First: figure out what kind of "wrong" it was
Open the call in Calls → Logs, find the moment, and ask yourself which bucket it falls into:
| Type of mistake | Example | Fix path |
|---|---|---|
| Under-informed | "I don't think we accept that insurance" when you do | Add the real info to your Knowledge Base |
| Stale data | Wrong hours, wrong holiday closure, wrong provider name | Fix the underlying setting (Locations → Hours, Team, Services) — the AI reads those, not a prompt |
| Wrong scope | Quoted a price you don't publicly quote, or commented on a clinical question | Add a Knowledge Base entry stating the rule ("we don't quote prices" or "clinical questions → message + callback") |
| Hallucination | Invented a service, a fee, or a policy that doesn't exist | Same as wrong-scope — give it the actual policy in writing. The AI prefers knowledge-base text over guessing. |
| Persona / tone | Sounded too casual / too formal | Settings → AI Voice → Conversation Tone |
The fix loop, in 4 steps
- Find it. Calls → Logs → open the call → read the transcript. Copy the exact wrong phrase.
- Categorize it using the table above.
- Fix the source of truth. The AI doesn't have a "correction" knob — it has inputs. Update the right input:
- Wrong facts about your practice → Settings → AI Voice → Knowledge Base (you can paste text, upload a PDF, or crawl your website)
- Wrong hours/services/team → fix the actual setting (Locations, Services, Team)
- Wrong tone → Settings → AI Voice → Personality & Behavior
- Wrong greeting → Settings → AI Voice → Custom Greeting
- Test it. Call your number and try the scenario that broke. If the new answer is right, you're done. If not, the knowledge entry probably needs to be more specific — see below.
Writing knowledge-base entries that actually fix problems
Vague entries don't work. The AI reads them like a person would.
Weak — "We have flexible scheduling." Strong — "Same-day appointments are available Monday–Thursday between 9 AM and 11 AM if a slot is open. Call back the same morning or check the booking link."
Weak — "Don't talk about prices." Strong — "Do not quote prices over the phone. If a caller asks about pricing, say: 'Pricing depends on what you need — I can have someone email you a quote, or you can book a consultation. Which would you prefer?'"
Concrete, actionable, and ideally including the exact phrase you want the AI to say.
When the AI is making things up that you never told it
That's a knowledge-gap symptom. The AI's general training contains a lot of medical/business-y patterns, and when it doesn't have your specific policy it will sometimes fill in with a plausible-sounding default.
The fix is to take away the gap: add the policy. Even a one-line entry that says "For [topic X], take a message and have the office manager call back — do not answer questions about X directly" is enough to stop the hallucination on that topic.
See Train on your FAQs for the full Knowledge Base flow (paste, upload, or crawl).
If a mistake harmed a patient
Open a support ticket immediately with the call ID. We will review the transcript and the recording, and if the issue points to a model-level problem (not a knowledge gap) we'll escalate to engineering.