From Copilot to Claude: six builds in six months for a broker-dealer
A capital-raising broker-dealer replaced Microsoft Copilot with Claude after two hands-on sessions, then kept building: a meeting-prep agent that assembles a full GP pack in 7.5 minutes, one search across 103,825 contacts, and market-data research that costs a few credits instead of hundreds.
01The situation
The firm is a federally regulated broker-dealer that raises institutional capital for private fund managers, a senior team where partner time is the scarcest resource. Its only sanctioned AI tool was Microsoft Copilot, while teams explored the same ideas in parallel with no coordination. The work itself ran on manual assembly: every warm introduction meant searching five membership networks and five databases one at a time, each GP meeting pack was stitched together by hand from around fifteen PDF exports, and an inherited market-data integration burned roughly 200 PitchBook credits on every query.
"The salesperson needs to be talking to people, not doing research."
R.D.Managing Partner
02What we built
Claude, firm-wide two hands-on sessions covering Claude Code, Projects, Skills and MCP connectors, ending in a live build of the team's biggest pain point. The firm then replaced Microsoft Copilot with Claude and connected PitchBook and Dakota to it through MCP.
A meeting-prep agent on Claude running on Claude Sonnet 5, it turns CRM notes and a couple of LinkedIn profiles into a 14-page brief and a 7-tab workbook, pulling live Preqin and web research as it goes.
One search, cleaner data a secure app that unifies five membership networks and five databases, 103,825 contacts, into one index, plus a discovery engine moved from Azure OpenAI onto Claude Opus 5 that reads incoming prospect lists for matching against the CRM, with a person checking every questionable match.
Built with Claude Code, owned by the firm a PitchBook integration rebuilt to cost a few credits a query, and a new website rebuilt off Squarespace, both on the firm's own Vercel team. An LP news agent on Claude is next.
03Before and after
Microsoft Copilot was the only sanctioned AI tool, while separate teams explored the same ideas in parallel.
Each GP meeting pack was stitched together by hand from around fifteen PDF exports.
An inherited market-data integration burned roughly 200 PitchBook credits a query, about $23,000 a year.
A dedup process so unreliable one prior run mismatched 500-plus records, fed by a 25-record export cap that forced dozens of manual pulls per campaign.
Claude replaced Copilot after two sessions, and the training fee was credited into the builds that followed.
A 14-page brief and a 7-tab workbook from two uploads, including data the manual pull missed, like key departures and full investor lists.
Free search endpoints first and targeted paid calls second, validated on a sandbox contract before any production credit was spent.
Confidence-interval matching replaced the 50-60% process, with every questionable match checked before it reaches the CRM.
What was stitched together by hand from about fifteen PDF exports now arrives as a 14-page brief and a 7-tab workbook in about seven and a half minutes, on Claude Sonnet 5.
"This is the machine doing the work for me. Love it."
Smaller project? Hire a Claude-certified engineer on YAID Hire: one brief to up to three engineers, each with published prices.
