Where they were
The client had built and operated a custom CRM for several years. Records were structured, the team was disciplined, and every meaningful interaction was already logged.
What the system could not do was reason. Looking up a contact returned a row, but not the surrounding context. The team wanted to apply AI to that gap, but knew an AI layer on top of a flat data model would not work.
The first conversation was not about the agent. It was about the schema underneath.
What we did
The work rebuilt the foundation and the layers above it as one system: connected entities, nightly intelligence pipelines, a custom workflow dashboard, and an AI agent operating on the same data.
Created a connected-entity data model
We created a Postgres schema for companies, contacts, investors, funds, transactions, meetings, opportunities, and engagements as connected entities rather than isolated tables. A single company can play multiple roles without being duplicated, and row-level security enforces confidentiality at the database.
Built a nightly intelligence pipeline
The four-stage pipeline collects from email, calendars, meeting recordings, shared drives, chat, and outreach tools. It cleans and deduplicates records, reads transcripts and email content, extracts deal signals, relationship intelligence, action items, and topics, enriches records, and recomputes priority and warmth scores.
Built a custom dashboard around the operating workflow
Pipeline kanban, opportunity timelines, meeting summaries, contact warmth, and relationship trajectories all run off the same source. The progressive web app works on laptop and phone, with role and confidentiality rules layered above database-level access.
Layered an AI agent on the same data
The agent has tool access to the database and workspace integrations, so it can read live context and write structured output back. It summarises meetings, drafts outreach, enriches counterparty profiles, answers plain-English database questions, produces firm-level reports, and gathers data continuously.
Shipped capabilities the old CRM could not deliver
The system added relationship-health scoring, counterparty matching, structured meeting intelligence, and full provenance so every fact carries its source and timestamp.
Where they are now
The custom CRM has been replaced with a system that holds the same disciplined record but reasons over it. Scattered firm knowledge is now reusable institutional memory.
Meeting prep, debriefs, counterparty research, and outreach drafts that used to take hours take seconds and improve with use rather than going stale.
It was a rebuild of the operating model: a connected-entity data foundation, a custom interface for the actual workflow, and an AI agent operating on the same source of truth.
The result was not an AI feature added to a CRM.