AI works well for data extraction, summarization, drafting and routine reporting, but still needs human review for accuracy-critical outputs like financials and LP communications. The most reliable stacks pair specialized tools with clean data, clear ownership and defined review checkpoints.
Most ops and platform teams I talk to are past the first wave of experiments. Someone tried a chatbot on the quarterly report. Someone else wired up a Zapier flow that summarizes founder emails, and a partner forwarded a demo of an "AI analyst." Some of it stuck. A lot of it quietly stopped running (and nobody noticed for a while).
And you're usually the one who owns the output when it goes wrong. If an LP letter goes out with last quarter's ARR, nobody blames the model. So let's be specific about where AI has earned a place in fund operations, and where it still needs a person standing behind it. But first, the thing that decides most of it.
Start with the data, not the agent
If I had to name one thing that separates teams with working AI workflows from teams still experimenting, it's the state of their data. Agents built on messy data produce slop. If your portfolio data lives in four spreadsheets with different company names, an agent will happily treat "Acme" and "Acme Inc." as two companies. AI makes clean data faster to use. It doesn't make dirty data clean.
Then there's coverage. A model only knows what it can see. Most of what matters about a portfolio company sits in private places: board decks, email threads, founder texts, Slack and WhatsApp. If the AI only sees the CRM and public news, every summary and every draft quietly misses things.
I got this wrong at first. When I started, I thought a general assistant like Claude would be enough on its own. I was pretty quickly incorrect about that. A context window isn't a database. You can't throw 40 portfolio companies and four years of history into a chat and expect consistent answers. You still need a system of record underneath: one list of companies with one name each, one place where metrics live, and one owner per workflow.
It's not glamorous. But it's the groundwork for everything below, and for any firm aiming at a more agentic fund. The post on how AI is changing the platform team covers the people side.
What has worked
The wins have a pattern. The task is high volume, the input is messy text, and a person can check the output faster than they could produce it.
Pulling numbers out of documents. Founders send metrics in every format: a board deck PDF, a screenshot of a spreadsheet, three lines at the bottom of an email. Models are good at reading those and returning revenue, burn, runway and headcount in a consistent shape. Good isn't perfect, though. I once assumed metric extraction was a solved problem, and I was shown otherwise fast. The time saved is still real, because the old version of this job was retyping. Just make sure every number links back to the page it came from.
Summarizing long inputs. Board decks, founder updates, call transcripts and news coverage all compress well. A two-paragraph summary of a 40-slide deck is a fine first read for a partner who has five minutes before a call.
First drafts. LP update sections, portfolio newsletter blurbs, event follow-up emails, meeting recaps. A draft that's 70% right still beats a blank page, as long as everyone treats it as a draft.
Routine reporting. Anything recurring with a fixed shape, like a monthly portfolio snapshot or a weekly digest of what happened across the portfolio, is a good fit. The structure stays put and only the inputs change.
Searching your own history. Asking "which of our companies hired a VP Sales this year?" against your notes and CRM works well, once the data lives in one place.
What has failed
The failures have a pattern too. Someone asked the model to be the source of truth instead of a fast first pass.
AI use that lives in one person's head. I see this one constantly. Giving every partner a ChatGPT login and calling it done gets you scattered work nobody can repeat. Nobody can see what was asked, what came back, or whether it was right. And when the one enthusiastic person leaves, their prompts and workflows leave with them. Individual use should feed a shared layer the whole firm can build on.
Mistaking a demo for a product. In my experience, a demo or a weekend build gets you 60–70% of the way. The last 30% takes months, and that's where the edge cases live: the odd currency, the restated quarter, the founder who reports bookings as revenue.
Letting the model do the math. Extraction is reliable. Calculation inside a prompt is less so. Ask a model to compute net IRR or a cap table waterfall from scratch and you'll sometimes get a confident wrong number. Pull the figures with AI, then do the math in a spreadsheet or your fund admin system.
Unreviewed external communication. LP letters and founder emails sent straight from a draft are where the damage happens. A wrong number in an internal summary costs you a correction. A wrong number in an LP letter costs you trust.
Tools without owners. Honestly, the most common failure is the boring one. Someone builds a workflow, gets busy, an API key expires, and the workflow fails silently for two months.
A simple way to sort your tasks
Before you add AI to a workflow, place it on two axes. How costly is an error? And how easy is it to check the output?
| Easy to check | Hard to check | |
|---|---|---|
| Low cost of error | Automate fully. Meeting recaps, news digests, tagging inbound requests. | Automate with spot checks. Summaries of long documents, first-pass research. |
| High cost of error | Automate the draft, require sign-off. LP update sections, extracted metrics. | Keep a human doing the work, AI assists. Valuations, capital accounts, legal terms. |
Most fund ops work lands in the bottom row, and that's fine. For the bottom-left box, you're trying to shrink the human part from "do it" to "check it." That's still a big change in hours.
Review checkpoints that people actually use
A review step only works if it's specific and fast. "Someone should look at this" turns into nobody looking at it. Here's the checklist I'd use for any AI output that leaves the firm:
- Numbers trace to a source. Every figure links back to the deck, email or file it came from. If you can't click through to the source, it doesn't ship. The reviewer's job is to spot-check those links, not redo the work.
- Period check. Confirm the quarter or month on every metric. Wrong-period numbers are the most common extraction error I see.
- Names and tone check. Company names, founder names, and anything that sounds like a promise to an LP.
- One named approver. Put a person's name on the workflow, not a team name.
- A log. Keep the draft and the final version so you can see what the reviewer changed. After a few months, that log tells you where the model is weak.
Say your fund has 25 portfolio companies and sends a quarterly LP update. Before AI, the ops lead spent most of that effort collecting and retyping. With extraction and drafting in place, the flow becomes extract → draft → check sources → send, and the same person spends their time on the checking and the tone. Fewer hours overall, and the ones left go to the part an LP actually notices.
This is the workflow Cura is built around. It sends metric requests from your own inbox, pre-fills what it already has, and asks founders only to confirm or fill gaps. It follows up automatically, extracts the numbers from replies and documents, links every number to its source, and drafts the LP update so your team reviews instead of retyping. You still approve it. There's more in the LP update guide.
Which AI tools are VC ops teams using?
Most ops teams run a mix of a general assistant like Claude or ChatGPT, the AI features inside tools they already pay for, and one or two specialized products for portfolio data or fund admin. The setup that holds up is a small stack with a named owner for each tool.
In practice, the stack usually looks like this:
- A general assistant for drafting, summarizing and one-off research. Pick one for the whole firm so people can share prompts and outputs, and connect it to your portfolio data so it isn't answering from memory.
- AI inside existing tools. CRMs like Affinity, Attio or HubSpot, and workspaces like Notion, now ship AI features. These are often the cheapest place to start, because the data is already there.
- Glue. Zapier or a similar no-code tool to move data between systems on a schedule.
- A specialized portfolio tool for metric collection and monitoring, because that workflow has enough edge cases to justify a product built for it.
I'd hold off on a fifth category until the first four have owners and actually work.
Where does AI still need human oversight in fund ops?
Anything a partner, LP, auditor or founder will treat as fact needs a human check: financial figures, capital account numbers, LP letters, valuation inputs and anything sent outside the firm. AI can extract and draft, but a named person should sign off before the output leaves the building.
Internal, low-stakes outputs are different. A weekly news digest or a meeting recap can run with the occasional spot check. Here's a test I like: if this output were wrong, who would find out first? If the answer is "an LP," a person reviews it before it goes.