AI in VC Deal Sourcing: A Practical Guide to What Works
Where AI genuinely helps VC deal sourcing (enrichment, monitoring, triage, drafting), where it is theater (autonomous selection), and the MCP-plus-agent weekly loop that runs discovery on proprietary data in under an hour of human time.
Key Takeaway
AI collapsed analysis cost but not discovery cost. This practical guide maps the four jobs AI does well in deal sourcing, exposes the circularity failure of autonomous deal selection, and details the working pattern: an MCP server feeding company data to an agent runtime that monitors, enriches, and drafts while the human keeps judgment, all in a weekly loop under one hour.
AI has collapsed the cost of analysis and left the cost of discovery almost untouched. Every fund can now summarize a deck in seconds; almost none have fixed the harder problem of knowing which company to look at next week. This guide is a sober map of where AI genuinely helps in deal sourcing, where it is theater, and how to wire the useful parts into a weekly loop with agents and APIs.
Where AI actually helps in sourcing#
Four jobs, in descending order of proven value. Enrichment: filling company records from public data, agents are excellent at this. Monitoring: watching a universe of companies and flagging changes, velocity, hiring, product, reliably mechanical. Triage support: ranking a shortlist against a written thesis, useful when the rubric is explicit. Outreach drafting: first drafts at volume, with human review. The AI deal-sourcing tools landscape and the Claude/Cursor sourcing walkthrough show the tooling in practice.
Where it is theater#
Autonomous deal selection. The failure mode is circularity: the model ranks what the funding databases already contain, so the "AI-sourced" list is the announced-deals list with better formatting. Discovery, the job that determines returns, still runs on proprietary observables, engineering momentum among them, fed to the model as data rather than delegated to it.
The agent-stack pattern that works#
The working pattern pairs an MCP server for data access with an agent runtime for the workflow. The VC MCP server exposes trending, sector search, single-company lookup, and methodology as tools; Claude, Cursor, or any MCP host orchestrates: monitor, fetch movers, cross-check against thesis, draft the outreach. Setup takes minutes: `npx -y @gitdealflow/mcp-signal` in any MCP host. The agent-native sourcing walkthrough covers the full installation across runtimes.
The weekly loop#
Monday: agents pull weekly movers by sector and compute the delta. Wednesday: human reviews the top decile against thesis; agents enrich the survivors with job postings and product surfaces. Friday: outreach to the clears, with the agent drafting against the evidence file. Total human time under an hour; the agents do enrichment and drafting, the human does judgment. This mirrors the weekly sourcing workflow with the agent layer doing what spreadsheets used to.
Key takeaways#
Use AI for enrichment, monitoring, triage support, and drafting; keep discovery on proprietary data the model cannot already see. The MCP-plus-agent-runtime pattern makes the useful parts a 20-minute install. And judge any AI sourcing tool by one question: does its shortlist contain companies the funding databases do not have yet.
A concrete agent session, walked through#
Monday, in Cursor or Claude Desktop with the MCP server connected: "Pull this week's top movers in developer-tools and fintech, and for the top five, summarize what changed week over week." The agent calls the trending and sector-search tools, returns a table with deltas. Follow-up: "For these three, draft outreach that references the specific repositories that accelerated." The agent drafts; you edit the one line that matters and send. Total elapsed time for a sourcing pass that used to take an afternoon: under twenty minutes, most of it your judgment. The agent-runtime installation guide covers setup across Claude, Cursor, and any MCP host.
The reason this works is the data boundary: the agent reasons over live panel data through tools, not over stale training-data memories of companies. An AI that "knows" a startup from its training cutoff is a liability in sourcing; an AI that fetches this week's numbers and computes deltas is an analyst that never sleeps.
Costs, limits, and failure modes#
The honest limits. Hallucination risk survives tool use: agents occasionally over-interpret a quiet week as a decline, so treat agent summaries as pointers to check, not verdicts. Context limits mean the cross-check layer, job boards, changelogs, still needs either more tool calls or your own eyes on the strongest candidates. Cost is negligible at sourcing scale; token spend is trivial next to any paid data subscription. The structural limit bears repeating: the agent can only reason over what its tools expose, which is why discovery still belongs to whoever owns the proprietary observable, and enrichment and drafting, the mechanical layers, are where the agents earn their keep.
What to build next#
For a fund, the natural next step after the weekly loop is a standing monitor: a scheduled agent pass that writes the weekly movers into a channel or sheet, so the Monday scan is waiting when you sit down. That is a half-day build with existing pieces, the API plus any scheduler, and it removes the last excuse, remembering to look.