Hebbia.

It is 9 p.m. on a deal. Nobody is reading.

Somewhere on Greene Street in SoHo, a private equity associate is staring at a virtual data room with ten thousand documents in it. Contracts. Cap tables. Five years of board minutes. The old job was to open each one and look for the landmine. The new job is to open a browser tab, type the question, and watch a grid fill itself in. The associate is not reading anymore. Hebbia is.

This is what Hebbia does now: it sits between the world's most expensive professionals and the documents that used to eat their evenings. Its product, Matrix, looks like a spreadsheet and behaves like a research department. Documents go in as rows. Questions go across the top as columns. In between, a swarm of AI agents reads everything at once and writes the answers - each one footnoted back to the exact page it came from.

The customers are not hobbyists. Hebbia says roughly 30% of the top 50 asset managers by assets under management use Matrix, alongside investment banks, private equity firms, law firms and Fortune 100 companies. These are organizations that have always been able to hire more humans. They are choosing, instead, to hire this.

Knowledge work is mostly looking for the answer, not having it.

Here is the inconvenient truth about high-end professional work. A great deal of it is not judgment. It is retrieval. An analyst spends hours finding the relevant clause, the buried number, the one footnote that changes the model - and then about ninety seconds actually deciding what it means. The expensive part of the brain waits while the tedious part does the digging.

The first wave of AI promised to fix this and mostly didn't. Chatbots were confident and frequently wrong. They summarized three documents nicely and fell apart at three thousand. Worse, they wouldn't show their work, which in finance and law is not a quirk - it is a dealbreaker. An answer you cannot trace is an answer you cannot use.

So the real problem was never "can a model write a paragraph." It was: can a system read everything an organization knows, reason across it, and hand back something a managing director will stake a decision on? For most tools the honest answer was no. That gap is the whole reason Hebbia exists.

A 25-year-old left a Stanford PhD to bet against the chatbot.

George Sivulka was on the academic conveyor belt - a Stanford PhD student with a background in math and astrophysics, the kind of resume that ends in tenure or a quant desk. In 2020 he got off. He founded Hebbia on a thesis that, at the time, sounded slightly backwards: the future of AI at work would not be a single clever model having a conversation. It would be many models, coordinated, each doing a small piece of a large reading job.

The name itself is a wink at the idea. "Hebbia" gestures at Hebbian learning - the neuroscience principle that neurons which fire together wire together. The company that wants to network thousands of AI agents into one answer named itself after the way brains build connections. Cute, and on-message.

Investors who have heard a thousand AI pitches found this one persuasive. Peter Thiel and Floodgate backed it early. Index Ventures led a $30M Series A in 2022. And in 2024, Andreessen Horowitz led a $130M Series B, with Index, GV, Thiel again, plus Eric Schmidt and Jerry Yang putting in personal money. The kicker that made the round unusual: Hebbia was already profitable, on about $13M of revenue, before taking the check.

Four lines that took most enterprise software a decade. The compression is the point.

Matrix: a spreadsheet that does the homework.

Open Matrix and you see a grid, which is a comforting thing to show a person who lives in Excel. The trick is what the grid is made of. Each row is a document - a 200-page credit agreement, a deck, an email chain, a scanned image. Each column is a question you would normally ask a junior analyst. Each cell is an answer, written by AI, with a citation you can click to land on the precise source.

The agent swarm

Under the surface, Matrix doesn't hand your question to one model and hope. It breaks the question into parts and dispatches a swarm of agents - drawing on frontier models from OpenAI and Anthropic - to work the pieces in parallel. One reads the indenture, another reconciles the numbers, another checks the footnotes. The results are assembled into a single, sourced answer. It is less "ask a genie" and more "spin up a team for ten seconds."

The numbers people actually repeat.

Demos are easy. The reason Hebbia gets taken seriously is that the customers are conservative institutions, and they keep coming back. Analyses that used to take two to three hours are reported to take two to three minutes. Diligence that occupied a team of analysts for weeks now gets a thorough first pass in an afternoon. Publicly referenced logos include the kinds of firms - KKR, Morgan Stanley, MetLife, FactSet - that do not put their name next to software lightly.

And the revenue line is the part investors whisper about: roughly $13M, profitably, before the Series B. In an era where AI companies treat losses as a personality trait, that detail did a lot of the convincing.

Put the world's knowledge to work.

Strip away the funding headlines and Hebbia's goal is almost plain: let people get a complete, trustworthy answer from everything their organization knows, instead of hunting for it by hand. The company talks about ending "busy work" and building "agent employees" - software that does the reading, gathering and cross-referencing so that humans are left with the part that actually requires a human.

It is a tidy ambition with a sharp edge. If the boring 80% of knowledge work can be automated reliably and with citations, the shape of a lot of white-collar jobs changes. Hebbia is, politely, betting that the people who own those jobs would rather direct the work than do it. So far, the asset managers seem to agree.

The data room never gets smaller. The night just gets shorter.

Deals are not getting simpler. Filings are not getting shorter. Every year there is more to read and the same number of hours to read it in. That trend has only ever had two answers: hire more people, or read less carefully. Hebbia is selling a third one. If it works at scale, the competitive question for a fund or a firm stops being "who has the most analysts" and becomes "who asked the better questions."

There are real caveats. Frontier models are expensive and improving under everyone's feet at once. Rivals like Glean, AlphaSense and Harvey want the same desks. Trust, in these industries, is earned slowly and lost in a single wrong citation. None of that is settled.

So return to that associate at 9 p.m. on Greene Street. The ten thousand documents are still there - they always will be. What's changed is who opens them. The grid fills in. The citations check out. The questions that mattered get asked out loud instead of buried in a folder. Hebbia didn't make the data room smaller. It made the reading disappear. And on a deal night, that is the whole game.