Sales-call transcripts

Paste or forward sales-call transcripts and Herald extracts objections, pricing signals, feature requests, and competitor mentions with timestamps. Works with any transcript source — Zoom, Fathom, Fireflies, Gong, Otter, Grain — as long as it reaches plain text.

Herald ships paste-first and forward-first in v1 (no vendor OAuths yet) because a 40-minute call contains 10× the signal of 40 Intercom messages, and lost deals teach more than won ones.

Two ingest paths

1. Paste in the web app

Dashboard → Add a call. Paste the transcript, optionally add a date and outcome tag, submit.

2. Email forward

Forward the transcript email from Fathom, Fireflies, Otter, or Zoom to your Product’s ingest address:

fwd+{product_id}@inbox.withherald.co

Herald sniffs the format — speaker prefixes, [00:02:14] timestamps, .vtt / .srt / .txt attachments — and routes to the sales-call pipeline.

What Herald extracts

For each call:

• Participants — inferred from “Name:” prefixes and sender/recipient email domains.
• Account candidate — the company the external participants belong to, matched against your user_entities table.
• Outcome — won, lost, next_steps, at_risk, or unknown. Inferred from language patterns; you can correct it in chat.
• Top labels per chunk — objection, pricing, feature_request, next_steps, intro, irrelevant.
• Summary — one paragraph.

The full transcript lives in R2 at r2:/{product_id}/calls/{call_id}.txt. Chunked metadata (speaker turns, timestamps, labels, embeddings) lives in the Durable Object SQLite.

Why this is a chat superpower

Once calls are ingested, you can ask things like:

“What are the top 3 objections across my last 20 sales calls?”

Herald returns clusters with timestamped pull-quotes — and, if you have Linear connected, it cross-references each cluster against your open issues to flag complaints that have no engineering response.

“Why did Bloomflow walk?”

Herald finds the Bloomflow call, pulls the stretch with the objection and next_steps labels, and tells you.

Quality notes

• Transcripts longer than ~90 minutes are chunked aggressively. The summary compresses hard — treat it as a table of contents, not a replacement for the transcript.
• Single-speaker audio (you on a Loom) still works but Herald can’t name sides — it treats the whole thing as one voice.
• Foreign-language calls ingest but accuracy drops. Herald’s classifiers are English-tuned today.

What’s coming

• Fathom OAuth — first OAuth source. Target: weeks 5-6 after your account is active.
• Fireflies OAuth — after Fathom.
• Zoom OAuth — last, because the cloud-recording filtering is a multi-week project we’d rather defer.

Paste + forward covers most of the value with none of the integration debt.