Give the model last quarter's letter as the template and a sourced data pack as the only input: each company's latest update, metrics with links, and public news. Ask for one short paragraph per company, with a citation on every number. Then a person checks each number against its source before anything goes to LPs.
Earlier this week I walked a fund through this workflow on a call. The thing that made the draft usable was last quarter's letter. Once the model could see how the fund already writes about its companies (the length, the order, what it leaves out), the output stopped sounding like a press release and started sounding like the fund.
That's the whole idea of this post. Most of the risk in an AI-drafted LP letter sits in two places: what you feed the model, and how you check what comes back.
Why the portfolio section is the right place to start
Q3 letters are due in the next few weeks. ILPA's Principles recommend quarterly reports within 45 days of quarter end, and most of a small team's time goes into the portfolio section: 30 or 60 short paragraphs, each pulled from a different founder update, deck or news item.
It's also the part of the letter that suits AI best. Each paragraph is a summary of material you already have. The fund-level numbers are a different job. NAV, IRR, TVPI and DPI come out of your fund admin and your valuation process, and a model should never be the source for any of them.
Step 1: Build the data pack first
Data comes before agents. If you point a model at a folder of half-read emails and ask for a letter, you'll get confident paragraphs with numbers you can't trace (NIST calls this risk "confabulation" in its generative AI profile). Fixing the input is cheaper than catching every invented number later.
For each company, the pack needs:
- The latest update or board deck, with its date.
- The metrics you'll mention, each with a link to the email, deck or document it came from.
- Public news since last quarter: funding, launches, key hires, press.
- A partner note, if there is one: "mention the new CFO", "keep it short, they're mid-raise".
If a company sent nothing this quarter, say so in the pack. Don't let the model fill the gap from last quarter's numbers without saying that's what it did.
This is also where the founder-reported metrics review pays off. A number that's already been checked against the company's history is safe to hand to a model. One that hasn't been checked isn't.
Step 2: Use last quarter's letter as the template
Upload last quarter's letter and ask the model to rewrite the portfolio section with this quarter's pack. That one step does more than any style guide. It gives the model your paragraph length, your section order, how you name metrics and how candid you are.
Then add the rules that the old letter can't show:
- One paragraph per company, no longer than the old one.
- Every number gets a citation to an item in the pack.
- If the pack has no source for a number, leave a placeholder like
[need Q3 revenue]rather than a guess. - No adjectives the source doesn't support. "Strong quarter" needs a reason in the pack.
Step 3: Check every number against its source
This is the step that makes the draft safe to send. Treat the draft as a draft and check it in two passes.
Pass one: numbers. Click every citation and confirm the number, the period and the unit. A number with no citation is wrong until proven otherwise. Watch for the quiet errors: a Q2 figure labelled Q3, annual revenue described as ARR, a monthly figure multiplied by twelve without saying so.
Pass two: judgment. Read it as an LP would. Is a company with bad news getting a cheerful paragraph? Did the model leave out something the partners would want said? Does any paragraph reveal something a founder shared in confidence?
Then a named person signs off. Humans stay in the loop here, every quarter.
Step 4: Keep the sources for next quarter
Save the data pack and the citations with the letter. Next quarter, the old letter and its sources are your starting template, and if an LP asks where a number came from, the answer takes a minute. (Auditors ask similar questions, and the same pack answers them.)
The teams I see doing this well treat the pack as the real product and the letter as one view of it. At Cura, every metric we pull from an email or a deck links back to where it came from, which is what makes the citation step cheap. But a shared drive with a dated folder per company works too, as long as every number in the letter points somewhere.
What should the AI get as input for an LP letter draft?
Three things: last quarter's letter, so it copies your structure and tone; a data pack with each company's latest update, metrics and the date and source of each one; and a short list of what the partners want said about specific companies.
Leave out anything you wouldn't want quoted back to an LP. Models summarize what they're given, and a sensitive detail from a board deck can end up in a paragraph without anyone asking for it. If you use a general-purpose AI tool, check its data settings before you upload anything confidential.
How do you check an AI-drafted LP letter before it goes out?
Check every number against the source it cites, and flag any number with no citation as wrong until proven otherwise. Then read for tone and for what's missing, like a company with bad news that got a cheerful paragraph.
A named person signs off, and the fund-level figures come from the fund admin, not the model. If the letter goes beyond your existing LPs or promotes a new fund, ask whoever handles compliance whether the SEC marketing rule applies. That depends on how the letter is used, not on who wrote it.
Which parts of the LP letter shouldn't AI write?
The fund-level numbers (NAV, IRR, TVPI, DPI) and the marks behind them should come from your fund admin and valuation process, not from a model. So should anything about a sensitive situation: a down round, a shutdown, a dispute.
AI can draft the routine company paragraphs well, and that's most of the page count. The judgment calls stay with the partners. (Honestly, they're also the paragraphs LPs read most closely.)