Goal
Train a chatbot that answers routine questions accurately from your content, refuses to invent policies, and escalates cleanly. This guide is the process we recommend inside Chatoo.
Step 1 — Choose source content deliberately
Include:
- Current service menus and “what’s included”
- Hours, locations, parking, arrival instructions
- Booking, cancellation, and deposit rules
- Service-area definitions
- FAQs your team already answers weekly
Exclude or rewrite:
- Outdated PDFs and old landing pages
- Marketing pages that contradict each other
- Prices you are not willing to honor in writing
- Anything legal/medical the bot must never assert
Assign an owner who updates sources when the business changes.
Step 2 — Remove conflicts before indexing
If one page says you open at 9 and another says 10, the bot will eventually pick the wrong one. Deduplicate. Prefer a single canonical FAQ.
Step 3 — Define escalation rules
Write rules for:
- Refunds and charge disputes
- Abuse / safety issues
- Requests for guarantees you do not publish
- “Speak to a person” / named staff requests
- Low-confidence answers
The correct behaviour is clarify → escalate with transcript — not guess.
Step 4 — Test with a fixed question set
Create 25 real questions from past emails/calls. For each:
- Expected answer source
- Pass / fail
- Should escalate?
Do not launch until the high-traffic questions pass.
Step 5 — Monitor the first two weeks
Review:
- Repeated escalations on the same topic → add a source
- Confident wrong answers → remove or fix the bad source
- Customers asking for bookings → tighten booking fields
Step 6 — Improve without “just adding more pages”
More content is not better training. Better training is fewer, clearer, owned sources. When you change a price or hour, update the source the same day.
Checklist
- [ ] Owner named
- [ ] Conflicting pages removed
- [ ] Escalation paths tested
- [ ] 25-question scorecard saved
- [ ] Weekly transcript review scheduled
Related: AI chatbot feature · Customer support solution