Use AI to sort, enrich and prioritize deal flow by pulling company data, flagging thesis fit and surfacing signals, while investment decisions stay with humans. Monitor for bias in screening criteria so promising founders outside familiar patterns aren't filtered out.
Deal sourcing is the partners' job. But the machinery behind it very often belongs to the ops or platform lead: the inbound form, the CRM, the enrichment, the weekly pipeline review. So when a partner says "let's use AI for sourcing," the work of making that real lands on you.
I hear this on a lot of calls, usually said with a bit of dread. I think it's better news than it sounds. The platform team sits on some of the best sourcing data a fund has, and the ops team is the one that can keep an AI pipeline honest. This post is about that role: what to build, what to measure, and how to keep the system from quietly narrowing the funnel.
The biggest win is speed, not a magic sourcing model
The highest-return use of AI in the front office isn't finding deals nobody else can see. It's moving faster on the deals you already have. Competitive rounds often go to the fund that gets to a confident yes first. Say AI assembles the diligence pack in a day instead of a week: the founder's history, a market map, the customers worth calling, and what your portfolio founders know about the space. Your partners spend that week with the founder instead of waiting on research. I'm pretty bullish on in-person time still winning deals, and speed is what buys it.
I also want to be straight about what doesn't work yet. Searching your network for the best warm path to a founder still isn't cracked. And most automated signals are noise. From what I see, teams throw away nine in ten of their news alerts, and a feed with too much in it gets tuned out. I'd take a short list someone reads over a complete one nobody opens.
What AI should and shouldn't do in sourcing
Here's how I'd split the work. AI takes the steps where speed and coverage matter. People take the steps where judgment and relationships matter.
| Step | AI does | People do |
|---|---|---|
| Capture | Pull inbound from email, forms and referrals into one list | Decide which channels to invest in |
| Enrich | Add company size, funding history, founder background, product details | Spot-check for wrong matches |
| Score | Flag thesis fit against a written rubric | Write and update the rubric |
| Route | Assign to the right partner by sector or stage | Take the first meeting |
| Decide | Nothing | Everything |
The last row matters most. An AI score is a sorting tool. Once it becomes the reason a company gets rejected without a person looking, you've handed an investment decision to a model no partner reviewed.
Step 1: Write the thesis down as a rubric
AI can only flag thesis fit if the thesis is written in a form it can apply. Most funds have a thesis in a pitch deck and a lot more of it in partners' heads. So the ops lead's first job is to get it onto one page.
A simple scoring rubric might look like this:
| Criterion | 0 | 1 | 2 |
|---|---|---|---|
| Sector | Outside focus | Adjacent | Core sector |
| Stage | Too late or too early | Edge of range | In range |
| Geography | Outside markets | Willing to consider | Core market |
| Signal of traction | None visible | Early users or pilots | Revenue or strong usage |
| Warm path | No connection | Second-degree | Referral from a founder or LP |
Total it out of 10. Anything at 7 or above goes to a partner that week. Anything at 4 or below gets a polite pass after a human glance. The middle band gets reviewed in the weekly pipeline meeting.
Notice what's not on the rubric: the founder's university, their last employer's brand, or how polished the deck looks. That's where bias hides, and a model will pick those up fast if you let it.
Step 2: Build the pipeline in small pieces
You don't need a custom platform. A CRM like Affinity, Attio or HubSpot, an enrichment tool like Clay, and a general AI assistant get you most of the way.
The flow is capture → enrich → score → route. A workable first version looks like this:
- One front door. Route every inbound source into a single CRM list, including the website form, the deals@ inbox and partners' own inboxes. Capture from partner inboxes automatically instead of asking partners to forward. People who rely on forwarding forget (and they always do, eventually).
- Enrich on arrival. When a company lands, pull public company and founder data automatically. Spot-check the matches in the first weeks. If enrichment is wrong 20–30% of the time, every score built on it is wrong too.
- Score against the rubric. Have the AI assign a score and, more importantly, write one sentence on each criterion, with a link to where each fact came from. That explanation is what lets a person check the score in seconds.
- Post a weekly digest. Send a Slack or email summary every Monday: top-scored companies, new referrals, and anything that changed status.
- Log the outcome. When a partner passes or takes a meeting, record it. That record is how you'll tune the rubric later.
If you're deciding whether to build this or buy a product for it, the build, buy or wait framework applies directly. For most funds, scoring is the one piece worth owning yourself, because the rubric is your thesis.
Step 3: Watch for bias on purpose
Screening systems drift toward the patterns they were trained or prompted on. If past investments skew toward a few universities or a few cities, a model that learns from them will push those founders up and everyone else down. Nobody has to mean for it to happen.
An ops lead can guard against this with a few habits:
- Review a sample of rejects. Each month, pull a random handful of companies the system scored low and have a partner read them cold. If a partner keeps finding ones they'd have taken a meeting with, the rubric is wrong.
- Track pass-through by source. Compare how often inbound from each channel makes it to a first meeting. A channel that never converts may be weak, or the scoring may be penalizing it unfairly.
- Keep proxies out of the prompt. Don't feed the model school names or past employers as scoring inputs. If you want a person to see them, show them after scoring.
- Write down every rubric change. A short changelog with the date and reason makes it easy to see when a change shifted who gets through.
Step 4: Use the data only your platform team has
This is the part people skip, and it's where platform people have an edge. Public data vendors sell the same company lists to every fund. Your portfolio is the highest-leverage asset most funds aren't using, and your platform function holds data nobody else has:
- Portfolio founder referrals. Founders know who's building something good in their space, often before it's public.
- Event and community attendees. People who came to your founder dinner or applied to your office hours.
- Expert and operator networks. Operators you've introduced to portfolio companies often hear about new companies early.
- Intro history. Who in your network has made good intros before.
Platform work turns into deal flow. A candidate you helped place, or an operator you introduced to a founder, may call you first when they start their own company. It just takes a while to show up.
Tag these sources in the CRM so the rubric's "warm path" criterion can see them. If you've already built the habit of turning platform interactions into intelligence, most of this data is sitting there waiting. And if your platform work is part of how you win competitive deals, the same records support both.
Can AI replace VC analysts?
No, but it changes the job. AI can handle the first pass of research, enrichment and memo drafting, which used to fill much of an analyst's week. Analysts who spend the freed time on founder calls, reference checks and sharper judgment become more useful to the partnership.
What changes in practice:
- Less compiling, more checking. An analyst who used to build a market map from scratch now reviews and corrects one the AI drafted.
- More founder contact. Time saved on research can go to first calls and customer references, which no model can do.
- Owning the rubric. A strong analyst becomes the person who notices when the scoring is off and proposes changes.
What data sources power AI deal sourcing?
The usual inputs are your CRM and inbound email, company and people data from LinkedIn and data vendors, funding announcements, product launches, hiring signals and news. Your firm's own network data, such as portfolio founder referrals and event attendees, is often the most useful source because no one else has it.
A practical split for a smaller fund:
- Internal: CRM history, the deals inbox, partner notes, and pass and meeting outcomes.
- Public: company websites, LinkedIn, news, product launch sites and job postings.
- Network: founder referrals, event lists, office hours applicants and LP introductions.
I'd start with internal and network data. They're cheaper, more accurate for your thesis, and much harder for anyone else to copy.