The hidden grammar of teamwork: Can AI learn how organizations work?

Nine hospitals changed how doctors hand over patients at the end of a shift, and medical errors fell 23%. What changed was how information moved between people. Two Stanford researchers now propose training an AI on workplace events, the way language models train on words, to learn the patterns behind results like that one. So far it is a proposal, and it raises hard questions about prediction and privacy.

At the end of every shift in a hospital, a doctor in training hands patients over to the next doctor. Things get lost in that conversation. Amy Starmer and her colleagues tested a fix at nine hospitals, called I-PASS. It gave doctors a memory aid for what to say, along with training and senior doctors who watched handoffs and gave feedback. Across 10,740 patient admissions, medical errors fell from 24.5 to 18.8 per 100 admissions, a drop of 23%. Preventable harm to patients fell 30%.

I-PASS gives each handoff five parts, from how sick the patient is to a read-back by the person receiving it.

Per 100 admissionsBeforeAfter

Medical errors24.518.8

Preventable harm4.73.3

Harm nobody could prevent3.02.8

Figure 1. Press the buttons. The numbers are from Starmer et al. (2014), across 10,740 admissions at nine hospitals, on one scale. The drawing is illustrative, and the pieces it drops are invented.

The program taught doctors how to hand over information, and it did not teach them new medicine. The study compared hospitals before and after the change without randomly assigning them, so something else could have helped. One detail points toward the handoffs. Harm that nobody could have prevented stayed flat, at 3.0 and 2.8 events per 100 admissions.

Coordination

Coordination is the order in which work happens and how information reaches the person who can act on it. Two teams with identical skills can coordinate very differently. An organization chart shows who reports to whom, and it leaves most of this out.

The same people can get very different results

Group projects work the same way. Suppose four students each have a different strength. If the programmer builds the model before the statistician checks the data, the model gets built twice. If the writer drafts before anyone agrees on the question, the draft gets thrown away.

Same four students, same five jobs. Put them in an order that wastes the least time.

Build the model
Write the report
Agree on the question
Check the data
Review it together

10 days of work12 days redone

In this order, 12 days of work get redone. Nobody got any smarter when you fix it.

Figure 2. Press the arrows. The jobs, the days and the rules are invented to show the idea. Work gets redone when a job comes before the job it depends on.

Researchers have measured this in groups. In 2010, Anita Woolley and her colleagues gave 699 people in small groups a range of tasks. Groups that did well on one task tended to do well on the others, so the authors called this shared factor collective intelligence. It was not strongly related to the average or the highest IQ in the group. It was related to how well members read each other and how evenly they took turns talking.

Which of these do you think predicted a group’s performance most strongly when researchers pooled the evidence?

Which predicted a group’s collective intelligence most strongly?

Collaboration process came first, then individual skill, then who was in the group.

Figure 3. Make a guess. The order is from Riedl et al. (2021). The heights of the podium are not to scale, and they show only the order.

That pooled analysis, by Christoph Riedl and his colleagues, covered 22 studies and 5,279 people. The idea still has critics. Marcus Credé and Jacob Howardson reanalysed the early data and argued that the evidence for a single group factor is weak, and Woolley’s team disputes that. The narrower point holds up across this work. Skilled members alone do not settle how well a group performs, which I also touched on in When answers are cheap, what makes a person valuable?

A language model for the workday

Language models learn by predicting. Given “The student submitted the,” a model estimates which piece of text comes next. It checks the real answer and adjusts itself. Do that across enough text and it picks up grammar that nobody wrote down for it. Those pieces of text are called tokens, and researchers write the task as P(next token | previous tokens), the probability of the next token given everything before it.

In 2026, Yankai Wang, a PhD student at Stanford’s business school, and his professor Amir Goldberg proposed a Large Coordination Model that changes what counts as a token. Wang got the idea from a paper that used the same technology to trace how one disease tends to follow another. “You can learn a lot from predicting the next word in a sentence,” he says. “Here we are just replacing ‘sentence’ with ‘a stream of work events,’ and replacing the next word with the event that is going to happen next.” The events would come from anonymized workplace records such as calendars and emails.

Question agreedData shared

Probability of the next event, given the last one

Analysis done75%

Draft revised25%

Press any bar to make that the next event, and the model predicts again.

Figure 4. This is a real, very small model, running in your browser. It counts which event follows which in 12 short logs and 10 short sentences that I invented, so it looks back only one step and has never seen a real workplace. A Large Coordination Model would use a transformer, the kind of neural network behind language models, and far more history.

The proposal beat more than 200 teams to win the AI for Organizations Grand Challenge from Stanford and Google DeepMind. The $100,000 prize includes running the study inside DeepMind’s own offices, with computing time and engineers. “DeepMind has the data,” Wang says. If it works, the model might warn a team about a bottleneck before it forms. Wang hopes it will find principles that hold when people move between companies and countries. The bet is that the patterns of good teamwork are regular enough to learn from data, the way the grammar of a language is.

A proposal, so far

As of the August 2026 article, there are no published results. Nobody has shown that this model predicts team performance. “We are not claiming that this will be the model of coordination,” Wang says. Everything below about what it might do is a hope that still has to be tested.

Workplace traces have already shown what an org chart cannot

Studying work through its digital traces is an older idea. A field called process mining has used event logs for years to find where cases get stuck and to predict how long they will take. The new step is learning across many kinds of work at once.

One large test came by accident. When Microsoft sent everyone home in 2020, Longqi Yang and his colleagues analysed the digital traces of 61,182 US employees, including their emails and calendars. Some of those employees had already been remote, which let the researchers separate the effect of a firm-wide move from the effects of the pandemic. Firm-wide remote work made the collaboration network more static and siloed, meaning people stuck to the same colleagues and had fewer links between groups.

Same people, and a different shape. The two gold people carry news between groups, and the org chart does not show that job.

Figure 5. Press the buttons. The ten people and their links are invented. The real result is from Yang et al. (2022), whose network of 61,182 people became more siloed, with fewer bridges.

Predicting the next event is the easy part

Suppose a model learns that teams with more meetings miss more deadlines. It would predict well. It would also give bad advice if the meetings were not the cause. Hard projects can need more meetings and also run late.

Across all 60 teams, more meetings go with more days late. A model that only predicts would stop here.

Figure 6. Press the buttons in order. These 60 teams are simulated by me, and I built meetings to help inside each group, so the data have the trap in them. Each line is a least-squares fit, the straight line closest to the dots. Real data do not come with a button that reveals the hidden cause.

Prediction also has limits of its own. In one large test, 160 teams built models to predict six life outcomes for children from a very rich long-term study of families. The best predictions were not very accurate, and they were only slightly better than a simple benchmark. Human behaviour may contain less predictable pattern than the richness of the data suggests.

Wang also imagines a virtual replica of an organization, often called a digital twin, where researchers could try a change before making it. A twin only earns trust by predicting what happens after real changes, and only experiments can show that. The small one below follows rules I wrote, so it can only tell you what those rules imply.

One approver finishes 10 tasks in 40 days, and the queue keeps growing.

Figure 7. Press a button to run 40 simulated days. The rules and rates are mine. A real digital twin would have to earn trust by predicting what happens after real changes, and this one has never been tested against anything.

A forecast leaves the goal open

Much of a manager’s work is prediction, such as planning budgets and spotting trouble early. Goldberg and Phanish Puranam, a professor at INSEAD, argue in a 2026 paper that this is exactly what AI does best, and that setting priorities is a different job. A forecast can tell you which project is most likely to publish a paper by June. It cannot tell you whether a quick paper is what your lab should want.

The forecast for each project stays fixed. What do you want the lab to care about?

A paper by June

Long-term science

Public need

A quick, safe paperForecast: 85% chance of a paper by June0

A hard open questionForecast: 20% chance of a paper by June0

An urgent public problemForecast: 45% chance of a paper by June0

Care only about papers, and the quick paper wins. Move the sliders, and the winner changes while the forecast stays put.

Figure 8. Drag the sliders. The three projects and all their numbers are invented. The idea that forecasting and choosing goals are separate jobs is from Goldberg and Puranam (2026).

People may also refuse to hand the job over. “Humans will be resistant to it,” Goldberg says, “not simply because managers will resist their replacement, but because we won’t culturally agree to cede that responsibility to AI.” Their argument covers what managers do today. A machine’s priorities are set by its designers, for now.

The same data can help a team or score a worker

In 2020, Microsoft added a feature to its Productivity Score that showed employers each worker’s name and actions over 28 days. After a week of public debate, the company removed names and kept only organization-wide numbers. Measured workers also respond to the measure. In 2024, Wells Fargo discharged more than a dozen employees after reviewing allegations of “simulation of keyboard activity creating impression of active work.”

Worker Aspends the afternoon helping a colleague think

Worker Bsends many updates, and owns a mouse jiggler

Score keyboard activity, and Worker B looks busier than Worker A. Now try the other two scores.

Figure 9. Press the buttons. Both workers and their numbers are invented. A mouse jiggler is a cheap device that keeps a computer looking active, and coverage of the Wells Fargo case says staff were accused of faking activity. Economists call this Goodhart’s law. Once a measure becomes a target, it stops being a good measure.

Anonymizing records helps, though it protects less than it sounds like it should. Yves-Alexandre de Montjoye and his colleagues studied phone location data on 1.5 million people and found that four points were enough to identify 95% of them. Patterns of work may be distinctive too. Few people email the finance chief at 3 a.m.

Names are gone from these 5,000 work records. What might a colleague already know about the person you are looking for?

With all four clues, 5,000 anonymous records come down to one person.

Figure 10. Press the clues. The workers and the clues are invented, each clue cuts the pool by an invented share, and each dot stands for 50 people until few are left. The measured result is from de Montjoye et al. (2013), on phone location data.

Where the limits lie

The phone study used location data, so its 95% does not apply directly to calendars and emails. A rare pattern of work could still point to one person. A model built to study teams can later be used to rank individuals, and nothing in the data itself prevents that. The decision belongs to whoever runs the system and whoever holds them to account.

How far the idea has come

The idea has cleared the first two rungs of a long ladder, and the study is still ahead.

Figure 11. The first two rungs are from the Stanford article, and the other four are my reading of what would have to come next. The researchers may run them in a different order.

The Stanford article says that “findings remain on the horizon.” Meanwhile, the nine hospitals saw their 23% drop with no model at all. They used a five-letter memory aid. They added training and had senior doctors watch handoffs.