Hebbia CEO George Sivulka argues Microsoft Excel is the right analogy for AI in finance, having reached 90 percent penetration in financial services within a year of its 1985 launch.
Sivulka says Hebbia's product is designed to say when it is not sure and return the decision to the user, rather than deliver a single answer.
Sivulka predicts agents will do most of the work inside companies, with employees becoming managers and orchestrators of AI agents before any replacement occurs.
Hebbia's CEO says AI will turn workers into agent managers
Hebbia is betting that the value in AI sits in the painful document work of finance and knowledge work, not in the models themselves. Its product breaks hard questions into steps, shows its work and hands the decision back to the human. Sivulka expects this to turn employees into managers of AI agents before it replaces them.
Takeaways
Hebbia targets document-heavy finance work, not model research. Sivulka was a Stanford PhD student in meta learning when GPT-3 arrived in June 2020. Friends in finance and private equity described pain, so he judged the bigger opportunity was document-heavy knowledge work. Hebbia began as a product studio with early retrieval-augmented generation.
Excel is the template for Hebbia's ambition. Sivulka calls Excel the most important product of the last hundred years. It reached 90 percent penetration in financial services within a year of its 1985 launch and is now used by 2bn people. He offers it as the right analogy for Hebbia.
Product shows its work and admits uncertainty. A diligence question becomes steps such as customer concentration and management strength, drawn from data rooms and public material, since a pitch deck is biased by the CEO. Matrix shows every step, surfaces disagreement and says when it is unsure, returning the decision to the user.
Employees become managers of AI agents first. Sivulka predicts agents will do most of the doing, with nearly all roles shifting to orchestrating them before replacement. He likens it to creating more managers. Hebbia was among the first to build multi-step agent workflows.
Sivulka says models already learn and outthink people. In-context learning lets models generalize and build something like a world model, which Sivulka calls learning even though the computation differs. He says foundation models are smarter than any living person, citing the MCAT and LSAT. Turchin pushed back, and Sivulka replied that brains are stochastic.
Useful AI distils signal from noise. Sivulka dislikes generative AI being used mostly to produce more noise, such as marketing copy and emails. He prefers tools that distil signal, and Hebbia takes in whole worlds of information, not just sanctioned content, to produce less biased analysis.
Hebbia's growth is fast, but adoption is early. Hebbia raised $130m in July at a $700m valuation, and revenue grew 15 times in eighteen months. Customers include Centerview Partners and Charles Bank. Sivulka says almost all customers are still experimenting with AI.
What comes next for Hebbia
Hebbia is expanding from finance into legal, working with large law firms.
Tap a name to go there · tap a line to see why they are linked
On this episode 2
Show 1
It is moving into insurance across the board, including reps and warranty insurance and auditing policies.
The current frontier is an effectively infinite context window, so models can learn how a firm works.
Sivulka expects firms to differ from each other because their custom processes differ, so AI will not run everything.