Meeting and Communication Agents: Summaries, Follow-Ups, and Drafts
Meeting and communication agents turn transcripts, threads, and calendars into structured summaries, action items, and draft messages.
They save hours of note-taking when grounded in the source text and gated before send.
Use this cheatsheet to scope a v1 that people trust.
How to Use This List
- Start with A to pick the job-to-be-done (summary vs follow-up vs draft).
- Apply B as required output fields in your schema.
- Use C before enabling any send/write tool.
- Keep D next to privacy reviews.
A - Job Cheatsheet
| Job | Input | Output | Autonomy default |
|---|---|---|---|
| Recap | Transcript / recording notes | Agenda-aligned summary | Auto OK for internal |
| Action extract | Transcript + attendees | Owner, task, due date, quote/timestamp | Auto draft; human assigns |
| Decision log | Transcript | Decision, context, dissent | Auto draft |
| Follow-up email | Recap + actions | Outbound draft | Human send |
| Thread triage | Email/chat thread | Priority, suggested reply | Human send |
| CRM note | Call transcript | Structured fields | Human approve write |
| Standup digest | Multiple updates | Team brief | Auto internal channel |
| Customer call coach | Sales call | Tips + risk flags | Private to rep/manager |
If you only ship one job first, ship action extract + recap before any auto-send.
B - Required Fields (Do Not Ship Freeform Only)
| Artifact | Must include | Nice to have |
|---|---|---|
| Meeting summary | Purpose, attendees, key points, decisions, open questions | Links to docs mentioned |
| Action item | Verb + object, owner, due date or "TBD", source timestamp/quote | Priority, dependency |
| Decision | Statement, who decided, date, alternatives considered | Revisit date |
| Follow-up draft | To/cc suggestion, subject, body, linked actions | Tone preset |
| CRM update | Account id, fields changed, evidence quotes | Next step date |
| Risk flag | Claim + transcript span | Severity score |
Reject outputs that list actions without owners or quotes that cannot be found in the transcript window you stored.
C - Safety and Send Checklist
- Consent and policy allow recording/transcription in this region and context.
- Access control: only attendees (or permitted roles) can read the recap.
- PII redaction path for pastes into external models when required.
- No auto-send to external recipients in v1.
- No silent CRM overwrite; show diffs pre-commit.
- Retention rules for audio, transcript, and summaries are defined.
- Speaker attribution errors are acknowledged (diarization is imperfect).
- Secret meetings (HR, legal, M&A) are on a deny or special-review list.
- User can edit every draft before it leaves the system.
- Audit log of generations and sends exists.
D - Quality Rubric (Score 1-5 Each)
| Criterion | 5 looks like | 1 looks like |
|---|---|---|
| Faithfulness | Every bullet traceable to transcript | Invented attendees or decisions |
| Actionability | Clear owners and next verbs | "Follow up on stuff" |
| Compression | 10-20% length, keeps decisions | Either transcript dump or empty |
| Attribution | Correct speakers on quotes | Swapped speakers |
| Tone fit | Draft matches audience | Casual slang to enterprise customer |
| Completeness | Open questions listed | Silent on unresolved conflicts |
Run the rubric on a golden set of 10 meetings before wide rollout.
E - Tool Allowlist Patterns
| Tool | Use | Gate |
|---|---|---|
get_transcript | Source of truth | ACL by meeting id |
get_calendar_meta | Attendees, title, time | Read-only |
search_prior_notes | Continuity across series | Same ACL family |
draft_email / draft_chat | Create draft only | No send scope |
send_email | Outbound | Explicit human click / approval token |
update_crm | Field writes | Diff + approve |
create_tasks | Tickets/todos | Project allowlist |
Prefer two tools (draft_* and send_*) over one "communicate" supertool that hides side effects.
F - Pipeline Stages
- Ingest - transcript + metadata; store immutably.
- Clean - filler removal optional; keep timestamps.
- Structure - summary, decisions, actions via schema.
- Ground - verify quotes/timestamps still match source spans.
- Draft - optional communications in a separate step.
- Review - human edits in UI.
- Act - send/write only after approval.
- Learn - optional: store user edits to improve templates (with privacy care).
Skipping stage 4 is how confident fiction enters CRM systems.
G - Common Failure Modes
| Failure | Symptom | Mitigation |
|---|---|---|
| Hallucinated decisions | "We chose vendor B" never said | Require quote spans; refuse if missing |
| Owner vacuum | Actions with no names | Default owner = speaker or "UNASSIGNED" |
| Infinite politeness | Drafts too vague to act on | Template with explicit ask + date |
| Context bleed | Prior meeting mixed into recap | Scope retrieval to meeting id first |
| Over-notification | Every recap pings a huge channel | Audience rules by meeting type |
| Diarization mess | Wrong speaker on commitments | Surface confidence; allow rename UI |
FAQs
Is a meeting summarizer an agent?
A single-shot summary can be a pipeline. It becomes agentic when it chooses tools (CRM lookup, prior notes, task systems) across turns based on content.
How long should summaries be?
Default to one screen plus an action table. Offer "detailed" as a second artifact, not the default email body.
Should the agent join live meetings?
Live bots add consent and product complexity. Many teams start with post-meeting transcript processing only.
How do we handle multiple languages?
Detect language, summarize in the meeting language by default, and only translate as an explicit second step with review.
What about highly technical discussions?
Preserve domain terms; do not "simplify" away API names or metric definitions. Link to internal docs when mentioned.
Can we auto-create Jira tickets?
Yes after precision is high on action extraction. Start with draft tickets in a staging project.
How do communication agents differ from personal email agents?
Work agents center on shared meetings, CRM, and team ACL. Personal agents center on one user's inbox preferences and devices.
What model tier is enough?
Long transcripts often need larger context windows or chunk-then-merge strategies. Verify limits at build; map-reduce summaries carefully to keep actions.
How do we evaluate follow-up drafts?
Score: correct recipients, correct commitments, no new promises, tone, and presence of scheduling ask when needed.
What if the transcript is empty or tiny?
Return a failure artifact, not a creative story. "Insufficient transcript" is a valid product outcome.
Should action items include due dates the agent invents?
Only if spoken or present in calendar. Otherwise mark due: TBD and ask the owner.
How do we prevent leaking confidential meetings?
Strict ACL mirroring calendar permissions, optional meeting-type classifiers, and admin denylists for sensitive series.
Related
- Business & Research Use Cases Basics - structured outputs and review packets
- How Knowledge-Work Agents Differ from Coding Agents - human review defaults
- Internal Knowledge-Base Agents with RAG Grounding - grounding prior notes and policies
- Business & Research Use Cases Best Practices - accuracy habits for knowledge work
- Email and Calendar Agents - personal-side cousin
Stack versions: Pins from the category manifest (verify at build): OpenRouter (~315+ models, July 2026 pricing/fees); LangGraph 1.0+; CrewAI 1.14+; Microsoft Agent Framework 1.0; Vercel AI SDK 6; Pydantic AI (latest); LlamaIndex (latest); OpenAI Agents SDK (latest + MCP); MCP (Linux Foundation governance); A2A (HTTP+SSE+JSON-RPC 2.0); Solana
@solana/web3.js+@solana/spl-token.