Venture capital, deal sourcing
AI deal sourcing at a venture firm works only as well as the thesis it is given.
Venture capital firms use AI for deal sourcing in two layers. Data platforms such as Affinity Sourcing, Harmonic and Specter watch hiring, web traffic and founder moves and surface companies that match a written thesis. Models like Claude or ChatGPT read what comes back: the deck, the founder's history, the firm's own past memos. A small VC team should start with the second layer, on its own thesis and documents, before buying the first.
By Tim Miner, founder and CEO · Published
Deal sourcing is how a venture firm finds companies before a round turns competitive: inbound decks, warm introductions, demo days, and outbound work against a thesis. AI touches each of those in a different way. Some tools find companies. Others read about them. A third kind keeps the CRM current, so a partner knows who has already spoken to a founder.
The data products are large. Harmonic says it holds data on more than 35 million companies and over 195 million people. Affinity says its sourcing dataset covers over 200,000 high-growth startups. For a team of four, the limit sits after the search: deciding which of those companies deserve a partner's call.
The current practice
How are venture capital firms using AI for deal sourcing?
Start with where a new company first enters your pipeline today, and who on the team notices it.
The vendors' own pages describe four uses, and one platform often covers several of them.
Thesis search in plain language
Harmonic's agent, Scout, takes an investment thesis described in natural language and returns evaluated results rather than keyword matches. Specter lets a team save a search written in plain English that keeps returning new matches as the market moves.
Growth and talent signals
Affinity Sourcing scores companies on web traffic, hiring trends and leadership signals, and alerts teams through automated watchlists. Specter tracks stealth founders, promotions, leavers and joiners, which it puts at more than 1,000 signals a day.
Warm paths through the firm's network
Affinity, Harmonic and Specter each say they map the firm's own relationships to show a warm introduction path into a company. Specter describes connecting LinkedIn and the team's inboxes to do it.
Questions asked in a chat window
PitchBook launched Navigator in November 2025 to answer prompts with its own data, and made that data available to subscribers inside Claude and ChatGPT.
None of these tools makes the call. They produce a longer, better-ordered list, and someone at the firm still reads the deck and decides whether to take the meeting.
The failure points
Where does AI deal sourcing go wrong for a venture team?
Look at the last deck a model summarised for your team, and at which figures anyone checked.
Most failures sit after the search. A signal says a company hired fast or its web traffic jumped. It says nothing about whether the company fits the fund's stage, check size or ownership target. A model asked to summarise a deck will write a confident paragraph on market size that came straight from the deck's own slide, unchecked.
A sourcing tool ranks companies by the signals it can see. Only the firm knows why it said no to the last ten.
The second failure is a vague request. "Find interesting AI companies" returns whatever the tool ranks highest. A thesis written as criteria, with the reasons the firm passed on similar companies, returns a list a partner can act on. Most disappointing output is a specification failure: the request did not carry the firm's judgment.
A small team
How would a small VC team get started with AI for deal sourcing?
Assume three to six people, one licensed model, and a CRM that is only partly current.
Four steps, in order, before any new subscription.
Write the thesis as criteria
Stage, check size, sector, geography, the signals that matter and the ones that do not. Add five companies you passed on and the reason for each. That page becomes the instruction every tool receives.Run it on your own inbound
Take last month's decks and ask the model you license (Claude, ChatGPT or Copilot) to screen them against the page, citing the slide behind each finding. Compare the results with what the partners decided.Fix the instruction, then save it
Where the model and the partners disagree, tighten the criteria. Save the prompt as a project or skill so everyone on the team screens the same way.Then trial a data platform
With a working screen, trial a sourcing platform on the same thesis. Harmonic and Specter both offer MCP connections that bring their data into a model such as Claude, so your screen can read what they find.
Keep a person on every pass and every first call. The model drafts the screen, and a partner signs it.
Platform or model
Does a small VC firm need a sourcing platform, or is a general AI model enough?
They do different jobs. A platform such as Harmonic, Specter or Affinity Sourcing brings data the firm does not have: company records, people records, funding rounds and signals refreshed over time. A general model reads and drafts on whatever the firm gives it, including its own memos and pass notes, which no platform holds.
A team that sources mostly through its network and inbound can start on the model it already licenses. A team running outbound against a thesis in a crowded sector will want the data as well. Either way, the screen belongs to the firm: the criteria, the checks, and the judgment about what a good founder looks like at seed.
| Sourcing data platform | General AI model the firm licenses | |
|---|---|---|
| What it brings | Company, people and funding data refreshed over time, with growth and talent signals | Reading, summarising and drafting on the documents the firm supplies |
| Examples | Affinity Sourcing, Harmonic with its Scout agent, Specter | Claude, ChatGPT or Microsoft Copilot, whichever the firm has licensed |
| What it knows about your firm | Your network and CRM records, where connected | Your thesis, memos and pass notes, once you give them |
| Best first use | Outbound search against a written thesis in a crowded sector | Screening inbound decks against the firm's own criteria, each finding cited to a slide |
| What still needs a person | Deciding which surfaced companies fit the fund | Checking every figure against the source before anyone acts on it |
| How the two connect | Harmonic and Specter offer MCP connections to models such as Claude | Can read platform data through MCP or vendor integrations, such as PitchBook's in Claude and ChatGPT |
Where we come in
Who teaches a venture team to use AI on its own deal flow?
Usually the team itself, with some help at the start. A sourcing platform knows its own data. It cannot know how your partners screen: the thesis, the pass notes, the checks a model's summary needs before anyone acts on it.
We run a Prompt Engineering Workshop for one team on its own material: ninety minutes to two hours, on whichever tool the firm has licensed, with the screen above built in the room. It is the first of three engagements. The AI Adoption Retainer keeps practice going month to month, and a forward deployed engineer can sit with the team on one process, such as first-pass reading of a deal pipeline, which your people then run.
We have delivered over 150 workshops at more than 100 financial services firms. If we are the wrong firm for your team, we will say so in the reply.
Write the thesis down first. Every tool you buy afterwards works from it.
Questions on AI deal sourcing for venture firms
Can AI pick which startups a venture firm should invest in?
AI can shortlist companies, and a person should choose. Sourcing platforms rank companies by signals such as hiring, web traffic and leadership changes. A model can summarise a deck and test it against your criteria. Neither knows your fund's portfolio construction or why you passed on the last company like it. The partner who takes the meeting decides, with the source open.
What should a small VC team use for AI deal sourcing first?
The model it already licenses, on its own inbound. Write the thesis as criteria, screen last month's decks against it with each finding cited to a slide, and compare the result with what the partners decided. Once the screen works, trial a data platform such as Harmonic, Specter or Affinity Sourcing on the same thesis.
Can we upload pitch decks and founder data to Claude or ChatGPT?
Only under the firm's own data terms with that vendor, and after whoever owns compliance has said what may go in. A one-page usage policy settles it for the whole team. Our guide to a usage policy sets out what that page covers.
Do venture sourcing platforms connect to Claude?
Some do. Specter says its discovery tools can be brought into Claude over MCP, Harmonic says its MCP can connect to any LLM, and CNBC reported that PitchBook made its private market data available within Claude for subscribers in late October 2025.
How would we know AI sourcing is working for our fund?
Measure it on your own pipeline, before and after: how long a deck waits for a first read, how often a partner agrees with the model's screen, and how many first meetings come from outbound work. Counting logins or licences tells you little about any of those.
Sources
Tell us what you are seeing.
If your documents or your experience say otherwise, we would like to hear it.