Financial services / AI
Financial services / AI case study — how Meritt Dev built and runs Private equity firm: an assistant that answers from the firm's own files, with citations Private equity firm: an assistant that answers from the firm's own files, with citations
A private equity firm kept its institutional memory in four places: the deal CRM, about 146,000 SharePoint files, partners' mailboxes and meeting notes. Finding an answer meant someone digging through all four by hand.
Overview
We built a Microsoft Teams assistant that answers questions over those sources and cites where each answer came from. Delivered by Meritt's founders as engineering leads on the project, working with our US partner studio.
- [01]
The challenge
The firm wanted one assistant on the firm's files rather than every partner pasting documents into a personal chatbot, with firm-level control over what it may and may not answer, and an AI bill of about $2,000 a month.
Questions about carried interest, salaries, HR matters and investors' banking details had to be refused before any model saw them, and the CRM had to stay read-only.
- [02]
What we built
A Teams bot with an identity check against the firm's directory, an allowlist of users, and policy checks that run before the model is called. Specialist sub-agents handle the CRM, the document library, the mailbox and meeting notes, and a coordinator composes the answer with its sources.
Alongside it: a deal scout that finds acquisition targets from public data, ranks them against the firm's buy box and checks the CRM for existing records, and a weekly firm digest.
Built on Node with GPT models hosted in Azure and deployed on Azure Container Apps, with the firm's data copied into its own store so nothing is queried live from production systems.
- [03]
Where it is now
Kickoff on 22 July, testers using it on real work from 3 August, and production on 3 September: six weeks from kickoff to a pilot group using it daily.
On the firm's own test set the pass rate went from 60% to 100% in two days of tuning, with a median answer in 33 to 43 seconds. A board-composition question that used to mean opening several documents came back cited in 50 seconds; a request to change a deal record was refused in 7 seconds. A full test run costs $3 to $7 in model usage.
Related