A partner at a small firm prepares for a first meeting with a founder. The CRM shows two logged calls from two years ago. It doesn’t show the WhatsApp thread from last quarter, the Slack message where an associate flagged traction, the deck a portfolio founder forwarded, or the note left after a conference dinner. Meeting prep takes an hour, and half of what matters never surfaces.
That is the unstructured data problem in venture capital. The firm’s knowledge exists. It just isn’t findable.
Where a firm’s knowledge actually lives
Structured data is a small part of what a firm actually knows. Contact records, deal stages and logged calls live in a CRM. The texture of a relationship, the reasoning behind a pass, the conviction built over years of watching a founder: that lives somewhere else.
It lives in inboxes, where years of founder updates, references and co-investor threads never touched a deal record. It lives in calendars, where every coffee and board call carries context that evaporates by Friday. It lives in meeting transcripts, recorded and summarised but not connected to the right company. It lives in WhatsApp and Slack, where the fastest decisions happen and nothing is logged. It lives in decks and data rooms, in folders only the person who saved them can find. And it lives in founder updates, read once and filed.
Firms route information through many different tools, and few of them talk to each other. The result is a firm that knows far more than it can access.
What inaccessible knowledge costs
The costs are real, even when they’re hard to measure.
Missed introductions. A founder pitching one partner knows people another partner has been trying to reach. The connection exists in the firm’s network, but without a way to see it, the intro never happens.
Repeated diligence. An associate spends days building a market map for a space the firm has looked at twice before. The earlier work exists in a deck from two years ago. Nobody knew to look.
Stale portfolio numbers. A partner walks into a board meeting with figures from the last quarterly update. The founder sent a corrected model weeks ago. It went to one inbox and was never reconciled.
Slow meeting prep. Before a first meeting, a sharp associate can spend an hour pulling threads together. That’s an hour that doesn’t go to a portfolio company that needs help.
None of these is dramatic. Together they are the quiet friction of a firm running on knowledge it can’t see.
Why CRMs never fixed it
The CRM was supposed to solve this. It didn’t, and the reason is structural.
A CRM is built around a data model: contacts, companies, deals, notes. Everything a firm knows has to be translated into those fields before it is useful. That translation takes manual effort, and manual effort is exactly what a small firm in a fast market doesn’t have.
CRM adoption at small firms is low because the tool asks for more than it gives back, at least at first. Partners don’t log calls. Associates log some, inconsistently. Notes written in a hurry lose their context.
Even firms with good CRM habits hit the same wall: the CRM only holds what someone chose to put in it. The WhatsApp thread doesn’t go in. The forwarded deck doesn’t go in. The CRM ends up as the skeleton of the firm’s knowledge, not the body.
What’s changing
Language models have changed the problem in a specific way. It is now possible to read unstructured text and pull meaning out of it without anyone translating it into fields first. A transcript can be parsed. A thread can be summarised. A deck can be read and its numbers indexed.
That doesn’t solve the problem on its own. Summarising content is easy. Making the output reliable, correctable and durable is hard. The question for any firm evaluating these tools is whether the output can be trusted, fixed when it’s wrong, and kept as the firm grows.
What a solution looks like
A real solution has four properties.
It captures knowledge with sources intact. Every value traces back to the email, transcript or document it came from, down to the page. Without sources, there is no way to judge whether the output is right.
Corrections stick. When a partner fixes an error, the fix persists and doesn’t get overwritten the next time the source is read. Human judgment wins over automated inference.
It confirms before changing anything you set. Suggested changes to fields people entered should wait for a person to approve them. The firm is the source of truth, not the model.
The data is exportable. Years of institutional knowledge should be portable. Lock-in that holds a firm’s own history hostage is a liability, not a feature.
Kept is built around these four properties. It connects to where a firm’s knowledge already lives, extracts what matters, and keeps it current and searchable without anyone re-entering it.
Frequently asked questions
What is the unstructured data problem in venture capital?
The gap between what a firm knows and what it can access. Most of a firm’s knowledge sits in inboxes, chats, transcripts, decks and founder updates rather than in a structured system, so when a decision needs to be made, it might as well not exist.
Why is CRM adoption low at small firms?
A CRM needs consistent manual effort to stay useful. Partners and associates have to log every call, meeting and note for it to reflect reality, and at a small firm that overhead is hard to sustain.
What does inaccessible knowledge cost a firm?
Missed introductions, duplicated research, stale portfolio numbers and time lost on meeting prep. Each instance is small. Across a portfolio and a full investment cycle, the cost adds up in partner hours, missed opportunities and decisions made without full context.
How does AI change the problem?
AI makes it possible to extract meaning from unstructured text without manual tagging. The harder part is reliability: outputs need sources, corrections and stability over time. The technology solves for speed. A well-designed system solves for trust.