Quorum is the automation we shipped for an India-based venture fund — pulling transcripts from four AI notetakers, mapping each conversation to a startup in the portfolio, and drafting structured due-diligence memos directly into the analyst's ClickUp workflow. The interesting part wasn't the model. It was the trigger — which we got wrong three times before we got it right.
Quorum is a meeting-notes-to-IC-memo automation we built for an India-based venture fund. It pulls transcripts from four AI notetakers, maps each call to a startup in the portfolio, and drafts structured due-diligence memos directly into the analyst's ClickUp workflow. The engineering story is the trigger. We tried four — a cron job, contextual toggles, an explicit checkbox, and finally a ClickUp stage-change webhook with a ten-minute delay — before we landed on the one that produced near-100% adoption. The lesson generalized: in AI workflow tools, the trigger is the product. The model is the cheapest part.
Early-stage venture funds are not behind on AI. They're drowning in it. Every founder pitch is recorded by Fireflies. Every analyst call is captured by Read.ai. Every partner conversation lives in Fathom or Granola. The transcripts are accurate, searchable, and entirely orthogonal to the work the fund actually does, which is producing investment-committee memos that defend specific allocation decisions in front of partners on Fridays.
The brief, when we received it, was straightforward in the way a brief is straightforward right before it teaches you something. Take the transcripts. Run them through Claude. Produce a structured IC memo per startup, dropped into the right ClickUp doc, no human in the loop. The fund's analyst could stop spending Saturday afternoons writing memos by hand. We agreed in writing. We shipped it in three weeks.
“The first version was correct. Nobody used it. 'Correct' is not 'usable' — the difference lived entirely in the trigger.”
It worked the day we shipped it. Nobody used it. The transcripts were appearing in ClickUp on schedule. The memos were correct. The analyst was still writing memos by hand on Saturdays. The first version of this product taught us that 'correct' is not the same as 'usable' — and that the difference lived entirely in the trigger.
What followed was three rebuilds of the trigger and one durable observation: in AI workflow tools, the model is commoditized, the integrations are commoditized, the dashboard is commoditized, and the trigger — where it lives, what it watches, when it fires, how the user feels in relation to it — is the only thing left that compounds.
Four trigger designs. Only the fourth produced adoption. The full story is in our published essay.
Three things weighed in our favor. First, we'd shipped AI workflow tools before and were comfortable in the messy middle — integrating multiple third-party notetaker APIs, structuring LLM output into typed schemas, and wiring everything into a tool the analyst already lived in (ClickUp). Second, we made the right call to focus on the trigger early; most agencies would have shipped the spec and moved on. Third, we were small enough that the analyst became part of the loop — we watched her work, asked her to draw the path of a single deal through her day, and rebuilt around what we saw.
The deeper alignment was on philosophy. We believe — and we said this in the proposal — that the LLM is the cheapest part of any AI-workflow product. The expensive parts are the integrations, the data shape, the trigger, and the way the user feels in relation to the system. The fund wanted a partner who agreed with this read of the field, not a partner who would let them buy a longer prompt.
The trigger lives where the analyst already works — moving a deal-card between stages. When a card moves, a webhook fires, the system waits ten minutes (room for the analyst to undo or add context), and only then runs the memo generation. The trigger is the analyst's existing motion, not a new button.
Startups don't move forward linearly. They go EP → FM → back to EP → FM again. We track every conversation by which stage occupancy it falls under. The IC memo for a startup in FM is generated only from conversations that happened during the most recent FM occupancy. Smearing evidence across stages was the version of the bug that nearly broke the product.
Each notetaker (Fireflies, Read.ai, Fathom, Granola) has its own adapter under `server/src/services/`. Adding a new notetaker is a new adapter, not a code surgery. When Granola moved from MCP to a personal API key one Monday morning, we rewrote the adapter in 36 hours without touching the rest of the system.
Layer two of the trigger journey: a UI surface where the analyst can toggle which Google Chat spaces the AI is allowed to read for context. The act of explicit consent — moving a decision from us to her — tripled the memos she touched. Toggle as automation primitive.
Each IC memo follows a fixed structure of forty-two diligence questions grouped by category. The Claude call returns a typed JSON object that maps cleanly to ClickUp doc sections. The structure ensures the memo defends the same things every time, in the same order, with the same evidence anchors.
After we discovered the synchronous handler had failed for 104 consecutive deliveries (ClickUp's webhook timeout is 30 seconds; our pipeline was sometimes slower), we moved everything to a delayed queue with retries. The ten-minute delay turned out to be a feature, not a workaround — it gives the analyst time to undo a misclick. The bug picked the design.
MethodologyAdoption observed by the analyst and the fund's internal review of the IC memo workflow. 'Near 100%' is measured against the count of cards moved through diligence stages in a given week. The engineering story is the trigger; the underlying conversion of insight to outcome is documented in our published essay on the build.
The memos became defensible the moment the system was responding to my actual work — not running on a schedule somewhere I couldn't see.
A multi-fund deployment with shared notetaker integrations and per-fund prompt configuration. A retrieval layer over the fund's historical memos so the AI can cite the fund's own prior language and judgments. And a deeper closing-the-loop mode where the system surfaces, at memo generation time, the diligence questions that previously distinguished portfolio winners from losers.
Quorum is portable as a pattern. Sales tools, recruiting pipelines, status reports, customer-success automations — anything that turns conversations into structured work shares the same question: where does the trigger live. We can scope a Quorum-shaped automation for your team in 4–8 weeks.