“Teams that share context perform better” sounds like the kind of claim that dissolves under examination. It doesn’t. It is one of the better-evidenced findings in organizational research, and the evidence has a shape worth knowing precisely, because the shape is what makes the AI argument non-obvious.
Why this isn’t a soft finding
The finding is one of the better-replicated results in organizational research. Independent meta-analyses, run a decade apart by researchers who were not setting out to confirm one another, keep landing on essentially the same conclusion, across thousands of teams. Shared context predicts team performance, and it does so on top of good teamwork behavior, not instead of it — teams that already communicate well still perform measurably better once they also share deep context.
The result holds however it is measured — whether teams rate themselves or independent observers watch them work and score what they see. The behavior-based number, the more conservative of the two, is still substantial.
The best field evidence comes from a trauma bay
The strongest study available measures a costed operational outcome rather than a survey score. Researchers followed real trauma-team handoffs at a Level 1 trauma center, tracking how much shared prior experience each pair of clinicians brought into the room against how long their patients ended up staying in the hospital. Teams with more shared experience produced measurably shorter stays — fewer days in the ICU, fewer days in the hospital overall — even after accounting for how badly hurt the patients were, how large the team was, and how experienced each clinician already was individually.
Same people, same training, same equipment. The variable was how much context the pairs already shared.
A mechanism-level detail matters for how this gets used. Shared mental models have been found not to predict performance directly; the relationship runs fully through team process. Teams that converged on shared understanding coordinated better, and the coordination produced the result. Shared context is not a substitute for good teamwork. It is what makes good teamwork possible at the pace real work demands.
What AI actually changed
“AI helps people share information” is not true by default. The evidence that matters isolates a narrower mechanism.
A large field experiment inside a real customer-support operation deployed a real-time assistant whose suggestions were mined from the successful conversations of the company’s own top performers. Issues resolved per hour rose 15%. But the average hides the finding: agents with under a month on the job improved by roughly a third, reaching in two months the performance level that had previously taken six. The most experienced agents — who had already internalized that tacit knowledge themselves — gained almost nothing.
Gains concentrated exactly where the context gap was largest, and vanished exactly where context was already abundant. That is not a generic productivity signature. It is the signature of a context transfer.
The pattern replicated in software development: a further set of field experiments found a 26% increase in completed tasks, again concentrated among the least experienced developers.
The honest caveat
40% of surveyed desk workers reported receiving AI-generated “workslop” in the previous month — output that looks complete and carries nothing behind it — at roughly two hours of rework per incident. AI can manufacture the appearance of shared context as easily as the real thing, and polished emptiness is worse than silence, because it hides the gap instead of closing it.
Something upstream of all this shifted too, quietly enough to miss. Recording a conversation and turning it into a transcript used to be either impractical or the kind of thing that made a room go quiet at the record button. Neither is true now. Meetings get captured as a matter of course, and turning a half-hour of one — or a hundred of them — into what actually mattered is a machine’s job, not an afternoon’s. Nobody reads the transcript. They get the part that was worth having.
That is a real gain, and it is an individual one — a faster way for one person to know more. It is not the finding this piece opened with. The meta-analyses measure something a faster summary does not touch: whether a team walks in holding the same understanding, not whether any one person read a better recap beforehand. The trauma-bay result held because the whole team carried the shared experience, not because one clinician showed up better briefed.
Which is the actual question AI raises for a team, not an individual: does what gets captured stay wherever it happened to land — one inbox, one person’s notes, one faster read — or does it become something the whole team can run on. The first is a productivity tool. The second is the only version that moves the number this piece has been arguing about since the opening line.
Sources
DeChurch, L. A., & Mesmer-Magnus, J. R. (2010). Journal of Applied Psychology — meta-analysis of 65 studies (231 correlations, 3,738 teams); team cognition correlated with performance at ρ = .38, raising explained variance from 11.6% (process and motivation alone) to 18.4%. Mesmer-Magnus et al. (2017) replication (k = 128 studies): ρ = .35. Zhou & Pazos (2020), transactive-memory-specific meta-analysis (56 studies, 5,249 teams): ρ = .44.
Fausett et al. (2026) on measurement dependence in transactive-memory effect sizes — self-report correlations around r = .77 versus r = .38–.39 for observer/behavioral measures.
Argote, Haan, Guo, Rosengart, Teng & Kahn (2024/2025), Organization Science — 121 trauma resuscitations; teams with above-average shared experience produced roughly 1.9 fewer ICU days and 3.3 fewer total hospital days per patient, controlling for injury severity, team size, and individual experience. Mathieu et al. (2000) on process mediation.
Brynjolfsson, Li & Raymond (2025). Generative AI at Work. Quarterly Journal of Economics — field experiment, 5,172 customer-support agents. Cui et al. (2025), three field experiments, 4,867 developers.
MIT Project NANDA, The GenAI Divide (2025). BetterUp Labs & Stanford Social Media Lab via Harvard Business Review (September 2025), 1,150 U.S. desk workers.