---
id: "lib-2026-10-23-the-effects-of-generative-ai-on-collaboration-and"
type: "library"
title: "The Effects of Generative AI on Collaboration and Social Networks Among Knowledge Workers"
lang: "EN"
org: "arXiv (Bocconi, SDA Bocconi)"
source_type: "Paper"
date: "2026-10-23"
week: "2026-W44"
url: "https://nosov.ai/library/lib-2026-10-23-the-effects-of-generative-ai-on-collaboration-and/"
original_url: "https://arxiv.org/abs/2610.32461"
tier: "A"
tier_name: "Strong evidence"
tier_reason: "Bocconi researchers using passively logged behaviour of more than 12,000 knowledge workers over 13 months, with a staggered difference-in-differences design, a shift-share instrument and spillover checks; strong method, one firm."
summary_basis: "full-text"
original_minutes: "40"
tags: ["future of work", "enablement", "microsoft copilot", "collaboration networks", "field evidence"]
---

# The Effects of Generative AI on Collaboration and Social Networks Among Knowledge Workers

Tier A (Strong evidence): Bocconi researchers using passively logged behaviour of more than 12,000 knowledge workers over 13 months, with a staggered difference-in-differences design, a shift-share instrument and spillover checks; strong method, one firm.

Summary based on the full text.

## In short

A preprint (arXiv, 23 October 2026) by Giuseppe Soda and Alessandro Iorio of Bocconi University and Leonardo Rizzo of SDA Bocconi. It tests whether generative AI reduces the need to consult colleagues. Using Viva Insights logs from one business unit of a large multinational during its staged Copilot rollout, it finds the opposite: after adoption, workers spent more time collaborating, reached more colleagues, and had less clustered networks. They also concentrated more time on colleagues they already had strong histories with. The authors read this as validation-seeking: AI output is fast to produce but needs others to check and accept it.

## Key numbers

- 1.430 hours: average monthly rise in self-initiated collaboration time over the first three months after Copilot adoption (95% CI 0.451 to 2.410)
- 2.2: additional colleagues adopters collaborate with each month (overall ATT 2.179)
- 1,992: adopters in the estimation sample, against 9,876 never-adopters
- 13 months: observation window, September 2024 to September 2025

## What it reports

- Self-initiated collaboration hours rose by 1.430 hours a month on average in the first three post-adoption months; the overall ATT across all cohorts, 0.948, is not significant.
- Total collaboration hours, including interactions others started, rose by 0.865 hours a month.
- Both synchronous (meetings, calls: +1.257) and asynchronous (email: +0.173) collaboration rose, with most of the short-run gain in synchronous contact.
- Out-degree rose by about 2.2 collaborators; local clustering fell (ATT -0.009), meaning adopters reached beyond closed groups.
- Preferential reinforcement of strong prior ties rose (post-average +0.011), so broadening came with deeper reliance on trusted colleagues.
- Results held when never-adopters directly connected to adopters were dropped from the control group.

## How they got it

12,040 knowledge workers observed monthly; 172 always-adopters were dropped, leaving 11,868. Adoption was coded as Copilot use above the 25th percentile for two consecutive months. Collaboration and Copilot use come from separate logging systems. The main estimator is Callaway and Sant'Anna's staggered difference-in-differences with doubly robust weighting and covariates; a two-way fixed-effects IV model instruments use intensity with Copilot feature-release intensity times each worker's baseline share of Copilot-using colleagues (first-stage F 80.43). Ties count only above roughly one hour of monthly interaction.

## Limits

- One business unit of one firm; generalisation to other organisations and tools is untested.
- Short post-adoption window of [-3,+2] months; the longer-run effect is less precisely estimated and not significant for generated hours.
- Collaboration time is a volume measure; the paper does not show whether the extra time improved output.
- The validation-seeking reading is an interpretation consistent with the data, not directly measured.

## Why it matters for companies

Business cases that expect AI assistants to cut meeting load or free time from coordination are not supported here: in the first months, adopters coordinated more. If the extra contact is people checking AI output with trusted colleagues, review load lands on a small set of experienced staff, and that cost belongs in any rollout plan. Collaboration analytics already in place can track this effect without surveys.

Original: https://arxiv.org/abs/2610.32461