A network engineer in one of Cisco Systems’ offices has a question about a routing protocol. Instead of walking to a colleague’s desk or pinging the team chat, she pastes the question into a large language model and gets an answer in seconds. She does this dozens of times a day. She produces more than she used to. She also talks to her teammates far less, a pattern that internal data and employee surveys at the company now document with some precision, according to people familiar with the findings.
The pattern is not unique to Cisco. Companies across the technology industry are discovering that the same tools that make employees more productive can also make them more isolated. The internal research, described by people who have seen it, found that employees who use AI tools frequently produce more output but report significantly lower trust in their colleagues, and that everyday interaction between teammates has declined. Many workers now direct their questions to ChatGPT, Claude, and similar models rather than to the people sitting near them.
The dynamic has a name inside the company: technical isolation. The research suggests it carries real costs. Collaboration culture depends on repeated, informal contact, and much of what an organization knows is never written down. When an engineer asks a chatbot instead of a coworker, the question gets answered, but the knowledge does not circulate. The employee who would have learned by explaining, the newcomer who would have absorbed context by listening, both lose something that no model can replace.
The findings are awkward for Cisco in a specific way. The company is one of the industry’s most aggressive promoters of AI-powered collaboration tools, and its own products are built around the premise that software can make teams work better together. The internal data suggests that the adoption of AI, at least as currently practiced, can work against that premise. Executives have begun discussing how to design human interaction back into workflows, according to people familiar with the discussions.
The challenge is that the behavior is rational for individual employees. A chatbot answers instantly, without judgment, and without the social cost of interrupting a busy colleague. Asking a person takes time, carries the risk of appearing uninformed, and creates an obligation to reciprocate. In a workplace where output is measured and collaboration is not, the incentives all point one way. Companies that want to preserve the benefits of human interaction have to make it easier, or more valuable, than the alternative.
Some managers are experimenting with remedies: structured pairing of junior and senior engineers, dedicated time for knowledge sharing, and explicit expectations that certain questions be routed to people rather than models. The research suggests these interventions matter most for new employees, whose first months set the pattern for how they will work for years. An engineer who learns to consult a chatbot instead of a mentor is unlikely to rediscover the mentor later.
There is also a broader question about what the data means for the industry’s productivity story. If AI raises individual output while eroding the collective capabilities that make organizations effective, the net effect may be smaller than the headline gains suggest. Tacit knowledge, the kind that lives in conversations and in people’s heads, takes years to build and can dissolve quickly when the conversations stop. The surveys at Cisco offer an early measurement of a trade-off that most companies have not yet begun to quantify.
The Cisco research is part of a small but growing body of evidence on how AI changes workplaces. Studies of software teams have found similar patterns: developers who rely heavily on AI assistants write more code but engage less with code review, and the quality of review, where much of an organization’s shared knowledge is transmitted, declines as a result. The effect is easiest to measure in engineering, where output is visible, but managers say the same dynamic is likely at work in finance, marketing, and customer support, wherever workers can route their questions to a model instead of a person.
The stakes for Cisco are commercial as well as cultural. The company sells collaboration software, and its customers are asking the same questions its own employees are living through. A vendor that can demonstrate it understands the human side of AI adoption, that it has measured the effects and designed around them, has a product story that rivals cannot easily copy. The internal research, awkward as it is, may prove useful on that front, and people familiar with the company’s thinking say parts of it are being adapted into guidance for customers.
The company has not published the findings, and it has been careful not to frame them as a warning against AI adoption. The message to managers, according to people familiar with the internal discussions, is more subtle: use the tools, measure the effects, and design the workplace with the same care that goes into designing the products. The engineers who once asked each other questions are still producing; whether the teams they belong to are still learning remains an open question.
