Once Taskra is replying confidently, the next step is shaping what it can and can't say. That's what Guardrails and Custom Rules are for.
"Never say" phrases
Open AI Training → Guardrails and add phrases the AI must never produce. Examples that work well:
- "We guarantee delivery in 24 hours" — if you don't actually guarantee that
- "You'll get a full refund" — when refund eligibility depends on conditions
- "This is our cheapest product" — pricing changes, the AI shouldn't lock you in
- Competitor brand names — keep replies on-brand
Confidence threshold
Found in Settings → AI Behaviour. The AI only sends a reply on autopilot when its confidence is above this number. Lower it for more coverage; raise it for more safety.
- 85–90% — Conservative. Most replies flagged for your review. Use when learning a new voice.
- 75–85% — Default. Catches edge cases, lets routine answers through.
- 60–75% — Aggressive. High coverage but more variance — only use after weeks of clean data.
Custom escalation rules
Settings → AI Behaviour lists the built-in escalation triggers (complaints, refund requests, negative sentiment + low confidence). You can extend these:
- 1Order value > £200 → always escalate. Great for high-AOV brands.
- 2Keyword "lawyer" or "Trustpilot" in the message → always escalate.
- 3New customer (first message ever) → always escalate for the first 24h after launch.
Custom rules are evaluated before confidence — even a 99% confidence reply gets flagged if a rule matches.
Per-store tuning
If you run multiple stores, each one has its own threshold and rule set. The Stores → Settings link opens that store's AI Training so you can tune independently.
Auditing what the AI did
/inbox keeps every reply for the retention period (30 days on Starter, 90 days on Growth, 1 year on Scale). Each reply shows the confidence score and which rule fired — useful for debugging unexpected behaviour.
When to escalate everything for a while
After a big product launch or a return-policy change, flip Autopilot off for 48 hours. Every reply will queue for your review — you can spot whether the AI has picked up the change before it sends thousands of replies based on outdated knowledge.