Sales pipeline data extraction — durable entities

Stop feeding the agent raw call transcripts every time you ask about a deal. Instead, after each call, run a one-time extraction that pulls out durable entities: decision-maker contact info, role, decision criteria, agreements made, objections, next-step commitments. These entities live as structured markdown files per deal, per contact. Future operations read the entity store, not the transcript.

On a B2B sales motion with long decision chains, the same call gets re-analyzed dozens of times — for next-step prep, for handoff to CS, for the renewal motion six months later. Each re-analysis is paid LLM tokens plus drift risk. A B2B sales practitioner reports: "Why repeat the same computations every time?" Extracting once and reusing across operations also lets you spot cross-cycle signals — "this decision-maker moved to a new company; that's a cross-sell signal" — that pure transcript-on-demand workflows miss.

1. Define the entity schema for your sales motion. Typical fields: company, contact name, role, channel, last_call_date, decision_criteria[], objections[], commitments[], status.
2. Write a one-shot extraction prompt that reads a transcript and emits the schema as YAML / markdown.
3. Save extracts as deals/{Company}.md and contacts/{Name}.md keyed by stable IDs.
4. Build a per-deal merge step that updates existing entities when a new call comes in — never overwrite history; append timestamped diffs.
5. Wire downstream skills to read entities, not transcripts. "Next-step recommendation," "renewal-readiness scoring," "handoff brief" all consume the structured files.
6. Periodically re-scan: did any contact move companies? Cross-sell signal.

• Schema drift between sales reps. Lock the schema in one markdown file; if a rep wants a new field, they amend the schema, not their personal extracts.