How would you design AI-powered meeting assistants?

### Signal to interviewer

I can design meeting AI for execution quality, not transcription novelty, by optimizing decision capture and follow-through.

### Clarify

I would clarify meeting types, participant roles, and compliance constraints around recording and retention.

### Approach

Apply a workflow surface model: pre-meeting prep, in-meeting capture, and post-meeting execution. Each surface has explicit outputs and handoff points.

### Metrics & instrumentation

Primary metric: completion rate of AI-captured action items. Secondary metrics: summary edit rate, prep-time reduction, and follow-up turnaround. Guardrails: missed critical decisions, privacy complaints, and incorrect owner assignment.

### Tradeoffs

Broader capture increases recall but can dilute signal quality. Aggressive automation speeds output but can mis-assign accountability.

### Risks & mitigations

Risk: overconfident summaries; mitigate with confidence tags and edits. Risk: privacy concerns; mitigate with clear consent and retention controls. Risk: low downstream adoption; mitigate with deep task-system integrations.

### Example

In product standups, the assistant turns spoken blockers into assigned Jira follow-ups and flags unresolved decisions for leadership review.

### 90-second version

Design meeting AI across the full workflow. Prioritize decision fidelity and action execution, measure downstream completion, and keep controls transparent so teams trust and use the output.