AI and the Org: What the Researchers Are Actually Finding

Why this is worth your time: it was not punditry, and it was not a vendor deck. This was a closed-door (though recorded) gathering of the people actually building the frontier research on AI and work — presenting the newest empirical evidence, much of it still unpublished, on what AI is doing to jobs and wages.

Across two days, one message held, and it cuts against most of the breathless coverage. There is no single “AI effect.” The impact is heterogeneous across every dimension that matters — by task, by wage level, by seniority, by whether workers retrain — so any one headline number (“X% of jobs automated”) misleads. The right altitude is the skill-task level, because that is where AI actually lands. And the capability curve is steep and predictable, while the adoption curve is governed by something much slower and far more human.

The model is ready before the organization is. That gap is now the whole story.

Capability is scaling. Adoption is not keeping up.

The technical trajectory is the easy part to describe. Capabilities continue to scale with training compute and, increasingly, with dramatic gains in algorithmic efficiency. What does not scale at the same rate is the economic and organizational machinery required to put those capabilities to work. Fixed costs, workflow redesign, governance, and regulatory constraints all act as bottlenecks between what is technically feasible and what is operationally viable.

How fast is the capability itself moving? One of the cleaner ways to measure it is to ask not what a model scores on a benchmark, but how long a task a human would need for the work a model can now finish on its own. On that measure, the length of task an AI can reliably complete has been roughly doubling on a regular cadence — from minutes a few years ago to multi-hour software tasks today.

The early enterprise evidence is instructive precisely because it is modest. The wins are real but targeted. In a large-scale workplace study presented by a Microsoft researcher, a generative-AI assistant cut time spent on email by roughly 14% for the workers who used it — a clear, measurable gain. But the same study found essentially no change in time spent in meetings, and no change in time spent writing documents. The tool relieved the most structured, repetitive task and left the rest of the workday largely intact.

In regulated settings such as healthcare, deployment friction — trust, integration into existing workflows, governance — dominates the conversation far more than raw model performance.

The practical implication for leaders is a shift in where to look. The useful unit of analysis is not abstract capability but task-level economics: which specific, high-volume, high-friction tasks can be redesigned around the technology, and what organizational change is required to make that stick. First-order value may also be genuinely undermeasured — time saved and work shifted out of view rarely show up cleanly in GDP — which means traditional metrics will lag the real welfare gains.

Adoption is a social process, not a software rollout

If the bottleneck is organizational, the natural question is what actually moves people to use these tools well. The most striking adoption finding of the week: peer influence is the single most important driver of regular AI use — at least as powerful as access to the tools themselves or to formal training, and most powerful of all for the more advanced uses, like working with agents or teaching colleagues.

The implication for leaders is concrete. Buying licenses and running a training session is necessary but not sufficient. Adoption spreads through visible, trusted peers — so the levers that matter are seeding capable early users across teams, making good usage observable, and giving people room to learn from one another rather than from a manual.

Compliment, substitute, or simplifier?

The most useful reframing of the week was to stop asking whether AI "replaces jobs" and start asking how it reconfigures tasks. Labor-market impact, the researchers argued, is shaped far less by simple substitution than by task-level reconfiguration — job transformation — that interacts with a worker's existing skills. The effects are heterogeneous and distinctly non-linear: moderate exposure to AI can raise wages, while very high exposure can depress them.

That reconfiguration changes which skills get rewarded. As work is transformed, AI tends to raise the return to social skills and lower the return to analytical ones — a reversal of the pattern that defined the computer era, when analytical and technical skills commanded a growing premium.

One counterintuitive thread ran through several papers: because AI often works by simplifying tasks, the wage gains from job transformation accrue disproportionately to lower-wage workers, compressing the premium that experience and credentials used to command. The shock is, in the researchers’ word, mildly progressive.

The heterogeneity also runs by seniority. One body of research found roughly a 9% automation effect on junior roles against about a 4% augmentation effect on senior roles, with a modest ~2% productivity gain at adopting firms — juniors automated, seniors augmented. A separate analysis found an 8–9% decline in junior employment at AI-adopting firms, driven by a hiring slowdown rather than layoffs. But exposure is not destiny: workers in exposed roles who received training saw 10–20% earnings gains, which flips AI from a threat into an upgrade. Good news for broad-based earnings, and a real challenge for how organizations build entry-level talent and retrain the people they have.

It also matters where AI is actually being used today, which is far from uniform. Usage data shows AI conversations clustering heavily in computer, mathematical, and analytical work, while large swaths of the economy — jobs built on physical presence or human relationships, such as transportation, food service, and personal care — show strikingly low AI usage relative to their share of employment.

The signaling problem nobody priced in

A subtler finding deserves executive attention. AI is quietly degrading the signals managers have always relied on. A polished memo used to be costly to produce, and that cost made it informative — a classic Spence signal that told you something about the person who wrote it. When everyone can generate a polished memo, the signal collapses, and hiring becomes less meritocratic. In one study, cheapening writing led bottom-ability-quintile candidates to be hired 14% more often and top-quintile candidates 19% less often, with wages down about 5% and worker welfare down about 4%.

There is a perverse incentive lurking here, too. When AI use is visible and judged, workers may underuse the very tools that would make them most productive, for fear of looking like they leaned on the machine. The lesson is not to optimize existing evaluation practices but to rethink them — this is a period of signal instability, not steady-state efficiency.

One experiment made the point vivid by letting applicants choose whether a human or an AI screened them — and the choice itself turned out to carry information. Applicants who actively chose a human screener went on to receive offers and stay in their jobs at meaningfully higher rates than those simply assigned a human, while applicants who chose AI fared slightly worse than those assigned to it. The selection wasn't noise; it revealed something real about the candidate.

History says this takes longer than you think

The most grounding sessions were the historical ones. New technologies, even ones with an overwhelming and obvious productivity advantage, diffuse slowly because the constraint is rarely the technology. It is the cost of reorganizing everything around it. The clearest illustration is one of the most mundane machines imaginable: the elevator.

The technology was ready for thirty years before the organizations were. It is a great deal easier to automate a task than to reorganize the people around it.

The deeper point is that occupations do not vanish overnight; the mix shifts underneath us. Two centuries of labor history show entire categories of work receding while new ones expand — a reallocation, not an evaporation.

The aggregate hasn’t moved — yet

For all the firm-level churn, the topline number that everyone watches has not budged. Despite real and rising adoption, the overall unemployment rate shows no measurable AI effect: the pooled post-ChatGPT estimate comes in at about +0.002 with a standard error of 0.0019 — statistically indistinguishable from zero.

What the calm surface hides is that AI touches the day-to-day work of some roles far more than others. “Exposure” here measures the share of a role’s tasks that AI can assist with — not a probability of job loss. By that measure, roles like programming, customer service, and data entry have a high share of AI-assistable tasks, while many others have very little. Crucially, high exposure cuts both ways: the same tasks that can be automated are often the ones where AI augments the worker, and (as the evidence on training shows) exposure plus support tends to raise earnings rather than erode them. The aggregate looks quiet because these effects are concentrated and offsetting, and so far are showing up in how work is done and in hiring patterns — not in layoffs. It is exactly the kind of signal headline statistics are built to miss.

The repeated caveat from the researchers themselves: the data is still thin, the adoption is still early, and the honest answer to most long-run questions is that we do not yet know. That humility is itself a finding worth carrying back to the boardroom.

The conference’s parting message was less a forecast than a redirection. There is no single AI number to memorize; the interesting questions have migrated from the lab to the org chart. The model is ready. The work now is making the organization ready to meet it.