The email recap is the meeting feature people ask about most. It sounds simple: agent joins meeting, agent listens, agent emails the recap. The simple version works. The version that scales (multiple meetings a day, multiple participants, recaps that earn the inbox space they take) requires a few more layers of decision-making, all of which are interesting.
This post is what a real recap looks like, how the agent decides what to include, how action items get split by owner, and the deliverability problem most products skip until users complain about deliverability. We will keep it concrete.
What a real recap looks like
A representative one, for a Tuesday Q3 retro between four people.
Subject: Recap - Q3 Retro - May 6 - 4 people
DECISIONS
- Velocity number to be redrafted to include scope-change %
(Sara owns; due Monday).
- October launch is the focus; everything not on the launch
path is deprioritized for the next 6 weeks.
ACTION ITEMS
Sara:
- Redraft velocity definition (Mon).
- Schedule walk-through with leadership before next retro.
Mark:
- Pull the customer-call data for the launch dashboard (Wed).
- Sync with design on the loading-state question.
Dana:
- Review the H2 OKR doc by Friday.
You:
- Reply to Sara on the velocity question by Monday EOD.
- Pull the threading sequence question from your H2 doc and
include in the launch dashboard.
FOLLOW-UPS (carried from prior retros)
- The Stripe migration question is still open from the April
retro; Dana to check status.
DRAFTS
- Reply to Sara on velocity, draft attached.The recap is ~120 words. It loads in under a minute. The action items are split by owner. The follow-ups carried forward. A draft reply is attached for the immediate next thing.
This is what "earned its keep" means for an email recap. If it took 5 minutes to read, you would not read it. At 60 seconds, you do.
How the agent decides what to include
Three layers of decision-making.
- 1Transcript ingestionAgent reads the full meeting transcript via the meeting bot or browser invisible mode.
- 2Decision detectionAnything that sounds like a commitment or a decision is flagged.
- 3Owner attributionFor each action item, the agent determines who owns it (the speaker, the named target, or the implicit "team").
- 4CompressionThe full transcript becomes the 5-8 bullet structure. Filler removed; specifics kept.
- 5Draft attachmentsFor action items that have an obvious follow-up (a reply email, a calendar entry), the agent drafts it.
The interesting one is owner attribution. The naive version assumes "if Sara said it, Sara owns it." The reality is more complex: Sara might be assigning to Mark; the team might be agreeing without anyone explicitly committing; the user might be the implicit owner of anything-not-otherwise-attributed. The agent's accuracy on this is around 90% in practice; corrections are easy ("that one is mine, not Mark's") and the agent updates the recap.
The deliverability problem
The email recap has to land in your inbox to be useful. This sounds obvious; it requires real engineering.
Sender identity. The recap email is sent from the agent's domain by default. For users who want recaps to appear from their own address, the agent supports authenticated sending via Gmail or Microsoft 365 OAuth. The "from" matches your real address.
Spam scoring. A high-volume sender needs to maintain sender reputation. We do the work: standard email authentication, sender warm-up, bounce handling, complaint handling. Users do not see this; they see the recap landing.
Subject-line discipline. The subject is consistent (Recap - <meeting title> - <date>) so users can train Gmail to recognize and prioritize. Filtering becomes possible.
Attachment handling. PDFs, transcripts, draft replies, when attached, pass through the same authenticated path. They land cleanly.
What the agent does with action items
Three options, configurable per action item.
Auto-fire (low-stakes). Some action items are confidently auto-handled. Sending a thank-you to the recap recipients. Adding a meeting to the calendar. The agent fires these without asking.
Approval-queue (medium-stakes). Most are queued for your approval. The recap email lists them with numbers; you reply "do 1, 2, 4" to fire those; the agent does.
Always-ask (high-stakes). Some actions never auto-fire regardless. Sending an email to a senior person. Moving money. Making a commitment on someone else's behalf. These wait for explicit per-item approval.
The classification is automatic with manual override. You can change which category a specific kind of action goes in by editing your skill: "always treat email-to-leadership as always-ask, never auto-fire."
Multi-channel rendering
The recap goes to email by default, but the same content renders in your text-message thread with the agent (a more compressed version) and in your business CRM if integrated (the action items become tasks). The "one recap, many surfaces" pattern is the same as elsewhere in the agent.
Some users prefer recaps via WhatsApp instead of email. Configurable. Some users want the recap split (decisions to email, action items to Slack). Configurable.
- Meeting ends. You go back to your inbox.
- 10 minutes later, you write rough notes to send.
- Action items get lost. Some never go anywhere.
- Followers up are missed for 1-3 weeks.
- Meeting ends. Recap lands in your inbox in 4 minutes.
- Action items split by owner. Drafts attached for the obvious follow-ups.
- Approval queue lets you fire the right ones with one reply.
- Carry-forwards from past retros surface automatically.
Frequently asked
Does the agent join the meeting as a participant?
What about meetings the agent did not join?
Can multiple participants get the recap?
How private are the transcripts?
Can I customize the recap format?
Try it
Let the agent join your next meeting. Read the recap when it lands. The first one teaches you the shape; the second one teaches you what to tune.
