Seventy-four percent of employees who regularly use AI now bring questions to a chatbot that they would previously have asked a colleague. That's from a survey of more than 22,000 workers across 10 regions by MyIQ, reported by FM magazine this summer. Fifty-nine percent ask colleagues for second opinions less often than they used to, and 48% say spontaneous conversations during the workday have become rarer.
On its face, that's efficiency. Nobody gets interrupted, answers arrive instantly, people feel more capable. In the same survey, 71% said AI had made them more self-sufficient.
But 46% also said they now know less about how their colleagues think and approach problems. That second number is the one worth worrying about.
The Map Nobody Drew
Researchers from BetterUp and Stanford, including Kate Niederhoffer and Jeffrey Hancock, gave this effect a name in Harvard Business Review last week: "relationslipping," the gradual weakening of working relationships as people substitute AI for one another.
Their explanation rests on an old idea from psychologist Daniel Wegner called a transactive memory system — a team's shared, mostly unspoken understanding of who knows what. It's why a good team doesn't need everyone to know everything. You know that Maria understands the client's billing quirks and that Dev remembers why the last pricing change failed, so you ask them.
That map gets built one small question at a time. According to the HBR authors, AI erodes it in two ways. When people skip conversations to get a quick answer from a chatbot, they stop updating their sense of who knows what. And when colleagues offload their own thinking to AI, their reasoning becomes harder to see — so even the map people do have grows less accurate.
The researchers found that treating AI as a teammate eroded trust among team members and their ability to coordinate. They're careful to say AI isn't inherently bad for teams. The problem is what it replaces.
A Public Version of the Same Collapse
You can watch this pattern play out in public, at scale. Stack Overflow, the question-and-answer site that served as the shared memory of the software industry for over a decade, peaked above 200,000 new questions a month in early 2014. By December 2025, it was down to 3,862, a 78% drop in a single year.
Developers didn't stop having problems. They stopped asking each other. "There is as of now no replacement public data set available," Redmonk analyst Rachel Stephens told DevClass.
Your company has its own small Stack Overflow — the hallway questions, the "quick one" messages, the five-minute screen shares. If those disappear, the knowledge doesn't vanish from individual heads. It vanishes from the network between them.
Who Pays for the Silence
The cost doesn't land evenly. In the MyIQ survey, 38% said newer workers have fewer natural chances to build relationships because routine interactions are disappearing, and 43% said they find peers harder to read professionally.
That matters most for people who learn the business by asking. A new hire who never asks the senior account manager how a difficult client works will be fine on the easy tickets — and stranded the first time the chatbot's generic answer doesn't fit.
A fair caveat: the survey's authors themselves note that their findings "do not establish that AI use directly weakens workplace relationships." People reported the change; correlation isn't proof. But the self-reports align with the HBR research and what happened on Stack Overflow, and the direction is consistent.
It also connects to the problem we covered in October. When people stop talking, they also stop catching each other's AI-generated mistakes. The same review conversations that build trust are the ones that keep unchecked output from going out the door.
What This Means
The goal isn't to discourage AI. It's to keep the questions that build a team flowing while the routine ones go to the machine. The HBR authors recommend using AI to prepare for conversations rather than replace them; here's what that looks like in practice.
- Separate fact questions from judgment questions. "What's the formula for this?" can go to a chatbot. "How would you handle this client?" should go to a person. Say it out loud as a team norm.
- Make AI the prep step, not the last step. Use it to get to a sharper question faster, then bring that question to the colleague who knows. Two minutes of AI prep plus five minutes of conversation beats either alone.
- Show your reasoning, not just your output. When you share AI-assisted work, add a line on what you asked and why you trusted the answer. It keeps your thinking visible and the team's map current.
- Build small rituals for new people. Pair every new hire with two or three named "ask me" colleagues for their first 90 days, and make asking an expectation rather than an imposition.
- Protect a little friction. A weekly 20-minute open-questions slot, a channel for half-formed problems, a standing "who's dealt with this before?" at the start of projects. Cheap, and they keep the map alive.
AI can answer almost anything. What it can't tell you is who on your team has already solved this exact problem — and that's the knowledge that makes a group of individuals into a team.
Discussion