Automating a search fund's deal flow, from first email to closing

MBC backs entrepreneurs (“searchers”) who acquire and run small and mid-sized companies. Its small team of investment analysts takes each deal from a teaser in the inbox to a signed shareholder agreement, through research, scoring, searcher calls, diligence and investment committee review. Much of that time went on work that needs no judgement: triaging email, filing call notes, assembling research and drafting documents.

MBC wanted to grow its portfolio without growing its team. An AI platform had been prototyped, with 122 database tables and a set of agents, but it had never run on live data, and no one had defined which steps of the deal process it should take over.

1. Map the deal workflow. I ran working sessions with the deal and searcher teams to map the M&A process end to end, in eight stages: intake, research and scorecard, searcher engagement, roundtable, deep diligence, investment committee preparation, committee decision, and closing. For every step, the map sets out what AI does, what a person does, and where a person must approve.

2. Put AI on the repetitive steps. In the target design I set out, agents:

• classify inbound deals, draft NDAs and open the deal record
• research the market, the company and the seller, then build a scorecard with red flags
• summarise calls into the deal record and brief the team before each meeting
• populate the financial model, flag diligence risks and draft the investment memo storyline
• check shareholder agreements against standard clauses

At the end of each stage, the system detects that the work is complete and proposes moving the deal forward. No deal advances, and no email is sent, without an analyst’s approval.

3. Build the data foundation. I audited the whole platform (122 tables, 303 database migrations), linked every workflow step to the data it reads and writes, and closed the gaps with new tables for stage approvals and reminders. Every write goes through one gateway that decides whether an action runs automatically or waits for confirmation.

4. Hand the building to the team. I wrote a step-by-step playbook that lets non-technical analysts specify an automation in plain language, have Claude Code build and test it in a safe copy of the system, and submit it for review before it goes live. Each analyst started with a backlog of first automations for their own workflow.

5. Start with the inbox. I took the AI email triage dashboard from prototype to production. It reads every incoming thread across the team’s 16 mailboxes, classifies it (deal, investor, searcher, task or back office) with a one-line rationale, and lets analysts forward, reply or reclassify in one place.

• Deal decision cycles estimated to fall by more than 50%, as AI takes over research, filing and drafting between analyst decisions.
• Human approval is built in at every stage transition and every outgoing message.
• The team can automate its own workflows on a documented, governed foundation, without depending on engineers for each change.
• New email reaches the triage board 2–50 seconds after arrival, with mailbox-level privacy enforced in both the database and the API.