The model reads the transcripts. The analyst makes the call.

Asked about a company with nothing in front of it, a model answers from its training data, months old, uncited and fluent enough to pass.

Research synthesis starts from the fund’s own position. The analyst gives the model the filings, the latest transcripts and the team’s prior notes on the name, then states the thesis and what would break it. The model reports what changed against that thesis, each point cited to the passage it came from, and says plainly where the documents are silent. The analyst reads the exceptions and decides whether any of them matter.

Earnings season is reading at volume on a clock. The work a model does well is the comparison: this quarter’s release and call against last quarter’s, line by line, on the points the thesis depends on.

A workshop builds the instruction for each step on names the team covers, so it runs the same way across the list rather than one analyst at a time.

With a one-page usage policy the desk actually reads, written with the compliance officer before anyone starts. The page names the approved tools and says what may go in, in plain terms: public filings and transcripts, and the fund’s own notes on a name. Material non-public information about a listed name stays out, with personal data and anything held in a tool off the approved list. One named person answers the edge cases by morning.

Compliance usually also wants an audit trail and a written account of the rollout before the first session. Workshops run in the fund’s approved tool, or on public material and mock files, so nothing sensitive has to move. The retainer keeps the policy current as the tools change. The usage policy guide has the full page.

The engineer sits with the analysts and builds on the fund’s own data, scoped to one process with real volume: the earnings comparison across the coverage list, say, or the allocator DDQ the fund answers from a blank page each time one arrives, when most of its questions were answered last quarter.

Discovery comes first: interviewing the people who do the work and documenting the process as it runs, which is rarely the process as written. Then a design that says which steps a model takes and which decisions stay with the analyst, the PM or compliance. Then the fund’s own people build it, with us guiding the Claude skills, agents and workflow. Nothing depends on us after the last day, and no vendor holds the process.

With one desk and its own coverage, after compliance has signed the one-page policy. A workshop of ninety minutes to two hours shows which steps of the analysts’ week a model earns its place in, and which ones it does not.