Foaster.ai wants to redraw the map of how your company actually works

There is a document that does not exist at almost any company: the true map of how work moves through it. Not the org chart on the wall, but the real thing - who hands what to whom, where a request sits for three days waiting on one person, which spreadsheet secretly runs a department. Foaster.ai is a two-person San Francisco startup built on a single bet: that AI agents can go find that map, and that finding it is worth a large slice of a $500 billion market.

The company calls itself, without much hedging, the AI-native McKinsey. That is a big label for a company with two employees and a $130,000 seed check. But the founders, Raphael Dabadie and Alexandre Combes, did not arrive at consulting through consulting. They arrived through a game about lying.

Teaching machines to bluff

In the year before Foaster existed, Dabadie and Combes worked as AI consultants and, on the side, built benchmarks to test how large language models behave. Their sharpest idea was to make AI models play Werewolf - the party game where hidden werewolves lie their way through a village of players trying to deduce who among them is the killer. Most AI tests measure facts and math. Werewolf measures something slipperier: whether a model can bluff, coordinate, deceive on purpose, and hold its story together under pressure.

The results traveled. Across 210 games, six models took roles - two werewolves, four villagers with special powers like the seer and the witch - and fought through day debates and hidden night phases. GPT-5 finished on top with a 1,492 Elo rating, winning nearly all of its games, and sustaining a 93 percent manipulation rate on the nights it played a werewolf.

Greg Brockman shared it. So did Hugging Face's Clement Delangue and Thomas Wolf. A second study, on how leading models' political preferences line up with real election results across eight countries, got picked up by Elon Musk. Between the two, Foaster's research reached more than 12 million views before the founders had a product to sell.

"Most companies confuse AI adoption with AI transformation."Foaster.ai, YC Launch

What the benchmarks bought them was attention and, more usefully, a point of view. Spend a year watching how models actually reason and you start to see clearly what they can and cannot replace. Dabadie and Combes concluded that a great deal of consulting - the interviewing, the synthesis, the pattern-matching across a hundred conversations - was work an AI could do faster and at larger scale. So they built one.

The operational graph

Foaster's core move is to separate two things companies constantly blur. Adoption is buying tools - a chatbot here, a copilot license there. Transformation is redesigning the work itself around what AI can now do. The founders argue that most enterprises are drowning in the first: scattered pilots, tools bought without context, plans that never touched the reality of anyone's day. You cannot fix a workflow you have never actually mapped.

So Foaster maps it. AI agents run 30-to-45-minute interviews with employees across an organization, capturing the workflows, tools, handoffs, bottlenecks, and repetitive tasks that usually live only in people's heads. From those conversations the system rebuilds what the company calls an operational graph: a structured picture of how the business actually runs. Then human experts step in to prioritize - weighing impact, feasibility, and strategy - and decide where AI should be deployed first.

The pitch on time is the part that makes executives lean in. Foaster says it can deliver in roughly two weeks what a traditional consulting team might take months to produce: a company-wide read on how the place operates, where it leaks time, and what to automate next. It sells this as an initial AI Transformation Assessment, then as ongoing monthly support that keeps the map current, and, for larger clients, as enterprise deployment with forward-deployed engineers.

Where it fits

The incumbents are obvious: McKinsey, BCG, Bain, Deloitte, Accenture and their fast-growing AI-transformation practices. Those firms sell judgment, relationships, and the credibility of a name - and they staff engagements with expensive humans. Foaster's wager is that the labor-heavy middle of that work, the mapping and synthesis, is precisely what agents do well, and that keeping only a small number of experts for the judgment calls is both cheaper and faster.

"The $500B consulting market will eventually be run by our AI agents and a small number of expert humans."Foaster.ai

It is a crowded moment - internal AI enablement teams, workflow-mapping startups, and every big firm racing to slap "AI" on a service line. Foaster's edge, for now, is credibility earned in public. Few consulting startups can say their research was reposted by the people who build the frontier models. And the founders' track records are unusually operational for a company this young.

Two unusual resumes

Raphael Dabadie, the CEO, has been building audiences since he was a child. He started a tennis fan community at age 11 that grew past 400,000 people, and at 18 he joined Rafael Nadal's team as an agent. Alexandre Combes, the CTO, built Repondia, a system that handled more than 60,000 customer calls for French restaurants using AI agents. Between them they have the two halves this company needs: someone who knows how to make people pay attention, and someone who has already shipped agents into thousands of live, messy human conversations.

That second detail matters more than it sounds. Interviewing employees is not a clean API call. People ramble, contradict themselves, protect their turf, and describe processes they have never articulated out loud. Combes has already run agents through tens of thousands of exactly that kind of exchange. Foaster is, in a sense, that experience pointed at a new problem.

Foaster joined Y Combinator's Spring 2026 batch, with Nicolas Dessaigne as its primary partner, and published its Werewolf Benchmark as an open Hugging Face Space. The company is early - no roster of named customers, no disclosed revenue - and honest about its ask: warm introductions to enterprise and mid-market companies serious about AI transformation, and to the operators inside them who own enablement, operations, and process improvement.

The interesting tension in the whole plan is the one Foaster leaves in on purpose. It could have promised to automate consulting entirely; that would have been a cleaner story. Instead it keeps humans on the judgment - the part where you look at a map full of opportunities and decide which ones are actually worth doing. The agents find the terrain. People still choose the route.