No one in this scene exists. The brand planner, her draft and her Thursday are made up for this article, although they are shaped by what a new study measured. It is mid-morning in the strategy unit of a large consumer-goods company. Copilot has turned last quarter’s sales notes into a two-page summary for a product relaunch in less than a minute. The summary reads well. It says that one market grew faster than the others, and the planner cannot tell where that claim came from. She does not send the summary. She opens her calendar and books twenty minutes with someone from supply chain whom she has never met. Then she books another twenty minutes with the finance analyst whom she has asked for help on every launch for three years.

The assistant has saved her an hour of writing. It has also given her two meetings that she did not have before.

What the study measured

Giuseppe Soda and Alessandro Iorio of Bocconi University and Leonardo Rizzo of SDA Bocconi School of Management posted their paper to arXiv on 23 October. It has not yet been peer reviewed. They studied one business unit of a large multinational in the consumer products industry. The unit implements strategic initiatives across the organisation, and the authors describe its work as highly interdependent and reliant on digital communication tools. They observed the unit from September 2024 to September 2025, the period when the company rolled out Microsoft Copilot.

The data came from Microsoft Viva Insights rather than from surveys. Viva Insights records calls, emails, meetings and chats across Microsoft 365 applications, and they were anonymised at the source. Copilot use was logged by a separate system. An action counted only when a worker actually engaged with what Copilot produced, for example by expanding a document summary or by reviewing a meeting recap that it had generated. The authors coded a worker as an adopter in the first month in which their use stayed above the 25th percentile for two consecutive months. Of 12,040 employees, they set aside 172 who already met the adoption threshold in the first month, because those people had no “before” period to compare with. That left 1,992 adopters and 9,876 workers who never crossed the threshold and who served as the comparison group.

The main outcome was what Viva Insights calls generated collaboration hours. These are the meeting, call, email and chat hours that a worker initiates with at least one other person in a month. The measure counts the worker who calls the meeting or starts the email thread. An invitation from someone else does not count. To map each worker’s network, the authors counted a tie only when one person directed at least an hour of collaboration to another person in a month.

Two designs pointing the same way

People who adopt a new tool early often differ from people who do not, and that difference alone could explain changes in how they work. The authors used two methods to deal with this problem. The first is a staggered difference-in-differences design, built on an estimator that Brantly Callaway and Pedro Sant’Anna published in 2021. It compares each group of adopters with workers who never adopted, over the same months. According to the authors, access followed organisational processes such as budget availability rather than individual initiative, which made staggered timing usable. In the months before adoption, the adopters and the comparison group moved in parallel, and the authors report pre-adoption effects close to zero.

After adoption, the two groups diverged. Over the first three months, adopters initiated about 1.43 more hours of collaboration per month than the comparison group would predict. The increase began in the month of adoption. Most of it came from meetings and calls rather than from email. The authors also report a broader average over all cohorts and months, and that average was smaller, at 0.95 hours, and not statistically significant. They explain that this average gives more weight to the earliest adopters, and they treat the three-month figure as their main result.

The networks changed as well. Adopters directed collaboration to about 2.2 more colleagues than before. Their networks became less clustered, which means that fewer of the people they worked with also worked with one another. At the same time, adopters concentrated more of their current collaboration time on colleagues with whom they had a long history. In other words, they reached further out and also leaned harder on their most trusted colleagues.

The second method asks a different question: does using Copilot more lead to more collaboration? It uses an instrument that combines two factors. One is the monthly pace of Copilot feature releases, which the authors built from Microsoft’s release notes. The other is the share of a worker’s collaborators who already used Copilot when that worker first appeared in the data. The authors report that the instrument predicted Copilot use strongly. By their estimate, a tenfold increase in monthly Copilot actions, for instance from one to ten, corresponds to about 3.3 extra hours of initiated collaboration per month. That estimate is only marginally significant. The same increase corresponds to roughly 20 additional distinct colleagues per month, and that estimate is firmly significant.

Why a writing tool would add meetings

The authors explain the result with an idea that Irene Unceta, Paula Subías-Beltrán and Oriol Pujol set out in Nature Machine Intelligence in August. They call the idea epistemic debt. When people use generative AI for cognitive work without taking time to understand the output, “a gap opens between what they present and what they can defend.” The gap grows as further work is built on top of it. Unceta and her co-authors locate the remedy in individual effort. The Bocconi team adds a social remedy: colleagues who question an output and ask its author to explain it. Copilot makes a draft cheap to produce. In the authors’ reading, it makes the draft more expensive to get accepted, because, as they write, generative AI “has raised, rather than lowered, the stakes of social acceptance.”

This also contrasts with an earlier natural experiment that used similar data. In 2022, Longqi Yang and colleagues published a study of 61,182 US Microsoft employees over the first six months of 2020, when the pandemic moved them to full-time remote work. They found that firm-wide remote work made collaboration networks more static and siloed, with fewer bridges between different parts of the company. Communication also shifted from synchronous to asynchronous channels. The Copilot adopters at the consumer-goods company showed a different pattern. Their networks became less clustered rather than more siloed, and most of their added collaboration came through meetings and calls.

The study also adds a different kind of evidence to the best-known workplace studies of generative AI. Erik Brynjolfsson, Danielle Li and Lindsey Raymond studied 5,179 customer-support agents and found that an AI assistant raised the number of issues resolved per hour by 14 percent on average, and by 34 percent for novice and low-skilled workers. That research measured output. The Bocconi paper measures the people around the output. The question is becoming more pressing. In a speech on 13 October, Philip Lane of the European Central Bank said that 52 per cent of euro-area workers now report using AI at work, compared with 26 per cent in 2024.

What the data cannot see

The authors are careful about what their data can show. They know who interacted with whom and for how long, but they do not know what people talked about. They list other explanations that fit the same pattern. Workers may have spent the time that they saved on drafting on meeting colleagues. Some outputs may simply require more coordination, even when no one doubts their quality. And some conversations may have been about Copilot itself rather than about its results. The study covers one business unit, and its effects are measured over the first three months after adoption. The authors note that in jobs where people depend less on one another, such as software development, the effect might go the other way, and they cite earlier research on software teams that points in that direction.

They also raise a concern about their own strong-tie finding. A colleague who scrutinises an AI-assisted draft helps to pay down the debt. A colleague who simply endorses the draft can spread the debt further, because others then build on it without checking it. Trusted colleagues offer informed judgment, the authors write, but “trust also lowers the scrutiny people apply to each other’s work.” Their data cannot separate checking from approval. “Distinguishing validation from endorsement is therefore a priority for future research.”

The paper’s most direct advice to employers is a single sentence: “Organizations that want collaboration to offset GenAI’s opacity need norms that reward questioning AI-assisted work and disclosing when it was used.” For a manager who reads the Copilot dashboard as a measure of time saved, the study suggests a second place to look. That place is the calendar, which is where this study found the change.

The made-up planner sends her summary on Thursday afternoon. The fast-growing market is still in it, but now it has a footnote that the supply-chain colleague supplied, about a warehouse that opened in the middle of the quarter. The finance analyst read the draft in five minutes and approved it. The planner knows which of the two meetings changed the document.

References

  • Brynjolfsson, E., Li, D., & Raymond, L. (2025). Generative AI at Work. The Quarterly Journal of Economics, 140(2), 889–942. nber.org
  • Callaway, B., & Sant’Anna, P. H. C. (2021). Difference-in-differences with multiple time periods. Journal of Econometrics, 225(2), 200–230. doi.org
  • Lane, P. R. (2026). The AI opportunity for Europe. Dinner remarks at the International AI Summit, Dublin, 13 October. European Central Bank. ecb.europa.eu
  • Soda, G., Iorio, A., & Rizzo, L. (2026). The Effects of Generative AI on Collaboration and Social Networks Among Knowledge Workers. arXiv. arxiv.org
  • Unceta, I., Subías-Beltrán, P., & Pujol, O. (2026). The epistemic debt of generative AI. Nature Machine Intelligence, 8, 1328–1330. nature.com
  • Yang, L., Holtz, D., Jaffe, S., Suri, S., Sinha, S., Weston, J., Joyce, C., Shah, N. P., Sherman, K., Hecht, B., & Teevan, J. (2022). The effects of remote work on collaboration among information workers. Nature Human Behaviour, 6, 43–54. nature.com