Cómo OffDeal usa Claude para cerrar 91 millones de dólares en operaciones de M&A con un equipo de cuatro banqueros

OffDeal is an AI-native investment bank built on the belief that the middle market — founder-led businesses with $1–20 million in annual profit — deserves the same quality of M&A advisory as large corporations, but has historically been too expensive to serve well. The firm pairs bankers with software engineers and uses AI to do the work that would otherwise require a large team.

"I think we’re the only investment bank in the world that has more engineers than bankers," said CEO Ori Eldarov. With five engineers and four bankers, OffDeal runs every stage of a deal through software: originating clients, sourcing buyers, building seller and buyer lists, writing confidential investment memoranda, and running diligence.

The early architecture was a collection of 12 independent agentic workflows, each designed for a specific task. Every API call was smart in isolation, but the system was brittle. When a standard operating procedure changed, the corresponding workflow broke. The deeper problem was context: a single deal can involve more than a hundred million tokens of information, and the models couldn’t hold enough of it to handle problems outside their predefined scope.

Migrating to the Claude Agent SDK and redesigning around MCP principles changed the architecture entirely. Twelve workflows collapsed into Archie, a single general-purpose agent that holds the full context of a deal and can move fluidly between tasks. Eval accuracy — measured against an internal benchmark — jumped from 25% to 85%. Failed API calls from context overflow dropped to zero overnight. The buyer sourcing agent now runs autonomously for up to four hours, researching potential acquirers across 10 different sourcing methods without human intervention.

In its first year, OffDeal closed 8 deals totalling $91 million in transaction value, with each banker managing 5–8 concurrent deals and meeting 2–3 new potential clients per day supported by AI-generated preparation materials. The firm is proving that a small, software-augmented team can operate at the output of a much larger bank.