AI in a VC workflow: co-pilot, never an authority
Where AI belongs in a venture fund's operations and where it does not: source and confidence on every enriched field, a review guardrail before anything reaches IC or LPs, and a propose-review-apply loop for an agent that never acts alone.
Venture is a judgment business run on thin data and thick relationships. That makes it a strange fit for the current wave of AI tooling, which is excellent at producing plausible text and poor at knowing whether the text is true. Most of the AI features I have been shown for investors fall into one of two failure modes: they are a chat box bolted onto a CRM that answers confidently and wrongly, or they are so hedged that nobody uses them.
The version that works, in my experience of building and using one, rests on a single principle: the model is a co-pilot, never an authority. It drafts, fills, proposes and summarises. A person decides, and the system is built so that the person cannot skip the deciding. Here is what that means in practice, feature by feature.
Enrichment: every field carries its source and confidence
The first useful thing AI does for a sourcing list is fill it in. Given a company name and a website, a model with web access can return the founding year, headquarters, team size, funding history, a description and a sector classification in under a minute. Done by hand this is an hour nobody spends, which is why most deal databases are empty past the company name.
The failure mode is obvious: the model returns a founding year that is wrong, the wrong year ends up in an IC memo, and someone has to explain it. The fix is not a better model. It is provenance.
Every enriched field in our system stores three things: the value, the source it came from, with a link where one exists, and a confidence level, high, medium or unverified. The interface shows all three. A team size of “42, from a LinkedIn company page, medium confidence” is information a partner can weigh. A bare “42” is a trap.
Two rules make this durable. First, manual edits win. When a person types over an enriched value, the record stores the person’s value and marks it manual, and the enrichment source is no longer shown for that field. Second, a field is only marked manual when the typed value actually differs from the enriched one; retyping the same value does not turn a sourced fact into an unsourced opinion. Small rules, but they are what keep the provenance meaningful after six months of edits.
The same applies to research the model performs on demand, recent funding rounds, news coverage: each item links to where it was found, is labelled as AI research, and is cached with a date so you know how stale it is.
Signals: computed, not asserted
A momentum score that ranks your sourcing list is helpful right up until someone asks why a company is at 74. If the answer is “the model said so”, the score is decoration.
Ours is a weighted roll-up of signals that each carry their own source and freshness: hiring velocity from public job boards, GitHub activity for developer tools, product launches, filings, plus a derived fundraise-likelihood estimate that is labelled derived, not sourced, because it is arithmetic on the last round date and stage norms. The breakdown is visible on hover: the top three contributing signals, with their readings. Search-interest data is shown but deliberately excluded from the weighting because at seed stage it is noise. None of this is intelligent. All of it is explainable, which is the property that matters.
Memos: drafted by AI, gated by a person
An investment memo is the document where an AI mistake does the most damage, because it is the document that gets read by the people making the decision. It is also the document AI drafting helps with most, because a first pass from your meeting notes, the enrichment record and the diligence checklist takes seconds and gets the partner to the interesting part, the argument, faster.
So we draft, and we gate. A memo drafted by the model is stored with a status that says so, and the interface shows “review before finalizing” until a named person marks it reviewed, which is recorded with their name and the time. Until that happens, the memo cannot be presented as the IC document. The agent, when asked to move a deal forward, reads the memo status and says whether the move is consistent with it.
The same gate applies on the LP side. The manager letter in a quarterly report can be drafted from the fund’s actual figures, which the model is given and told never to invent. It arrives labelled as an AI draft with the model name, is editable in place, and is published only by a person through the publish review. Nothing reaches an LP that a human has not read.
The agent: propose, review, apply
The most requested and most dangerous feature is the agent: ask in plain language, get things done. We built it, and the trust architecture is the whole design.
Reads answer instantly. “How much dry powder is left in Fund II” returns the figure and a small fund table with an “open in app” link. There is nothing to approve because nothing changed.
Every write is a proposal. “Mark this company as Series A” produces a card that shows the field, the current value struck through, the new value, and two buttons: apply and dismiss. Nothing mutates until a person applies. The server, not the model, computes the diff from the live record, so the card can never claim a “from” value that is not true, and it refuses values that are not in the allowed vocabulary rather than writing them.
Batch writes are checklists. “Add these three companies from my notes” produces one card with three rows and a checkbox each. You apply the ones you want.
Guarded actions never execute from chat. Anything LP-facing, publishing a report set, sending notifications, produces a card whose only button opens the application’s real confirmation dialog. The agent says this out loud: it does not publish LP material itself. A model that can be talked into emailing your investors is not a feature.
Provenance and audit follow. A field the agent writes carries source “Agent” with the date, in the same chip system as enrichment, and the manual-edit rules apply unchanged. Every applied action lands in the activity feed as “Agent did X, asked by
One API for humans and the agent. The agent’s tools call exactly the same endpoints the buttons in the interface call, with the same validation and the same audit hook. It is a client, not a side door. This is the single most important engineering decision in the whole feature, because it means the agent cannot do anything a person could not do, and everything it does is logged the same way.
What we deliberately did not build
No autonomous or background agent that acts while you sleep. No AI that reads email bodies to score relationships; metadata only, who met whom and when. No AI-generated numbers anywhere in performance reporting; TVPI, DPI and IRR are computed from ledgers by ordinary code. No AI writing to LPs without a person pressing publish.
Each of these would demo well. Each would eventually produce a story that ends with a manager apologising to an LP or a founder.
The test
If you are evaluating AI features in any tool for your fund, three questions separate the co-pilots from the liabilities. Can I see where every AI-produced value came from and how confident it is? Can anything AI-produced reach my IC or my LPs without a named person reviewing it? When the AI acts, does it propose first, and is the action logged as the AI’s, asked for by me?
If the answers are yes, yes-with-a-gate, and yes, the tool respects the shape of the job. That is the standard we built Investryx to, and you can try it for fourteen days with your own companies to see whether it holds.