: review, coordination, intent.
Faced with a weak link, there are only two options: strengthen it or remove it.
- Strengthen review instead of drowning it: smaller changes, tests, assisted review.
- Write down what wasn’t: intent, decisions, the team’s memory. Invisible links become visible, AI included.
- Remove handoffs: smaller teams, roles that overlap more.
- Measure the chain, what gets shipped, not the link, what gets produced.
It’s not inevitable. Electricity took decades to transform factories: the shop floors had to be redesigned around it. At the end of 2026, we’re probably at the bottom of that curve.
Faster alone, then. Not yet faster together, until we redesign the chain.
So in your team, what’s the weakest link?