The objection I hear most often is not will it work. It is what happens the day it says something stupid to my customer.
That is the right question, and most people selling AI agents have no real answer to it. They demo the happy path: a polite customer, a simple job, a clean booking. Nobody demos the caller who is furious, or the one who says the word gas, or the one who asks for a firm price on a full system replacement.
I came into this work from risk and QA, so I will tell you plainly: an AI agent is not risky because it is intelligent. It is risky because nobody wrote the test cases.
Below is the checklist I run before an agent is allowed to speak to a real customer. Fourteen cases. If a vendor cannot walk you through their answer to each one, they have built a demo, not a front desk.
What an untested agent costs
An untested agent does not fail quietly. It fails in public, in writing, with a timestamp.
The costs stack in a predictable order. First a customer is annoyed. Then a booking is wrong, so a truck rolls to the wrong address or the wrong job type, which costs a half day. Then a review appears. Then your team loses confidence and starts working around the agent, which returns you to voicemail with an extra subscription attached.
None of that is a model problem. Every one of those failures traces back to a scenario nobody wrote down before go-live.
The checklist, fourteen cases
Run every case against the live configuration, not a sandbox. Write down the expected behaviour first, then see what actually happens.
Re-run cases 3, 7, 12 and 13 after every configuration change. Those four are the ones that cost real money when they break.
What changes once it is in
Before: an agent that performed beautifully in a demo and had never once been asked a hostile question.
After: fourteen documented cases, an expected behaviour written for each, and a record of what it actually did. When something does break later, and something always does, you have a baseline to compare against instead of a debate.
The team change matters as much as the technical one. Staff stop working around the agent once they have watched it handle the angry caller and escalate the gas leak correctly. Confidence is not a feeling here, it is the output of a test run.
Nothing ships until it has been stress-tested against a real customer conversation. Including the one where the customer is shouting.
The measurable: escalations that were correct, and cases re-run after each change. An agent that escalates cleanly is not a weak agent. It is the only kind worth putting in front of your customers.
