From transcript to thesis: an earnings research workflow

Pass 1: the watchlist round

One question per name, the evening the call happens: "Summarize today's call for X: figures, guidance, and the dominant Q&A topic." Ten names take twenty minutes, and each summary arrives with quotes and sources, so nothing needs re-verification later.

Anything surprising gets a flag, not a follow-up. The point of the first pass is coverage, not depth; depth is pass two.

Pass 2: drill into the flags

For each flagged name, open a fresh conversation and drill: what exactly was said, what changed against the previous quarter, what the tone did. This is where the model tier matters: switch to deep reasoning for the questions that cross several calls, keep the fast model for lookups.

Ask for the development as a chart when a trajectory matters. Three quarters of revenue on one axis settles debates faster than three summaries.

Pass 3: the sector read

Companies do not report in a vacuum. "Give me a read of the sector this quarter: who raised guidance, who cut, and what themes repeat across calls?" turns the individual flags into context. If two competitors named the same constraint, that is a sector fact; if only one did, that is a company fact. The distinction is the beginning of a thesis.

Pass 4: the thesis check

Write the claim down, then attack it: "What in the recent calls contradicts the view that X?" An answer grounded in transcripts either produces the counter-evidence with quotes or says the calls do not contain any. Both outcomes are useful, and neither is available from a model that fills gaps with plausible fiction.

Exporting without losing the ground

Every answer copies out as markdown with its sources attached; charts flatten to text bars, tables stay tables. Pasted into notes, an investment memo, or another model, the finding keeps the speaker, the call, and the date behind every claim. A workflow is only as good as what survives the handoff, and the handoff here loses nothing.