AI Inbox
Triage
Automatically categorise, prioritise and surface the right messages for each user — so no one misses a critical update buried in project noise.
01
Problem
Context · click to edit
Root cause · click to edit
02
Requirements
| ID | Requirement | Priority | Notes |
|---|---|---|---|
| R-01 | Classify every inbound message into: Urgent / Follow-up / Mention / FYI | MUST | Core feature |
| R-02 | Classification latency < 800ms p95 | MUST | Inline, not async |
| R-03 | Users can correct a label — feedback loop to improve model | SHOULD | Phase 2 |
| R-04 | Per-user sensitivity settings (quiet hours, channel overrides) | SHOULD | Settings panel |
| R-05 | Digest view: daily summary of FYI messages | NICE | Email or in-app |
03
Risks
HIGH
False negatives on urgent messages
If the classifier misses an urgent message, user trust collapses immediately. Mitigation: conservative threshold — err toward urgent. Measure precision/recall weekly.
MED
Latency on large context windows
Long threads passed to classifier may exceed 800ms target. Mitigation: summarise thread context to last 10 messages before classification.
MED
User distrust of AI labelling
Some users resist AI-driven sorting. Mitigation: ship with an easy one-click disable. Make label logic visible on hover.
04
Open Decisions
DECISION NEEDED · Classification model
Do we use our own fine-tuned classifier (lower latency, higher infra cost) or claude-haiku-4-5 via API (faster to ship, per-token cost at scale)?
DECISION NEEDED · Rollout strategy
Opt-in beta for power users (lower risk, slower feedback) vs. default-on for all new signups (faster learning, higher support load)?
05
Design Ideas
✦Click "Generate ideas" to get 3 design directions from Claude.