Network pathways to belonging: How AI reveals the architecture of employee connections

Featured Researcher

Sameer Srivastava

Professor, Management of Organizations

Michael Blanding

Photo of diverse group of professionals chatting and networking at a business event in a modern office building — Photo: Jacob Lund/Adobe Stock

The idea that employees define themselves through their companies is becoming less common, a trend that’s been building for decades and accelerated during the pandemic. “My father spent 40 years at IBM,” says Berkeley Haas Professor Sameer Srivastava. “That was a time when people very strongly identified with a company—they would talk about themselves as ‘IBMers.’ Especially after the pandemic, a lot of those bonds really started to fray.”

This change could have significant repercussions: Prior research has tied strong organizational identification to lower turnover, higher engagement, and improved employee performance for employers, as well as greater work satisfaction and lower stress and anxiety for workers. But traditional surveys aren’t built to track the way identification ebbs and flows over time. “Some weeks you might feel really connected, and some weeks you might feel more distant,” Srivastava says.

In a new study published in the American Journal of Sociology, Srivastava and co-authors—incoming Haas PhD student Sarayu Anshuman and Lara Yang and Amir Goldberg of Stanford—developed a language-based machine-learning model to detect company identification in employee email. The results offer insights that may help companies build greater employee alignment.

Identification through networks

The researchers’ model analyzed the contextual relationship between the pronouns “I” and “we.” The closer the words appeared in what’s known as a “semantic vector space”—essentially, the degree of overlap in meaning that people attached to these terms—the more identified the employee was presumed to be.

The researchers measured the fluctuations in worker identification at three companies: a technology firm, a design-services company, and an IT staffing and outsourcing firm. They anonymized data from 14.7 million email messages to create results that were person-, company-, and time-specific. At the same time, the researchers mapped the metadata of internal communications (i.e., who sent messages to whom, and when) to analyze employees’ networks—specifically the density of connections to fellow employes, referred to as cohesion, and the extent to which those ties reach globally across the company, referred to as range.

Cohesion has long been associated with emotional support and camaraderie, while range brings access to information and influence. Traditionally, network theorists have focused on the tradeoffs associated with these network structures. But Srivastava and colleagues found that both have a positive relationship with organizational identification.

“These two network structures seem to work together,” Srivastava says. Cohesion helps employees form strong bonds with their immediate colleagues, while range offers a big-picture view of the organization.

That’s because identification, the study suggests, isn’t just about being ensconced in a mutually supportive community—it’s about feeling connected to something larger. “What seems to shape identification is having a supportive, cohesive group of people that you can connect with, while at the same time ensuring that you have sufficient bridging ties across the organization,” he explains.

Improving connections

For employees, that bridging often takes intentional effort. “The cohesion part happens more naturally,” Srivastava says. “The harder part is global reach, because that often requires reaching beyond your typical set of colleagues.”

He encourages employees to invest in “cross-cutting ties” that link them to other units, such as by joining cross-functional projects, volunteering at recruiting events, or mentoring employees outside their department. “Those kinds of activities can be powerful for creating identification with the organization as a whole.”

For organizations, the implications are trickier. While the machine learning approach could theoretically track individual employees’ identification levels in real time, Srivastava is cautious about its potential misuse and the unintended consequences that could ensue. “It’s not hard to imagine firms starting to use such tools to inform decisions about whom to promote and give opportunities to. Without careful safeguards and protocols, as well as transparency of communication, a certain creepiness factor can quickly set in.”

Instead, he envisions using such models to calibrate traditional survey data. By combining occasional surveys with digital trace data, a company can monitor changes in employee sentiment with greater nuance, while at the same time reducing survey fatigue. “This opens up a possibility for companies to do surveys much less often, and in more targeted ways,” he says. “It allows you to get at variation in psychological beliefs and attitudes without having to ask people all the time.”

Ultimately, the research supports investing in organizational structures that both foster belonging within groups and encourage meaningful cross-boundary connections. In other work, Srivastava has tested interventions like reshaping professional development programs to bring together people from different internal communities—resulting in measurable increases in employees’ sense of belonging.

“These kinds of interventions can take siloed organizations and make them more cohesive,” he says. That, in turn, could make people feel more connected to a company’s overall mission—boosting motivation, sense of purpose, and performance.