Agent Phone Call Debrief
Purpose
Turn raw post-call inputs (transcript, agent summary, user notes, or “the call just ended”) into a structured debrief — outcome vs goal, commitments, risks, and next actions. Built for AI agent phone platforms (PollyReach, Agent Phone Call, OpenPhone, Twilio voice agents).
Pairs with phone-call-prep-brief (pre-call). This skill handles after-call.
Does not place calls, record calls, or access telephony APIs.
When to use
Use when:
- An AI agent finished an outbound or inbound phone call
- The user pastes a transcript or agent-generated call summary
- Booking, support, or scheduling calls need follow-up tasks
- The user asks “how did the call go?” or wants CRM notes
Safety and boundaries
Do not invent facts not present in the provided transcript/summary.
Do not request passwords, full payment card numbers, or government IDs.
Sensitive calls (medical, legal, financial disputes): flag for human review; do not provide legal/medical advice.
Recording: If transcript provenance is unclear, note that fact-checking may be needed.
Mark uncertain extractions with [unclear] or [verify].
Required inputs
Ask only what’s missing:
- Call direction — outbound, inbound, or callback.
- Original goal — from prep brief or one sentence (what success meant).
- Call artifact — transcript, bullet summary, or user recap (required).
- Counterparty — who was called or who called (role/org).
- Outcome signal — user’s gut: success / partial / failed / unknown.
Workflow
- Parse artifact for: stated outcome, commitments, dates/times, names, reference numbers, objections.
- Compare against original goal → verdict (met / partial / missed).
- Output debrief (format below).
- List follow-ups with owner (user / agent / human) and due hints.
- If verdict is missed or partial, suggest one retry angle (not a full re-call script).
Output format
Call debrief — {counterparty or role}
| Field | Value |
|---|---|
| Direction | outbound / inbound / callback |
| Goal | … |
| Verdict | met / partial / missed |
| Duration | if known |
| Reference #s | confirmation, ticket, reservation ID |
What happened (3–5 bullets)
Factual, sourced from transcript. No fluff.
Commitments made
| Who | Commitment | When | Confidence |
|---|---|---|---|
| Agent | … | … | high / verify |
| Other party | … | … | high / verify |
Risks & escalations
- Promises that exceed authority
- Hostile/abusive interaction
- Compliance issues (recording consent, PII spoken aloud)
- Items marked
[verify]
Follow-up tasks
- Task — owner — due (relative or date)
CRM / ticket note (copy-paste)
≤120 words, third person, past tense.
Retry recommendation (if partial/missed)
One paragraph: what to change before a second attempt.
Quality bar
- Grounded — every commitment traces to transcript text.
- Actionable — follow-ups have owners, not “monitor situation.”
- Honest verdict — partial when only some must-cover items landed.
- Concise CRM note — scannable in a ticket system.
Examples
Good commitment row: Agent — “Table for 4 Tuesday 7pm under Taboada” — Tue 7pm — high (confirmed twice in transcript).
Bad commitment row: Agent — “Reservation handled” — soon — high (too vague).
Good follow-up: [ ] Email restaurant confirmation screenshot to user — agent — within 1h
Bad follow-up: [ ] Follow up on reservation