AI Employee can genuinely answer half an agency's inbound conversation volume. It can also confidently invent a discount that does not exist. The difference is entirely in setup: what you feed it, what you fence it from and where you route the handoff. Here is the setup that works in production.
Feed it facts, not vibes
The knowledge base is the product. Write it as answers to real questions from your last hundred conversations, not as marketing copy. Include prices, hours, service areas and the top ten objections with their honest responses.
Anything not in the knowledge base is something the model may improvise. If you would not want it improvised, write it down explicitly.
Fence the failure modes
Cap the conversation turns before human handoff. Route anything containing refund, cancel, legal or angry sentiment straight to a person. Disable the AI on conversations a human has already joined, or the two will talk over each other.
Test adversarially before launch: ask for discounts, ask it to contradict your policy, ask the same question five ways. The transcript of that session is your remaining setup list.
Measure it like an employee
Review a sample of conversations weekly, score resolution versus handoff versus error, and update the knowledge base with every miss. Teams that treat the setup as finished at launch are the ones whose AI drifts into nonsense within a quarter.
Ten minutes of weekly review keeps resolution rates climbing. Zero review guarantees decay.
Key takeaways
- The knowledge base is the product - write real answers, include prices
- Hard-route refunds, cancellations and anger to humans
- Weekly conversation review or the quality decays within a quarter
Want this looked at on your account?
Free 15-point health check with an ex-HighLevel engineer - findings with proof, fix list you keep.
Book a free diagnosis