AI is reshaping how venture firms find, evaluate, and support companies—but only where the work is structured enough to model.
Blake Aber · Predicate Ventures · 2026
Where AI Fits in the Venture Workflow
Venture capital runs on three repeatable activities: sourcing deals, evaluating them, and supporting the companies you back. Each has a different ratio of structured data to human judgment, and that ratio decides where AI earns its keep.
Sourcing is the most data-rich part of the job. Diligence sits in the middle. Portfolio support is the most relational, and the hardest to automate.
The firms getting value from AI are not chasing a single tool. They are mapping each task to the right method and accepting that some parts of the job stay manual.
Deal Sourcing
Most firms still source through warm intros and inbound. That approach scales poorly and skews toward networks the partners already have.
AI changes the math by reading the entire market instead of the slice you already see. Models can scan company registries, hiring data, product launches, code repositories, and funding announcements to surface companies before they raise.
Signal extraction
The useful work is signal extraction, not list generation. A long list of startups is worthless. A short list of startups that just doubled engineering headcount, shipped a second product, and started hiring a head of sales is a real lead.
The models that work combine multiple weak signals into a ranked view. No single data point predicts a good investment. The combination narrows the field.
Ranking against your thesis
Generic sourcing produces generic results. The advantage comes from training the ranking on your own outcomes—which deals you took, which you passed, and how they performed.
This turns a partner's taste into a reusable filter. It does not replace the partner. It lets one analyst cover ten times the surface area.
Diligence and Evaluation
Diligence is where AI saves the most time and creates the most risk. The time savings are real. The risk is that a confident summary hides a shaky source.
Document review
Data rooms are dense and repetitive. Language models read contracts, financials, cap tables, and customer agreements faster than any associate and flag the items that need a human.
Use them to find the questions, not the answers. A model that surfaces an unusual liquidation preference or a missing assignment clause has done its job. A model that tells you the company is a good investment has overstepped.
Market and competitive analysis
AI compresses the early hours of market work. It maps competitors, estimates market size from public data, and assembles a first draft of the landscape.
Treat that draft as a starting point. The numbers need checking, and the framing reflects the model's training, not your conviction.
The hallucination problem
A model that invents a fact in a diligence memo is worse than no model at all, because it produces false confidence. Every claim that feeds an investment decision needs a traceable source.
The firms doing this well require citations on every generated statement and reject outputs that cannot point to a document. This discipline is the difference between a tool and a liability.
Portfolio Support
Once you invest, the value of AI shifts from selection to operations.
Monitoring
Firms manage dozens of companies with thin teams. AI reads board decks, financial updates, and product metrics to flag the companies drifting off plan before the quarterly call.
This is pattern detection, not advice. The flag prompts a conversation. The partner decides what to do.
Talent and customer matching
Much of a firm's post-investment value is connection. AI matches a portfolio company's hiring needs against the firm's network and matches their product against potential customers inside the portfolio.
The matching is mechanical. The introduction still requires a person who knows both sides.
Build Versus Buy
Most firms should buy. A handful with engineering teams and a clear edge should build.
When to buy
Vendor tools cover sourcing databases, document review, and CRM enrichment. They improve quickly and spread the cost of model development across many customers.
If your firm's edge is judgment and relationships rather than data infrastructure, buying lets you stay focused on the work that wins deals.
When to build
Build when your proprietary data is the asset. A firm with twenty years of outcome data and a distinct thesis can train ranking models that no vendor can replicate, because no vendor has the data.
Building also makes sense when your workflow is unusual enough that off-the-shelf tools force you to change how you operate.
The middle path
Most firms land between the extremes. They buy the infrastructure and build a thin layer of proprietary logic on top—custom prompts, scoring weights, and connections to their own deal history.
This keeps engineering cost low while preserving the part that reflects the firm's view.
What AI Does Not Do
AI does not pick winners. It widens the funnel and speeds the review, but the judgment about a founder, a market shift, or a moment in time stays human.
It does not build conviction. A model can tell you a company is growing. It cannot tell you whether you believe in the team enough to wire money.
It does not replace relationships. The best deals still come from trust built over years, and trust is not a feature you can ship.
How to Start
Pick one task with clear inputs and outputs. Deal sourcing or document review are good first choices because both have measurable results.
Run the AI process alongside your existing process for a quarter. Compare what it surfaces against what your team finds. Keep the tool only if it adds deals or saves hours you can prove.
Expand from there. A firm that adds one working AI process per quarter will operate differently in two years than one that waited for a single platform to solve everything.
The firms that win with AI treat it as a set of specific tools for specific jobs, not a replacement for the judgment that defines the business.
Top comments (0)