AI didn’t shrink workplace ties—it made workers lean on more people, and trust old ones more
AI use coincided with more collaboration and wider workplace networks—but workers also leaned more heavily on colleagues they already trusted.
11,868 workers, one striking pattern: AI users reached more colleagues and leaned harder on trusted ties.
Workers who adopted Microsoft Copilot spent more time collaborating with colleagues, reached out to more people, and built less tightly clustered networks. Yet they also concentrated more of their collaboration on colleagues with whom they already had strong histories. In a study tracking more than 11,800 knowledge workers through a staged rollout, generative AI did not appear to make human collaboration obsolete. It changed what workers needed one another for—and which colleagues they trusted to help.
That is a more complicated story than either “AI replaces coworkers” or “AI makes everyone collaborate more.” A system that drafts text, summarizes meetings, or produces an analysis can reduce the need to ask someone for a first answer. But an answer generated quickly is not necessarily one a worker understands, can defend, or can persuade others to use. The research suggests that as AI takes on some of the work of producing ideas, people may turn to one another more for checking, interpreting, and making those ideas credible.
The study’s central estimate is an average increase of 1.430 hours in monthly collaboration initiated by adopters after they began using Copilot, compared with the change among workers who never adopted it. That increase was statistically significant, and the researchers found no evidence that adopters’ collaboration was already rising faster before adoption. The result is not proof that AI will increase collaboration in every workplace. It is evidence from one large organization that, at least during the first months of adoption, AI-assisted work can bring more human interaction rather than less (Soda et al., 2026).
The Science
Giuseppe Soda, Alessandro Iorio, and Leonardo Rizzo studied a large multinational consumer-products company, focusing on one business unit whose work involved substantial coordination across the organization. The researchers followed employees from September 2024 to September 2025, as Microsoft Copilot was introduced in stages. Copilot is an AI assistant built into Microsoft 365 tools such as Outlook, Teams, and Word.
The analysis began with records for 12,040 employees. After excluding workers who were already sustained Copilot users at the beginning of observation, the main adoption analysis covered 11,868 people: 1,992 who became sustained users and 9,876 who did not meet the study’s adoption threshold. The authors defined adoption as exceeding a minimum activity level for two consecutive months, rather than counting a single experiment with the tool as adoption.
They linked two kinds of passively collected workplace records. One captured Copilot activity; the other recorded digital collaboration through meetings, calls, emails, and chat. The authors say the systems record these separately, reducing the risk that an apparent link between AI use and collaboration is simply an artifact of how the data were collected. The records were anonymized at source, and the study received institutional ethics approval.
One central measure, “generated collaboration,” counted time the worker initiated with other people—for example, by arranging a meeting or starting an email exchange. It included overlapping digital activity only once. This measure matters because it captures the worker actively creating opportunities for collaboration, not just being contacted by others. The researchers also examined total collaboration, including interactions initiated by colleagues.
The team represented collaboration as a changing social network: workers were the nodes, and an interaction of at least roughly an hour in a month counted as a tie. Three measures described how those networks changed. Out-degree counts how many colleagues a person reaches out to. Clustering estimates how often those colleagues are connected to one another, so lower clustering means a person’s contacts are less likely to form one close-knit circle. Preferential reinforcement tests whether someone devotes more time to colleagues with whom they have invested heavily in the past.
To separate the effect of adoption from the possibility that especially collaborative employees were simply more likely to use Copilot, the authors used two complementary quasi-experimental approaches. The first compared changes for workers who adopted at different times with changes for never-adopters, while accounting for stable differences between people and common shifts over time. This staggered difference-in-differences design depends on an important assumption: without adoption, the two groups would have followed similar collaboration trends, conditional on measured characteristics.
The second approach used a fixed-effects instrumental-variable model. In plain language, it used the timing of Copilot feature releases and workers’ prior exposure to Copilot users in their collaboration networks to predict changes in their own use. The authors ran diagnostic checks among never-adopters, where their own Copilot use could not explain a relationship between the instrument and collaboration. Those checks did not find evidence that the instrument predicted collaboration in that group. Such checks strengthen the case for the design, but they cannot prove that every assumption behind it is true.
What They Found
The clearest result is that adoption coincided with more worker-initiated collaboration. In the difference-in-differences analysis, the average pre-adoption effect was 0.074 hours and statistically indistinguishable from zero. After adoption, the average effect was 1.430 hours; its reported 95% confidence interval ran from 0.451 to 2.410 hours. The estimate rose at adoption and remained positive the following month: 1.263 hours in month zero and 1.796 hours in month one. These are changes in monthly collaboration time, not a claim that every adopter spent exactly that much extra time with colleagues.
Collaboration rose after Copilot adoption
Difference-in-differences estimates for average pre- and post-adoption effects on worker-initiated collaboration time, in monthly hours.
| Label | Value |
|---|---|
| Pre-adoption average | 0.074 hours/month |
| Post-adoption average | 1.43 hours/month |
The finding also appeared when the researchers widened the measure to include interactions initiated by other people. Total collaboration rose by an estimated 0.865 hours on average after adoption, with a reported 95% confidence interval of 0.557 to 1.174 hours. That consistency matters: the result was not limited to workers sending more invitations or initiating more exchanges themselves.
The instrumental-variable results add a more detailed, but less straightforward, picture. In its model, the coefficient for Copilot activity and generated collaboration across all modes was 1.552, but its confidence interval crossed zero and the estimate did not reach conventional statistical significance. The estimates for synchronous generated collaboration—such as meetings and calls—were also not statistically significant. By contrast, the estimate for asynchronous generated collaboration, including email and chat, was 0.223 and statistically significant. For total collaboration, the reported estimates were positive and statistically significant across all modes, synchronous activity, and asynchronous activity. These coefficients describe the model’s relationship per unit of Copilot activity; they should not be read as a simple, universal number of hours gained by every user.
Estimated collaboration effects varied by channel
Fixed-effects instrumental-variable coefficients for Copilot activity across collaboration modes. The coefficients are not directly interpretable as a universal number of hours per user.
| Label | Value |
|---|---|
| Generated · all modes | 1.552 |
| Generated · synchronous | 1.329 |
| Generated · asynchronous | 0.223 |
| Total · all modes | 1.593 |
| Total · synchronous | 1.234 |
| Total · asynchronous | 0.417 |
The networks themselves changed in three distinct ways in the instrumental-variable analysis. Workers’ out-degree increased, with a reported coefficient of 9.396. That points toward outreach to more colleagues. Clustering fell by 0.030, suggesting that a worker’s contacts became less tightly connected to one another. Preferential reinforcement increased by 0.031, indicating a stronger tendency to allocate collaboration time to people with whom the worker had built stronger ties in the past. All three estimates were statistically significant in the reported models.
Together, these measures describe a network that broadens without becoming indiscriminate. A worker may reach out to a wider set of colleagues, including people outside the usual tight circle, while still leaning especially heavily on a few trusted relationships. AI appears associated both with more open pathways through the organization and with a deeper reliance on established ties.
The paper’s figures make the pattern visible at different scales. Figure 1 plots workers’ outcomes around the month they began using Copilot, showing the trajectories the authors examined before and after adoption. Figure 2 presents the event-study estimates, including the near-zero average pre-treatment result and the positive post-adoption collaboration estimates. Figure 3 compares average post-adoption and overall treatment estimates across collaboration and network outcomes; the paper standardizes those coefficients for comparison, while the underlying outcomes remain different measures.
Why This Changes Things
The debate about workplace AI often treats a colleague and an AI assistant as competing sources of help. Need a draft? Ask the system, not a person. Need a summary? Let the model produce one. On that view, AI’s organizational effect should be fewer consultations and weaker social networks. The study identifies a different possibility: replacing some exchanges involved in producing an initial answer may create or intensify exchanges involved in judging what that answer means.
A generated answer is not automatically an organizationally usable answer. Someone may need to verify it, explain how it was produced, adapt it to local knowledge, or persuade a team that it is safe to act on. The authors connect this to “epistemic debt”: the gap that can open when a person accepts an output faster than they can understand or defend it. That debt is not necessarily settled alone. Colleagues can challenge an assumption, supply missing context, or help turn a plausible-looking result into a decision others will stand behind.
This gives AI a social role that is easy to miss if we measure only individual productivity. A model can shorten the time between a question and a first draft, while leaving intact—or even increasing—the time needed for interpretation, checking, and agreement. The bottleneck may move rather than disappear. If that is happening, organizations that judge AI’s success only by how quickly individuals produce documents could overlook the human work required to make those documents useful.
The network results also matter because collaboration is not simply a count of messages. Who connects to whom determines how knowledge moves. A tightly clustered network can support trust and rapid coordination, but it can also keep information circulating within a familiar group. A broader set of contacts can expose people to expertise beyond their immediate circle. The observed mix—more contacts, less clustering, stronger reliance on historically important ties—suggests expansion across the network alongside reinforcement within it.
This is not automatically a public-good outcome. Stronger ties can help people verify unfamiliar AI outputs with trusted colleagues, but concentrating attention on established relationships may also burden those colleagues or make it harder for newer workers to gain access. Broader networks may spread knowledge, but they can bring more coordination costs. The paper establishes changes in observed interaction patterns, not whether those changes improved the quality of decisions, employee well-being, productivity, or fairness.
The study also helps distinguish AI from earlier workplace technologies. Email and collaboration software mostly made it easier to transmit information and find colleagues. Generative AI can produce a first version of the information itself. That makes the relationship between technology and social interaction less predictable: a tool may substitute for one kind of consultation while generating new reasons to seek another. The results imply that “AI versus people” is often the wrong unit of analysis. The consequential question is how AI reshapes the chain of work connecting an initial output to a shared decision.
There are important limits. This is one business unit in one multinational company, with collaboration observed through Microsoft’s digital tools. Conversations in hallways, in-person problem-solving, or work in other platforms may be missed. The study’s network ties require a substantial level of recorded interaction, and its outcomes reflect the first several months after adoption rather than a mature, long-run equilibrium. It cannot tell us whether workers eventually settle into new routines, whether other AI systems produce the same pattern, or how the results vary across occupations and organizational cultures.
Causal claims also depend on assumptions. The difference-in-differences design is more informative than a simple comparison of users and non-users, but it still relies on treated and untreated workers having comparable underlying trends. The instrumental-variable strategy offers a second lens, and its placebo checks are reassuring, but its interpretation depends on feature releases and prior network exposure affecting collaboration through the worker’s own AI use rather than some other route. The authors acknowledge that diagnostic evidence cannot settle that question definitively.
What’s Next
The next step is to find out what the extra collaboration is for. Are workers asking colleagues to check factual claims, explain an AI-generated analysis, adapt a draft to a client, or simply manage more work because AI has raised expectations for speed and output? The current data show the pattern of interaction, not its content or value. Interviews, task-level records, and measures of decision quality could reveal whether the additional contact is productive validation or avoidable coordination overhead.
Longer follow-up would also show whether the network effects persist. A short-term burst of questions after a new tool arrives may settle into a different rhythm. Or AI may permanently change who becomes a go-to expert, how newcomers find help, and which teams act as bridges between organizational groups. Replicating the design across industries, job types, and AI systems would help establish how much of this result reflects Copilot, this company’s rollout, or a more general feature of generative AI at work.
For managers, the implication is not simply to encourage more meetings. It is to recognize that trustworthy use of AI needs social infrastructure: clear ways to check outputs, space to explain uncertainty, and credit for the colleagues who help turn machine-generated material into sound judgment. Workplaces should watch not only how many tasks AI completes, but also whether expertise travels more widely and whether validation work becomes concentrated on a small number of trusted people.
The study offers a grounded counter to the fear that AI will automatically hollow out workplace relationships. In this setting, the early evidence points instead to a reorganization: workers reached more colleagues, their networks became less closed, and trusted ties mattered more. As generative AI spreads, the key question may not be whether humans remain in the loop, but how organizations design the loop so that broader collaboration leads to better shared decisions.
Technologies that reduce the effort required to produce answers may simultaneously increase the effort required to have those answers recognized as meaningful and credible by others.
Sign in to join the conversation.
Comments (0)
No comments yet. Be the first to share your thoughts.