Seventy-four percent of regular AI users now put a question to a chatbot that they would previously have put to a colleague. Fifty-nine percent ask coworkers for second opinions less often than before. Forty-six percent say they know less about how their colleagues think. That is from a survey of more than 22,000 employees across ten regions, fielded in June and July 2026 by the cognitive assessment platform MyIQ (FM Magazine, 2026).
Asking AI instead of colleagues is now the default reflex, and here is the part that should concern a Head of Operations: none of that shows up on your dashboard. Meetings still happen. Message volume is flat or up. Scheduled collaboration is untouched. What is disappearing is the unscheduled consult โ and that is the layer your onboarding and your promotion calibration were quietly running on.
The metric you are watching cannot detect this
Most mid-market collaboration measurement is volumetric: meeting hours, Slack or Teams message counts, cross-team channel activity, calendar density. Those instruments capture scheduled and semi-scheduled interaction because that is what leaves a structured trace.
The behavior described in the MyIQ data leaves almost no trace. It is the thirty-second turn to the desk next to you. The "quick sanity check before I send this." The "how does this process actually work, because the doc is wrong." Those exchanges were never on a calendar and rarely in a channel. When they migrate to a chat window, the volumetric dashboard registers nothing โ or, worse, registers improvement, because the same headcount is now producing more visible output with fewer interruptions.
This is a measurement failure with a specific shape: your collaboration metrics have become insensitive precisely where the change is happening. Flat numbers are no longer evidence of a healthy team. They are evidence that you are measuring the part that did not move.
What changes when employees ask AI instead of colleagues
The survey figures cluster into a coherent pattern rather than a scatter of complaints (MyIQ / Cooperative Agency, 2026):
- 74% direct questions to a chatbot they would previously have asked a colleague
- 59% ask colleagues for second opinions less often
- 48% report fewer spontaneous conversations during the workday
- 46% know less about how colleagues think and approach problems
- 43% find peers harder to read professionally
- 38% say newer workers have fewer natural opportunities to build relationships
Read the sequence as a causal chain rather than six separate statistics. Fewer questions asked (74%) produces fewer spontaneous conversations (48%), which produces less exposure to how colleagues reason (46%), which produces reduced ability to read peers professionally (43%). The last figure is not a soft-culture complaint. It is a degradation in the raw material that both apprenticeship and promotion judgment are made of.
The same respondents also report real gains: 71% feel more self-sufficient and 62% feel more confident deciding independently. That is not a rounding error, and it is why this will not present as a problem in any engagement survey. Individuals experience the change as competence. The cost lands on the system, not the person โ which is exactly the class of cost an operations function exists to catch.
One honest caveat before you act on this
This is a vendor survey from a cognitive assessment company, self-reported, released through PR, with no published methodology beyond sample size, geography, and field window. The release itself states that the findings do not establish that AI use directly weakens workplace relationships โ only that employees associate greater AI reliance with a change in how often they consult and informally learn from one another.
So size the effect from this data and carry the causal weight on the organizational psychology literature, which is considerably older and considerably better identified. That is the next section.
The system that is degrading already has a name
What thins when people stop asking each other things is the team's transactive memory system โ the shared directory of who knows what, first formalized by Wegner and developed over three decades of team research. It is maintained by use. Every time someone asks a colleague a question, two things happen: the asker gets an answer, and both parties update the directory of expertise. Remove the asking and you keep the individual knowledge but lose the index.
That index is not decorative. A meta-analysis of 76 empirical studies covering 6,869 sampling units found transactive memory systems reliably associated with team performance, with effect strength varying by context rather than the association itself being in doubt (Bachrach et al., 2019). Teams that know where expertise lives outperform teams that do not, holding individual capability constant.
An AI assistant is an excellent substitute for the answer. It is not a substitute for the index. When your senior operations analyst answers a question, the organization learns something about your senior operations analyst. When Claude or ChatGPT answers it, the organization learns nothing at all.
The counter-evidence, and what it actually shows
There is credible research pointing the other way, and it deserves a fair hearing. Gensler's 2026 Global Workplace Survey of more than 16,400 office workers across 16 countries found that the 30% who qualify as AI power users spend less time working alone (37% of the workweek versus 42% for late adopters), more time learning (12% versus 8%), and more time socializing (11% versus 9%), with stronger reported team relationships (Gensler, 2026).
Both findings can be true, and I think they are. AI absorbs the low-stakes informational question โ the kind that carries no relationship value anyway โ and frees time that heavy users redirect into learning and social contact. The aggregate time budget improves. What the Gensler data does not tell you is who benefits from that redistribution.
The MyIQ figure that answers it is the 38%: newer workers have fewer natural opportunities to build relationships. Reallocated social time accrues to people who already have a network to spend it in. A new hire in month two has no such network, and the interactions that would have built one were exactly the low-stakes informational questions now going to a chatbot. Workday's 2026 research is consistent with this asymmetry โ Gen Z employees were 12 times more likely than Gen X to report feeling completely disconnected from colleagues, with more than one in five saying AI tools had made their peer relationships worse (Workday, 2026).
The net effect is not "AI isolates everyone." It is narrower and more actionable: the benefit is concentrated among the already-embedded, and the cost is concentrated among the newly-arrived. That is a distributional problem, and distributional problems are solved by design, not by sentiment.
Onboarding: expertise location has to become an artifact
If ambient consultation was your apprenticeship mechanism, it is now partially unfunded and nobody defunded it deliberately.
For a 50โ500 FTE operation, the fix is not a mentoring program with a launch deck. It is making expertise location explicit, because it can no longer be emergent:
- Publish the directory the asking used to build. A maintained map of who owns which process, system, exception path, and customer segment โ with names, not team boxes. This is a one-page artifact per function, reviewed quarterly. Most mid-market orgs have this information distributed across the heads of four people and written down nowhere.
- Make the first ninety days require named humans. Structure ramp around deliverables that cannot be completed without consulting three specific colleagues. Not "meet the team" โ a task with a dependency.
- Instrument the right thing. Stop reading message volume as connection. Ask new hires directly, at day 30 and day 90: who would you go to for X? Track how many names they can produce and how fast that number grows. That is a measurable proxy for transactive memory, and it costs one survey field.
The cheapest version of this is a fifteen-minute weekly session where a new hire has to bring one question they resolved with AI and re-ask it to a person. The point is not the answer. The point is the introduction.
Promotion calibration: your peer signal is losing validity
The 43% figure โ colleagues harder to read professionally โ is a measurement problem wearing a culture-problem costume.
Mid-market promotion decisions lean heavily on peer impression: what the calibration room believes about someone's judgment, based on observing them reason in unstructured moments. Those moments are thinning. The input is degrading while the process that consumes it stays exactly the same, which means calibration confidence is holding steady on a weaker signal. That is the precise condition under which bias expands, because when observation thins, familiarity and visibility fill the gap.
The remedy is well-evidenced and unglamorous. Structured assessment โ defined competencies, consistent evidence requirements, the same questions asked of every candidate โ has long shown materially better predictive validity than unstructured judgment. Sackett and colleagues' 2022 re-analysis of personnel selection validity, which corrected decades of systematic overcorrection in the meta-analytic literature and generally lowered published estimates, still placed structured interviews near the top of the usable predictors at an operational validity around .42 (Sackett et al., 2022). Unstructured impression does not compete with that even when observation conditions are good โ and yours are getting worse.
Concretely: require written evidence of judgment for every promotion case โ a decision the person made, the tradeoff they priced, the outcome. If nobody can produce that evidence for a candidate, that is data about your observation conditions, not about the candidate.
The decision for this quarter
You are not going to reverse the 74%. Asking AI instead of colleagues is faster, it is judgment-free, and it makes people feel more capable โ 71% of them say so. The behavior is rational at the individual level and will not respond to a policy asking people to talk more.
What you can decide this quarter is whether the two systems that were silently running on informal consultation get an explicit replacement. Both are cheap. Publish an expertise map per function. Add one written-evidence requirement to promotion cases. Neither needs budget approval, and both stop working the moment you assume the ambient version is still running.
Your collaboration dashboard is going to keep looking fine. That is the whole problem.
