Case study · X& Financial · sanbo.io

Claude Managed Agents for private-market operations

An 18-month engagement building production AI for X& Financial, whose sanbo.io platform manages private-market investments. Three Claude-powered agents now work alongside the firm's analysts every day.

Context

X& Financial runs sanbo.io, a platform that manages private-market investments across six VC funds. The work is document-heavy: fund reports, investment memos, due-diligence notes and filings, all of which depend on analysts reading large volumes of material carefully and on time.

Over 18 months we built production AI into that workflow, starting with a shared assistant and expanding into purpose-built research and reporting agents. Each one is a managed Claude agent running on infrastructure the firm controls, with custom tools that reach into the sanbo.io product API.

What we built

A shared AI employee for the whole firm

A managed Claude agent that works inside Slack and inside sanbo.io, available to the whole team the way a colleague would be. It runs on a proprietary harness we built, with custom tools and an MCP layer that connect it to the sanbo.io API, a context engine that pulls in the firm's emails, shared drive, database, Notion and meeting notes, and a memory layer the firm controls. Model-provider processing runs with zero retention.

Investment-research and memo agent

An agent that drafts investment memos from source documents: company filings, data rooms and market material. Every claim traces back to a source, and each draft passes verification gates, red-team review and analyst sign-off before it reaches a partner. A first draft that used to take an analyst five hours now takes five minutes, with the analyst's time going to judgement rather than assembly.

Fund-report agent

An agent that prepares fund reports from live data, then routes every section through review before anything is issued. Preparation that used to take a week now takes 3.5 hours, and the reports go out on time because the analysis is already up to date when the review starts.

Controls and delivery

aHuman review for consequential output. Nothing consequential goes out without a person signing off.
bTraceability. Every generated claim points back to the source it came from, so reviewers can check rather than trust.
cRole-based access. Each agent sees only what the people using it are allowed to see.
dFirm-controlled context. Memory, documents and tool access stay under the firm's control, and model-provider processing runs with zero retention.

Results

$200m+

managed across six VC funds on the platform the agents serve

5 min

memo first draft, down from 5 hours of analyst time

3.5 h

fund-report preparation, down from a week

226

custom tools giving the agents governed access to the product API

Next step

If this sounds like your problem, tell us what you're trying to do.

Get in touch→