One Study Says AI Widened the Team. Another Says It Emptied the Room.

The study that measured conversations instead of output

Most of what you read about AI at work measures output: documents drafted, tickets closed, minutes saved. A paper posted on 23 October 2026 by Giuseppe Soda and Alessandro Iorio of Bocconi University and Leonardo Rizzo of SDA Bocconi asked a different question. When an AI assistant lands in a workplace, what happens to the links between people?

The setting is unusually good for that question. The authors got passively collected behavioural data — not a survey, not a diary — on 12,040 employees at one business unit of a large consumer-products company, from September 2024 to September 2025. During those thirteen months the company rolled out Microsoft Copilot in stages, which gave the researchers a natural before-and-after with a control group. Their final estimation sample is 11,868 workers: 1,992 who reached sustained use and 9,876 who never did. They defined an adopter as someone whose Copilot activity passed the 25th percentile of observed use for two consecutive months, so that a curious click does not count as adoption (arXiv:2610.32461).

The data come from Microsoft Viva Insights, which logs who meets, mails and messages whom. To keep incidental contact out, the authors count a tie between two people only when it clears roughly an hour of interaction in a month. Collaboration is then split into what a worker initiates and what a worker receives, because the two mean different things: the first is outreach, the second is being needed. They estimate effects two ways, with a staggered difference-in-differences design and with an instrumental-variable model that uses the timing of Microsoft's feature releases as a shove, and the instrument is strong enough to trust (Kleibergen-Paap F of 80.43 against a threshold of 16.38).

What they found is the opposite of the story most owners carry around. Copilot did not turn a team into a set of isolated people talking to a machine.

Collaboration went up, and the network opened out

Adopters spent more time in contact with colleagues after they started using the tool, not less. The average effect over the first three months is an extra 1.43 hours of self-initiated collaboration per month, with a confidence interval from 0.45 to 2.41 hours that does not cross zero. Count the contact other people started with them too, and the number is 0.87 extra hours a month. Neither is enormous in a working week. What makes it interesting is where the extra hours landed: the rise is concentrated in live conversation — meetings and calls up 1.26 hours — against a much smaller 0.17 hours in email.

The structure of the network moved further than the hours did. After adoption, the average worker reached 2.18 more distinct colleagues per month. The clustering coefficient, the measure of how much your contacts know each other, fell by 0.009. Both point the same way: people were stepping outside their usual circle instead of retreating into it.

Then comes the part that gets misunderstood, including by the researchers' own press coverage. Adoption also raised the share of collaboration poured into strong ties — colleagues with a long shared history — by a small but measurable amount. So the pattern is subtler than "everyone talks to everyone more": a widening at the edges and a thickening at the core at the same time. Workers reached further out for information and leaned harder on the two or three people whose judgement they already trust.

Read that as a warning and as a plan. A firm where knowledge ends up funnelled through a handful of trusted people is efficient right up to the week one of them leaves. The same data also says those trusted people are doing unpaid quality control, and nobody has put it in their job description.

The other study found the room going quiet

Now the counterweight. Michael Nagle and co-authors, in a Harvard Business School working paper, looked at 187,489 developers on GitHub across two years, from June 2022 to June 2023. Copilot's general launch that June created a threshold — free access for developers above a certain point in an internal ranking — which let them treat access as close to random. The result on collaboration is blunt: developers with Copilot worked in repositories with 17 fewer peers, against a baseline of 22. That is a 79.3% drop, reported in the paper's own appendix (HBS Working Paper 25-021).

Their time moved the same way. Coding rose 12.4% and project-management work — code review, issue triage, asking someone to look at a patch — fell 24.9%, from a base of roughly 37% of their working time. The mechanism the authors describe is not mysterious. In open source you contribute by forking someone else's project, changing it, and asking the maintainer to merge your change. That handshake carries friction by design. Copilot let developers skip it and contribute unilaterally instead, so requests for review, assignments to other people and issue subscriptions all fell.

Their summary line is worth keeping: developers with AI access "bypass collaboration frictions." If you run a firm where the work is done alone and the coordination is an obstacle, that is a feature you should be buying. If you run a firm where the coordination is the work, it is the thing you will end up paying for.

One more finding from that paper should stop an owner mid-sentence. The shift toward independent work was largest for the less-experienced developers. Nagle's own reading, given to MIT Sloan, is that replacing junior staff with AI to save money is "a profound strategic error": the people who learn the fastest from the tool are the ones you would be cutting. At 187,489 developers, this is the best evidence available that AI does not shrink the need for a team — it changes which part of the team's time is worth paying for.

Why both results are true at once

The two datasets do not contradict each other. They describe two different jobs that AI does to a team's calendar, and which one dominates depends on the work.

Every task contains two kinds of coordination. There is coordination during production: asking a colleague for a first draft, borrowing someone's phrasing, getting a template. AI eats that kind. It produces the draft, so you stop asking. Then there is coordination after production: getting someone to read what came out, agree it is right, and be willing to put their name near it. AI does not eat that kind at all. If anything it feeds it, because the output arrives fluent, plausible and occasionally wrong in a way that only a person with context can catch.

The developers in Nagle's sample were doing work where the second kind of coordination was an obstacle to be bypassed: code either compiles and passes review, or it does not. The consumer-goods employees in the Bocconi sample were doing work where the second kind is the job itself — a recommendation, a plan, a number that goes to a manager. One group lost coordination. The other gained it.

The Bocconi authors give the mechanism a name borrowed from an August 2026 Nature Machine Intelligence comment: epistemic debt. Spend time with generative AI without spending the time to understand what it produced, and a gap opens between what you present and what you can defend. Build more work on top and the gap widens. AI compresses the time it takes to produce an answer; it does not compress the time it takes anyone to understand that answer well enough to stand behind it. Borrow the time saved on production and you owe the difference at the other end.

Which is why the finding that matters most is not "more" or "less". It is that the bottleneck moved. In the Bocconi data the added hours went into live conversation rather than email — the form you use when you need to argue something through quickly. That is what paying down epistemic debt looks like on a calendar.

Run the arithmetic on a firm of ten

Abstracts do not tell you whether to buy a seat. Numbers for your own business do, so here is the sum worked out for a ten-person heating contractor that drafts three quotes a day. Every figure below is my assumption, plainly labelled, not a measurement from a study. Swap in your own.

The saving. Three quotes a day at 40 minutes of drafting is two hours of someone's morning. If the AI tool gets a first draft out in twelve minutes, you save 28 minutes a quote, which is 84 minutes a day, or roughly seven hours a week across the firm.

The bill. A quote is a price with your margin inside it, so nothing goes out unchecked. Give the estimator 20 minutes per quote to verify the parts list, the labour rate and the exclusions, and you have spent 60 minutes of the 84. Add a colleague's fifteen-minute read on one quote in three, because the expensive ones get a second pair of eyes, and the bill is 75 minutes.

The result. Nine minutes a day. On the best case, where the tool already knows your price list and your standard exclusions and the check drops to five minutes a quote, you net 69 minutes a day, about 5.75 hours a week, which is a working day a month. That is a real gain and it is not a headcount. The gap between 9 and 69 minutes is entirely the cost of checking, which is why the two questions worth asking a vendor are about the price list and the audit trail, not the model.

Now run the same firm as ten independent surveyors who each handle their own jobs from first call to final invoice. Their quotes need no second opinion, because the person who priced the job is the person who will do it. By the developer logic they should see collaboration fall, and they probably will. The trap is what leaves with it: the informal "how did you price that one?" conversation across the desk, which is how a twenty-six-year-old learns to price. You will not notice the loss for two quarters, and by then you will have attributed it to the hiring market.

Measuring it in thirty days, without buying software

1. Pick two people and one task. Not the whole firm. A task that crosses at least two people, like pricing, onboarding a client, or handling a complaint.
2. Count the conversations first. For one week before the tool arrives, have both people tally every time they ask a colleague something in order to finish that task. A shared note, one line each: date, who, how many minutes, what it was about. That is your baseline.
3. Log the checking separately from the saving. Two columns, not one. Minutes saved drafting, minutes spent verifying. Merge them and the whole exercise becomes a story you tell yourself.
4. Count who people ask. The Bocconi result was not just "more contact" but "more contact, concentrated on a few trusted people". If your tally shows every question going to the same two names, you have a single point of failure, and it predates the AI.
5. Look at day 30, not day 3. The measured effects peak in the first three months, which is the window when habits are still forming. A saving that only exists in week one is a novelty, not a change.
6. Write down the decision. Keep, change or drop, per task. A pilot that ends with "people seem positive" has taught you nothing you can act on.

When more conversation is the problem, not the win

A rise in collaboration sounds like good news. It is not automatically, and there is an experiment that shows why. Fabrizio Dell'Acqua, Bruce Kogut and Patryk Perkowski put teams of two into "laboratory firms" playing a cooperative console game, then replaced one human on a random subset with the game's built-in AI, which outperformed most people at the task. Team output fell. Teams that received the automated player logged 1.28 more coordination failures per round, and 2.51 more in the round immediately after the swap. Trust and individual effort dropped too, with no change in anyone's pay (Review of Economics and Statistics, July 2025).

The damage was not evenly spread. Low- and medium-skilled teams lost the most; high-skilled teams came out slightly ahead, because their routines were strong enough to absorb a new member who did not follow instructions the way a person does. Read that as a straight warning for a small firm: your best team will survive a botched rollout, your weakest one will not, and the weakest one has the least slack to absorb the disruption you just introduced.

Three more reasons to hold back.

The evidence has a horizon. The Bocconi paper follows one business unit for thirteen months and its authors are explicit that the collaboration effects are measured over the first three months after adoption, that they cannot see what anyone discussed, and that they cannot separate "a colleague validated this" from "a colleague rubber-stamped it". They say plainly it is unclear whether the results apply to other industries or job types (arXiv:2610.32461).

Borrowing a draft is easy; borrowing a process is not. In a six-month randomised experiment with 7,137 workers at 66 firms, heavy users of Copilot spent 3.6 fewer hours a week on email, a 31% cut, and gained four hours of focus time. Time in meetings did not move at all. The authors' explanation is the important part: the tool changed the behaviours a person can alter alone, and left untouched the ones that require everyone to agree (arXiv:2504.11436). If your plan for AI requires the team to change how it works together, a licence will not deliver it.

Some of this you are already being told to buy. A field experiment with 776 professionals at Procter & Gamble found that one person with AI matched a two-person team's solution quality, and that AI pushed people across functional lines: R&D staff started proposing commercially viable ideas, commercial staff started proposing technically sound ones (POID Working Paper 118). That is a genuine argument for the tool. It is also the argument every vendor will make to you, which is why the case against needs saying out loud: if the only thing your team's conversations do is relay information that the model now supplies directly, you are about to discover that some of those conversations were not worth the time. The honest test is whether a conversation ends with a decision somebody owns. If it does not, cutting it is not a loss.

So: do not roll this out across the whole firm in month one. Pick the task where the output has a verifiable answer, where two people already check each other's work, and where a mistake costs money you would notice. Leave the ambiguous, judgement-heavy work alone until you have watched the first task for a month.

Which one is your firm? A table you can answer

The two studies point in opposite directions because the firms in them were doing opposite kinds of work. Before you spend anything, place your business in one of these four rows. Be honest: most owners will find they are in two of them at once, which is exactly why a firm-wide rollout is the wrong shape.

The backdrop is not in doubt. In a survey of 1,015 SME leaders published on 19 October 2026, the Social Market Foundation found 79% of small and medium firms already using AI in some capacity, with micro firms trailing at 60% against 86% for small and 92% for medium. Of those using it, 87% reported a positive impact, time savings cited by 88% and quality of work by 79%. The same survey found AI adoption 1.8 times more likely to increase SME employment than to decrease it, and its authors conclude that small firms are using these tools to make existing staff more productive rather than to replace roles (SMF, Byte-sized growth).

Which leaves the question the studies cannot answer for you: when the collaboration does grow, is it validation or is it endorsement? A colleague who interrogates the output, asks for the reasoning and tests an assumption is paying down debt. A colleague who glances at it and says "looks fine" is spreading it. The difference is not in the software. It is in whether your firm treats "where did this number come from?" as a normal question or as an insult.

Frequently asked questions

Does AI make teams collaborate less?

Sometimes, and the deciding factor is whether the work is independent or interdependent. In a study of 187,489 developers, Copilot access cut the number of peers they worked with by 79.3%, because code has a checkable answer and a colleague's review was friction. In a study of 12,040 employees doing interdependent knowledge work, Copilot use added 1.43 hours of collaboration a month and 2.18 more colleagues. Neither result generalises to the other setting.

How much extra collaboration should I budget for?

Measured effects are small in absolute terms: roughly 1.4 extra hours per person per month in the interdependent case, concentrated in meetings and calls rather than email. What matters more is the shape. Networks widened and loosened, and time shifted toward the few colleagues people already trusted. In a firm of ten, that means one person's calendar absorbing the questions, which is a scheduling problem before it is a software problem.

Should I replace junior staff with AI if the tool does their work?

The researchers who measured the largest AI productivity gains say no. In the developer study the shift toward independent work was biggest among less-experienced people, and the lead author's own advice is that cutting entry-level roles to save money is a strategic error, because the people who learn fastest from the tool are the ones being cut. A separate experiment at Procter & Gamble found AI helped people with less domain experience reach the performance of experienced colleagues, which is an argument for keeping them, not removing them.

What should I measure to know whether AI is helping my team?

Two columns, kept apart: minutes saved producing and minutes spent checking. Then a third measure that most firms skip, which is who is being asked. If every question funnels to the same two people, you have built a single point of failure and it will not appear on any invoice. Thirty days is enough for a first read; three months is where the pattern settles.

Is AI worth it for a team of five?

For a narrow, checkable task, often yes, and the value shows up as time rather than headcount: 5.75 hours a week on a good case in the quote-drafting example above, nine minutes a day on a bad one. For anything where the output needs a judgement call and nobody in the firm has the depth to make it, the tool mostly adds confident text to a decision you were not equipped to make. That one is worth skipping.