Most leaders track AI adoption the way they used to track software licences. Seats in use. Prompts per week. Tickets closed with copilot help. The dashboards look tidy and they miss the whole point.

Every question your team asks the model is a question they did not ask a colleague. That is the number nobody is measuring, and it is the one that will matter twelve months from now.

The velocity you can't feel

A VP of engineering at a Series C B2B SaaS company, on a call last month. Roughly seventy engineers under her. She opened the session proud of the numbers: PR throughput up, mean time to merge down, on-call incidents flat. Then, halfway through, she said something that stopped me.

"We haven't done a real design review in about six weeks."

She could not say when the last one was. She could not name who noticed it had stopped. The pattern was gentler than a decision. Standups shorter. Pair-programming rarer. The Slack channel where seniors used to argue about tradeoffs quieter every week. The copilot had absorbed the small questions, and the big conversations went with them.

Nobody made a call about any of this. The dashboard could not see it. Every headline number was up.

What the research says

Stanford, Michigan, Oxford and CMU published a longitudinal study on September 7. They followed 1,182 CharacterAI users at baseline and 439 of them a full year later, mean 362.5 days apart. Consumer companion chatbots, not enterprise copilots. But the mechanism they measured is one senior leaders should recognise.

Sustained chatbot use predicted lower well-being. Then they tested why.

Results further support the social displacement pathway, indicating that these links were mainly explained by lower in-person social interaction.

The direct effect of chatbot engagement on well-being disappeared once in-person time entered the model (β=-0.06, p=0.074). The indirect effect through displaced human contact held. The exposure carried no measurable weight in isolation. The quiet replacement of face-to-face contact was where the cost sat.

Professional AI is a different product from a companion chatbot, but it has the same shape: always available, memory-carrying, personality-tuned, faster than a colleague. So the displacement question is the right one to ask, even though the paper did not measure it at work.

The practice is small and awkward. Measure in-person team time separately from AI interaction time. Count design reviews per month. When the ratio drifts, catch it in month two, not month twelve.

The bill for the drift arrives a year later. Your dashboard will still be green the week it lands.