AI That Ends Busy Work — Hebbia CEO on “Agent Employees”

Hebia's CEO bets general AI agents will replace chatbots and vertical software

Hebia is betting that the winning AI company is a general work platform rather than a set of vertical tools or chatbots. Its model is a spreadsheet-like suite where agents do the tasks and sit in org charts beside humans. Sivulka argues that accuracy, bought with heavy inference compute and smart routing, matters more than cost, and that this is what lets Hebia sell to the largest asset managers and banks.

Takeaways

• Generalist AI beats vertical AI. Sivulka calls the belief that verticalized software is the only answer a fallacy. Hebia builds capabilities that let agents move from investing or legal work toward general reasoning, because expertise and unusual backgrounds win in a profession, not shared prompts.
• Spreadsheets, not chatbots, are the right interface. Sivulka calls chatbots the TI-84 of AI. Matrix, Hebia's flagship, is a grid where every cell is a task completed by an agent. Hebia aims to build the equivalent of a new Microsoft Office suite around it.
• Agents will become employees in org charts. Sivulka argues agents will be nodes in org charts with their own email and Slack accounts. The mix of humans and agents, from 0 to 90 percent agents, will define output, and organization design shapes products, as at AWS.
• More compute at inference drives accuracy. Hebia found early that more LLM calls at runtime improve accuracy. It productionized retrieval augmented generation in 2020, then moved to ISD, which recursively reads sub-documents. Sivulka calls context window length AI's biggest problem.
• Accuracy matters far more than model cost. Sivulka says Hebia would spend $10,000 per user a year on models without minding. Its Maximizer routing system, likened to air traffic control, handles roughly 250bn LLM calls a month, with partnerships at OpenAI, Anthropic and Amazon.
• Scale is already large and growing fast. Hebia processed about 100m pages last year and expects four to five billion this year. It is deployed at 40 to 50 percent of the largest asset managers and some tier one banks. It has raised $160m.
• AI sales needs domain experts, driven by FOMO. Sivulka says AI buying is driven more by fear of missed upside than by pain, so sellers need domain expertise and consulting skills. Hebia builds value cases, such as a $100m board-level P&L impact, though revenue upside resists a firm ROI calculation.

Sivulka's forecasts

• In ten years AI will still be deployed largely as a chatbot, because adoption is slow, like the shift from cash to credit cards.
• By 2030 prompting and managing will blur, since people already act as AI managers when they prompt chatbots.