Auto-send governance
Rules governing Pro-tier conditional auto-send in Handlet.
Auto-send governance
Handlet Pro may opt into conditional auto-send for high-confidence replies when per-intent toggles and guardrails pass. Auto-send is disabled by default.
Configurable intents (Pro)
Operators enable automation per intent:
| Intent | Requirements |
|---|---|
| Follow-ups | Training-ready toggle, master switch on, transparency ack, confidence ≥85% |
| Quote requests | Training-ready toggle + pricing guardrails allow a firm price |
| Booking requests | Training-ready toggle + required fields complete |
| New leads | Training-ready toggle + guardrails pass |
| General enquiries | Training-ready toggle + guardrails pass |
| Review replies (publish) | Training-ready toggle; ≤2★ reviews always manual |
| Complaints | Training-ready toggle + separate complaint acknowledgement; high severity always manual |
Bulk configuration
- Enable all eligible / Disable all — bulk intent controls without changing master switch state. Enable only affects training-ready intents.
- Automate ready replies — explicit preset: master on + training-ready intents enabled in one save. Complaints require separate acknowledgement. Does not bypass Stage 2 training minimums.
Platform-vetoed intents (never auto-sent)
Regardless of settings, Handlet never auto-sends:
- Risk, noise, unknown, or unsupported intents
- Personal-suppressed conversations
- Sends when credentials are degraded or mandatory disclosures fail
Quote, booking, and complaint auto-send remain blocked when their guardrails fail — even if the intent toggle is on.
Confidence threshold
Accounts must set a minimum confidence threshold at or above the platform safety floor (85%). Thresholds below the floor are rejected.
Stage 2 calibration minimum (hard law)
Before auto-send unlocks per intent, operators must accumulate validated training evidence:
- 30 validated messages (global default) or 50 for complaints
- Counting starts 2026-06-23 — no retroactive credit for earlier sends
- Minimums are platform constants, not operator-editable in v1
- Settings show per-intent training progress by default for Pro workspaces
- Per-intent toggles remain locked until training is complete for that intent; send-time eligibility also blocks fail-closed if an untrained stored toggle or forged request reaches execution
Validated events: approve and send, edit before send, correctness confirmed, correctness rejected. Dismissed correctness feedback does not count.
See Calibration and confidence bar and Conditional autonomy policy for full gating conditions.
Transparency and control
Enabling auto-send requires acknowledgement of AI transparency terms. Configuration changes are audited with actor and timestamp. Disable auto-send instantly from Channels & Tone settings.
Legal and compliance
Auto-send remains confidence-governed, auditable, and subject to Conditional autonomy policy. End-customers retain a human-review path — see Human intervention and contesting automated replies.