Troubleshoot: AI saying the wrong thing
Tighten your AI agent's behaviour — system prompt, knowledge base, qualifying questions.
01
Step 1 — Read the call transcript
- Calls → click the offending call → read the full transcript top to bottom.
- Identify the EXACT phrase the AI said wrong.
02
Step 2 — Identify root cause
- **Wrong fact**: knowledge base missing or outdated. Add to /knowledge-base.
- **Wrong tone**: system prompt needs explicit guardrail (e.g. 'never quote prices').
- **Off-script**: temperature too high (try 0.3 instead of 0.7).
- **Hallucination**: the AI invented something. Tighten system prompt with 'NEVER answer questions about X — say
'.
03
Step 3 — Update the system prompt
- At /assistants → Edit → Advanced → System prompt override.
- Add explicit DO NOT instructions for the failure mode.
- Save, test, re-listen.
04
Step 4 — Test before going live
- Use the 'Test call me now' button on the agent edit page.
- Make 3 test calls covering different scenarios.
- Listen end-to-end before flipping back to production.
05
Step 5 — Lock in with rebuttals
- For repeatable objections (e.g. 'too expensive', 'not interested'), add explicit rebuttal entries.
- Trigger phrase + AI response — fully scripted, no improvisation.
Related articles
Building a voice agent
Pick a template, pick a voice, write a script. Live in 10 minutes.
Writing a script that converts
Practical advice from the calls that worked across our pilot tenants.
Testing an agent before going live
How to QA a new agent without burning real leads.
Troubleshoot: voice quality / latency
When the AI sounds robotic, choppy, or has long silences. Fix steps.
Still stuck?
Email hello@voxreach.com.au — we reply within 2 business hours during AEST hours.
Open a ticket