Write a PRD

PRD: AI Meeting Summaries with Decisions and Action Items

Status: Draft for engineering and design leads Product: Huddle

1. Problem

Huddle records and transcribes meetings with speaker labels, but attendees still write their own notes afterward. Notes are late, inconsistent, or never written, so decisions get lost and action items have no clear owner. We will generate a summary after each meeting that lists decisions and proposes action items with owners.

2. Goals and non-goals

Goals - Produce a summary within 5 minutes of meeting end: decisions, action items (task, owner, due date if stated), and open questions. - Never assign an action item to someone who was not in the meeting (hard constraint, see §4). - Keep the host's review to under 60 seconds for a typical meeting.

Non-goals (v1) - Fully automatic emailing of action items to owners (see §3). - Cross-meeting task tracking or integrations with Jira, Asana, and similar tools. - Summaries for meetings without a transcript.

3. Decision on the Sales request

Sales asked for action items to be emailed automatically to owners the moment a meeting ends. We are not building that for v1. Instead, we will send the draft summary to the host at meeting end. The host can send it to owners with one click, and owners are emailed on send.

Reasons: - Speaker labels are wrong about 8% of the time. An automatic email would sometimes tell a real person they own a task they never agreed to. Because it goes out with no human check, the error becomes visible to a customer's colleagues. - Email is irreversible, so a wrong or non-attendee assignment cannot be quietly fixed afterward. - The underlying Sales need is that owners hear about their tasks quickly and nothing falls through the cracks. A host prompt at meeting end with one-click send meets that need at the cost of a few seconds.

Path to automation: Auto-send will be an opt-in, per-workspace setting, considered after launch once the gates in §9 are met. It would send only high-confidence, presence-verified items and leave everything else for host review. Engineering should build the send pipeline so that auto-send is later a policy change rather than a rebuild.

4. The hard constraint: who can own an action item

Rule: An action item's owner must be a participant who was actually in the meeting when the item was discussed. Anyone else is ineligible.

This rule cannot rely on the calendar invite, and it cannot rely on the model behaving well.

Attendance source. The eligible-owner roster is built from Huddle's session join/leave logs, not from the calendar invite. This handles both failure modes: - Invitees who never joined are not on the roster. - The roughly 15% of meetings with uninvited joiners include those people, matched to a user account when possible.

Presence at time of item. A person is eligible for an item only if they were in the session at the timestamp of the relevant utterance. Someone who left at minute 10 cannot own an item from minute 40.

Enforcement is structural, not prompt-based. 1. The extraction step receives the roster as a list of participant IDs and must return an owner ID from that list or `null`. 2. A deterministic validator runs after the model. It rejects any owner ID not on the roster or not present at the item's timestamp, and it converts that item to unassigned. 3. The UI and email layers accept only validated owner IDs. There is no free-text owner field on the model output path. 4. Names mentioned in the transcript who were not present ("have Dana look at it") produce an unassigned item. The text may keep the speaker's wording, but Dana gets no assignment and no email.

Shared rooms. A room device usually appears as one participant, so individuals in the room are not in the presence log. We cannot verify that a named person is in the room. Such items become unassigned by default, and the host can attest who was in the room during review (§6). That attestation counts as human-verified presence.

Uninvited joiners without an account (guests) can be eligible owners, shown by display name. They receive email only if the host supplies an address during review. We never guess addresses.

5. Handling speaker-label errors

The 8% label error rate is concentrated in crosstalk and shared rooms. Eligible-owner enforcement does not solve it, because a mislabel can still pick the wrong person who was present. Mitigations:

Confidence tiers (High/Medium/Low) are stored on each item, shown to the host in review, and used later to gate auto-send. Thresholds are set empirically during the evaluation in §9. The tiers themselves are a v1 requirement.

6. User experience

Host flow 1. The meeting ends and the host gets a notification: "Summary ready to review." 2. The review screen lists: - Decisions as short statements, each linked to the transcript moment. - Action items with the task, owner chip, due date if stated, and a confidence indicator. - Unassigned items grouped at the top of the actions list, so gaps are visible. 3. The host can edit text, reassign from a picker limited to eligible people, delete items, or add room attendees ("Who was in the room?"). Adding room attendees makes them eligible for items in the room segments. 4. Send emails each owner their own items plus the decisions. Unassigned items go to no one.

Design requirements - Every item has a "jump to source" link that plays the transcript segment. Reviewing should be verification, not re-listening. - The owner picker never offers non-attendees. If the host wants to assign to an absent person, show a separate option, "Send as a follow-up (not from the meeting)," which is a different item type with clearly different wording and no "you agreed to" language. Design to decide whether this ships in v1; the default is that it does not. - Low-confidence owners appear as suggestions ("Suggested: Sam") and never as assignments. - Email copy says "Action item from Meeting name, date", includes the source excerpt, and offers a "This isn't mine" link that notifies the host. This gives owners a correction path. - If the host does not review, send a reminder at 24 hours. Nothing is sent to owners without host action in v1.

7. Functional requirements

1. Roster service: builds the per-meeting roster with join/leave intervals from session logs and links to user accounts where possible.
2. Extraction: from the transcript and roster, produce decisions, action items, and unassigned items. Each action item has task text, owner ID or null, due date or null, source timestamps, and a confidence tier.
3. Validator: enforces §4 as code with unit tests, independent of the model, and logs every rejection.
4. Review UI: as described in §6, on web and desktop. Mobile is view-only for v1.
5. Delivery: per-owner emails on host send, with a record of who was emailed. Sends are idempotent, and edits after sending trigger an explicit "send update" action.
6. Audit log: stores for each item the model output, validator result, host edits, and send events.
7. Admin setting: workspace-level toggle for summaries, plus retention aligned with existing transcript retention.

8. Edge cases to specify and test

• Invitee never joined but is mentioned as owner.
• Uninvited joiner takes an action; the guest has no account.
• Person leaves and rejoins (multiple intervals).
• Two people share a first name; ownership must resolve by roster, and ambiguity means unassigned.
• Whole meeting in a shared room (all items default unassigned until the host attests).
• Multiple languages, or a transcript with large gaps.
• Nobody commits to anything: show "No action items detected" rather than inventing some.

9. Success metrics and launch gates

Launch gates (must all pass) - Zero non-attendee assignments across a labeled evaluation set of at least 500 meetings, deliberately oversampled for uninvited joiners, shared rooms, no-show invitees, and crosstalk. Any production occurrence afterward is a Sev-1 incident. - Among High-confidence items, wrong-attendee assignment rate of 3% or less (target, to be validated against the evaluation set; propose a stricter figure if data supports it). - Median summary delivery within 5 minutes of meeting end.

Post-launch metrics - Share of meetings where the host sends the summary within 1 hour. - Host edit rate per item, and owner "This isn't mine" rate. - Share of action items left unassigned. Some is healthy, but a high rate means the model is too timid.

Auto-send criteria (future): two consecutive months of High-tier precision above the agreed bar and no non-attendee incidents. Sales and Support are consulted before enabling.

10. Risks and open questions

• Sales disagreement. Sales may see the host step as friction. We should share §3 with them early and offer a pilot of auto-send with a few design-partner accounts after the gates are met.
• Hosts skipping review. If sending rates are low, the fix is better reminders and faster review, not removing the check.
• Presence data quality. Do our session logs reliably capture join/leave for all clients and dial-in users? Engineering to confirm in week 1.
• Guest privacy. Can we email or display guest names without consent, given customer policies? Legal to advise.
• Room attestation. Is host attestation adequate for the constraint, or should room participants confirm individually? Recommend host attestation for v1.
• Follow-up items for absent people. Ship in v1 or defer? Design and Product to decide.

11. Milestones

1. Weeks 1–2: roster service, presence data audit, evaluation set labeling begins.
2. Weeks 3–5: extraction, validator, confidence tiers, offline evaluation.
3. Weeks 4–7: review UI and email delivery, in parallel.
4. Weeks 8–9: internal dogfood, gate review.
5. Week 10: limited beta with design-partner customers, including Sales' accounts.

Huddle AI meeting summaries

Product requirements for engineering and design leads

Purpose and product decision

Huddle will generate a reviewable record of meeting decisions and actions from its transcripts. The host checks the draft, resolves ownership and publishes it. Approved actions can then be emailed to their owners.

Huddle serves 3,000 companies whose users currently write notes themselves. Speaker labels are wrong about 8% of the time, and roughly 15% of meetings include an uninvited participant. Calendar membership and transcript attribution therefore cannot establish attendance or ownership.

The MVP will not fulfil Sales’ request for immediate, unreviewed owner emails. It will generate drafts automatically after transcription and send emails automatically after explicit publication. Faster delivery does not justify assigning work to absent people. All requirements and numerical targets below are proposed product decisions.

Scope and intended outcome

Primary users are meeting hosts reviewing records and attendees receiving actions. Success means less time producing dependable notes, with no assignments to non-attendees.

MVP includes English transcripts, decisions, action extraction, attendance verification, host review, publication and owner emails. Exclude live summaries, recurring reminders, task-system integrations and automatic reassignment. Use existing meeting access and retention controls.

An action has one verified owner or remains unassigned. Split genuinely separate responsibilities into separate actions. Do not infer deadlines, commitments or decisions from tentative discussion.

Attendance is a release dependency

Engineering must first confirm whether Huddle retains authenticated participant identities and join/leave events. The supplied calendar list is insufficient. If these records are unavailable, build attendance instrumentation before enabling assignments; summaries may still launch with unassigned actions.

An eligible owner must have a stable person identifier linked to an authenticated, human participant session with a recorded join event for this meeting occurrence. Calendar invitees who never joined are ineligible. Bots and room devices are ineligible. Deduplicate reconnects by identity. Someone who left early remains eligible, although a later discussion does not prove they accepted work.

Uninvited users are eligible when their authenticated join is recorded. Guest display names, shared-room speaker labels and email addresses typed by a reviewer are not identity evidence. Unverified guests and people represented only by a room device remain ineligible in MVP; explain how they can join individually with verified identity for future meetings.

This deliberately sacrifices assignment coverage. The invariant is enforceable against verified participation records, not proof of physical human presence behind an account. If literal physical presence is required, that remains an unresolved product constraint and assignments must not launch under a stronger guarantee.

Ownership and extraction rules

The model extracts structured candidates with supporting transcript spans. Each decision contains its outcome and evidence timestamps. Each action contains a task, supporting spans, optional explicit due date and a nullable candidate participant identifier. Preserve the original date wording alongside any normalised date, using the meeting timezone; ambiguous dates stay unset.

Only explicit commitments or clearly agreed assignments qualify as actions. Suggestions, questions and rejected proposals do not. Where later discussion reverses a decision, present the final supported outcome. If the resolution is unclear, flag it for review instead of declaring a decision.

A model candidate is internal metadata, never an assigned owner. Server validation rejects identifiers outside the eligible attendance set before anything is displayed. The review UI starts every owner field as “Unassigned”; reviewers choose an eligible attendee after checking evidence. Do not preselect names from speaker labels, including for “I’ll do it”.

Render ownership only from validated structured fields. Generated task and decision wording must not embed owner claims that bypass these controls. Unsupported person-specific claims are withheld for review. A request that absent Alex should do something can become an unassigned “Confirm responsibility for…” follow-up only when that follow-up was actually agreed; otherwise omit it as an action.

Review and publication experience

The meeting page shows “Preparing summary” after the call, then “Draft needs review”. The host is the default reviewer; existing authorised meeting editors may also review. If no reviewer is available, retain the draft without sending owner emails.

Display separate Decisions and Actions sections. Each item links to its transcript passage and recording timestamp. Offer edit, delete and add controls. Actions show task, owner and due date. An attendance-only owner picker explains why invitees or unresolved guests cannot be selected. Provide a visible “Needs owner” state rather than silently dropping useful work.

Reviewers must confirm each chosen owner and can publish with unresolved actions. The publish dialogue shows recipients and unresolved-action count, with an explicit “Publish and email owners” control. Actions without owners remain visible but produce no email. Keyboard navigation, labelled controls and screen-reader status announcements are required.

Published records show their revision and approval time. Subsequent edits create a new revision. Ownership changes require the same validation and explicit publication, then notify affected owners with a correction. Never silently replace a published record with regenerated content.

Engineering contract and failure behaviour

Use meeting occurrence IDs, not recurring calendar series IDs. Store transcript version, attendance evidence version, model/prompt version, summary revision, item IDs, evidence spans, owner ID, approval actor/time and delivery status. Keep internal candidates separate from approved owner fields.

Process states are waiting for transcript, generating, needs review, published and failed. Trigger generation only after transcript finalisation. Transcript changes invalidate unpublished approvals and create a fresh draft; published revisions remain intact until a replacement is approved.

Enforce owner eligibility server-side on draft writes, publication, exports and notification dispatch. Manual edits and direct API calls cannot bypass it. If attendance evidence is missing, contradictory or unavailable, block ownership and sending rather than falling back to calendar data. Record rejection reasons for support without exposing transcript content in routine logs.

Use transactional publication and an outbox. Delivery keys combine meeting, published revision and recipient so retries do not duplicate emails. Email only approved actions owned by that recipient, with a permission-checked meeting link. Resolve addresses from verified identities, never model output. Block delivery if permissions or identity validation fail; show the reviewer the failure. Retry transient failures with backoff for up to 24 hours, then offer manual retry.

Treat transcripts as untrusted data: spoken instructions cannot change permissions, choose recipients or trigger tools. Generation has no email capability. Preserve tenant isolation and apply meeting deletion and retention policies to summaries and queued notifications.

Acceptance scenarios

1. Invited Pat never joins but is named in an action: Pat is absent from the picker; API assignment is rejected; no owner email is sent.

1. Uninvited authenticated Lee joins: Lee is eligible and can receive an action after reviewer confirmation and publication.

1. Several people join through one room device: no individual becomes eligible from its label; actions remain unassigned unless independent authenticated participation exists.

1. Overlapping speech mislabels a commitment: no owner is preselected; the reviewer can inspect evidence and choose a verified attendee or leave it unassigned.

1. No decisions or actions are supported: show explicit empty states rather than inventing content. Failed or incomplete transcription shows a retry state and sends nothing.

1. A forged owner ID, missing attendance record or duplicate publish event cannot create an invalid assignment or duplicate email. Revoked access blocks queued delivery.

Evaluation and release

Build a consented, de-identified evaluation set of at least 200 meetings, deliberately including overlap, shared rooms, absent invitees, uninvited guests, duplicate names, reversals and ambiguous commitments. Human annotators label attendance, final decisions, actions, ownership and evidence; adjudicate disagreements independently. Report each difficult slice separately.

Release gates: zero non-attendee assignments or deliveries across end-to-end and adversarial tests; at least 95% precision and 85% recall for supported decisions and actions, measured separately. Zero observed violations is necessary but insufficient: eligibility enforcement must also pass API and notification-path tests. Model confidence alone never permits assignment.

Pilot with ten opt-in companies before staged rollout. Target draft availability within two minutes of final transcript for 95% of meetings up to 60 minutes, and a 50% reduction in median note-preparation time against pilot baseline. Track owner corrections, unassigned-action rate, publication rate and delivery failures. Any non-attendee assignment disables assignments and sending pending investigation. Expand only after two weeks without a critical violation and acceptable review effort; do not reduce abstention merely to improve coverage.