Yankai Wang, a PhD student in organizational behavior at Stanford Graduate School of Business, was reading a paper about how transformers — the architecture behind large language models (LLMs) like ChatGPT and Claude — might be used to study disease development by tracing how one condition tends to follow another. It struck him that the way people work together could be modeled in a similar fashion: as a sequence of discrete events unfolding over time. Could the same architecture that learned the grammar of human language by predicting the next word in a sentence earn the “grammar” of coordination by predicting the next event in an organization?
“Innovation comes when you borrow an idea from one realm and project it onto another,” says Amir Goldberg, the Amman Mineral Professor of Organizational Behavior and one of the faculty directors of the Computational Culture Lab. “The genius of Yankai’s idea is treating coordination as text, thinking about sequences of coordination as if they were textual representations of language and envisioning an architecture parallel to that of LLMs. But the language here, the words in the language, are not words: They are organizational events.”
When Wang came across the AI for Organizations Grand Challenge, sponsored by the Stanford Institute for Human-Centered AI and Google DeepMind, his nascent idea felt like a natural fit. Collaborating with Goldberg, Wang entered the contest, which offered a $100,000 prize to researchers studying AI tools in the workplace.
It was a winning bet: Out of more than 200 academic teams that submitted proposals, Goldberg and Wang’s was selected. In addition to the prize money, they will gain access to DeepMind’s resources so they can put their idea into practice. “DeepMind has the data,” Wang says, “and they’re on the frontier of building these kinds of models.”
Wang and Goldberg propose to use the transformer architecture to build a Large Coordination Model similar to an LLM. But instead of training it to predict the next word in a sentence, they’ll train it to predict the next event in a sequence of events, using anonymized workplace data — gleaned from calendars, emails, and shared documents, among other sources — so the model can learn the underlying patterns, or grammar, of coordination.
“You can learn a lot from predicting the next word in a sentence,” Wang 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.”
If this approach is successful, Wang imagines using the model to do things such as predict a team’s performance — anticipating where a bottleneck might occur, for example — or simulate a virtual replica of an organization so researchers can test how changes in structure or workflow might play out before they’re made in the real world.
“The hope here is that we find a transferable principle underlying how people coordinate, one that explains how they move to different companies or countries and continue to coordinate successfully,” Wang says. “We are not claiming that this will be the model of coordination, but we do think it’s going to be a useful lens to uncover some of the underlying principles of coordination as well as a very good prediction machine for a pretty broad set of scenarios, situations, and tasks.”
AI Is Changing Coordination
The project, Goldberg believes, could help shape the future of companies in the age of AI: Rather than employees being fixed in place by rigid org charts, AI could surface coordination needs that traditional bureaucracy misses. This would enable more fluid collaboration across projects and close information gaps that cause friction between coworkers.
In a related paper, coauthored with INSEAD professor of strategy Phanish Puranam, Goldberg argues that much of a manager’s work involves prediction-based tasks like planning, budgeting, and monitoring, which is exactly what AI does best. But prediction isn’t the whole job. Setting an organization’s priorities means deciding which goals are worth pursuing, and that draws on a survival-shaped motivational architecture that machines lack. (A machine’s priorities are set by its designers, for now.)
Goldberg suggests that in the future managers may become more important, since setting and assessing goals currently remains a task best done by people. That’s true even as what being a manager entails shifts from the supervisory, productivity-monitoring role of the 20th-century workplace to the coaching and collaboration-facilitating role that’s become common in knowledge work. It’s a job description that’s already transforming to include overseeing the output of AI agents, which is very different from supervising humans.
“At this moment, it’s not obvious the extent to which algorithms and AI are going to fully replace human managers,” Goldberg says. “And even if they’re capable of it, humans will be resistant to it, not simply because managers will resist their replacement, but because we won’t culturally agree to cede that responsibility to AI.”
Coordination, after all, is a distinctly human quality, one that enabled humans to succeed as a species. The modern organization was a revolutionary solution for coordinating people when tribal and kinship structures no longer sufficed. Goldberg wonders whether AI could push that evolution further still, allowing people to coordinate through a shared technological fabric rather than through formal organizations.
“The essence of coordination is changing,” Wang adds. “AI as a technology is going to change coordination with or without our project.”
Although findings remain on the horizon, Goldberg is optimistic that their work could deliver meaningful results. “Social scientists usually observe reality, and it takes a long time for their observations to then become integrated into and to impact reality. If our project is a huge success, then it might have far more immediate impact than you would normally expect.”
Read the full article by Michael McDowell / Stanford Business
