Hiring an AI Voice Agent Developer? Ask These 10 Questions

Ten questions to ask an AI voice agent developer, what a good and a bad answer sound like, and a 20-point scoring table that tells you whether to hire or pass.

By Carlos Familia, CRM & AI automation expert. Updated October 16, 2026. 7 min.

Before the HVAC company had a CRM worth booking into, it wanted its phones answered. That order shaped the whole build. The Retell receptionist I set up for them in California writes every service request to a dispatch board, one kanban card per call with the transcript and the recording on it, on a local number. The pipeline and the follow-up workflows came later.

So my first question for any voice developer is not about voices. It is where the booking lands, and whether your office can read what the caller actually said.

Put the ten questions below to every developer on your shortlist and score each answer 0, 1 or 2. Hire the one whose good answers come with a failure they found and fixed. Pass on anyone whose best evidence is a demo recording. If you are hiring on a marketplace, read the AI voice agent red flags too, and if you want the parts of a call explained first, start with how AI receptionists work.

Questions about the call itself

1. Where does a booking land, and can my staff read the call?

A good answer sounds like: "On your calendar or your job board, as a record with the transcript and recording attached. A booking that fails goes to a list someone checks every day."

A bad answer sounds like: "You'll get an email summary."

A summary is the model's version of the call. The transcript is the caller's. On the HVAC board, the dispatcher opens a card and reads the customer's own words.

2. What will you change if it sounds robotic?

A good answer sounds like: "Settings before voices: the model temperature, backchannel, and a voice, model and provider that match the language. If a voice has a flaw I can't tune out, I change provider."

A bad answer sounds like: "I'll pick the most realistic voice."

The HVAC client found the first version too robotic. Raising the temperature from 0.2 to 0.75 helped, and backchannel (the small "mm-hm" a listener makes) did the most. On a collision shop's agent, the voice had an echo, so I moved it to another provider instead of fighting it.

3. How does it confirm a phone number?

A good answer sounds like: "It asks for the number and confirms it digit by digit. It never assumes the number on caller ID is the one to use."

A bad answer sounds like: "It takes the number from caller ID."

On voice agents for a marketing agency's salon clients, the agent read callers their own number back to them. I fixed it on three agents the same way: ask for the number, then confirm it one digit at a time.

Questions about what it may say

4. Which facts about my business can it state, and who checked them?

A good answer sounds like: "Only facts you approved in writing, like hours, services, prices and service area. If something isn't verified, it stays out of the prompt, and the agent says it will find out."

A bad answer sounds like: "It learns everything from your website."

Loading a restaurant's menu into a phone ordering agent turned up five price conflicts that I had to confirm with the owner before launch. For a collision shop whose details I could not verify, the agent states no address, no certification and no warranty terms at all.

5. Which questions does it skip, and what reaches my staff?

A good answer sounds like: "It asks only what applies to this caller, and every call ends with a rating or a summary your team can act on. Anything urgent goes to a person, with the context."

A bad answer sounds like: "It runs the full script with everyone."

The collision intake agent I built branches on the situation. Price shoppers skip the empathy opener, towing comes up only when the car can't be driven, and insurance questions come up only when the job goes through insurance. Each caller ends scored hot, warm or cold, so staff know whom to call first.

Questions about testing and failure

6. How do you test without touching my real customers?

A good answer sounds like: "Mocks for every tool that writes to a live system, an isolated test contact, a test calendar, and a real phone for transfers and texts."

A bad answer sounds like: "I'll try it on your live line."

Retell's own testing documentation warns that any tool you leave unmocked, including custom functions, code tools, calendar tools and MCP tools, hits your production endpoint while a test runs. Transfers and texts go the other way: a simulation fakes them every time, so only a real phone proves them. On the HVAC build, the end-to-end certification ran on an isolated contact, so no real customer was part of it.

7. How will I know it stopped answering?

A good answer sounds like: "Alerts on silence, failed calls and a low balance, sent to a named person, plus a weekly count of calls on the platform against calls in your system."

A bad answer sounds like: "Check the dashboard now and then."

A tattoo studio's line started refusing calls, and the cause was billing: the voice account had no valid payment method. The agent itself was fine. The outages that hurt most look like that, and the red flags post collects more of them.

Questions about accounts and rules

8. Whose name is on the voice account, the number and the card?

A good answer sounds like: "Yours. I work as a user inside your accounts, and the low-balance alert goes to you."

A bad answer sounds like: "It all runs on my account, and you pay me monthly."

On a lead-gen agency's demo line, outbound calls stayed blocked until the voice account passed identity verification (KYC). The account holder is the one who verifies, so settle ownership before anyone signs up.

9. What has to be in place before it dials or texts anyone?

A good answer sounds like: "Consent on file, calling hours in the lead's own time zone, no dialing when the zone is unknown, a kill switch, and a carrier-registered number for any text that carries a link."

A bad answer sounds like: "It calls every lead the second they come in."

The federal rule on telephone solicitations keeps them to residential lines inside 8 a.m. to 9 p.m., by the local time of the person being called. Since an FCC declaratory ruling on February 8, 2024, AI-generated voices fall under the TCPA's artificial-voice rules, which generally means prior express consent from the person called, per Wiley's alert. My outbound build reads the time zone from the area code and skips numbers it can't place.

Texting has its own gate. A collision shop's agent could not send its tow link because the number was not A2P registered, and the pre-approved number pool sends fixed templates only. The steps are in my GoHighLevel A2P guide. Consent wording is for your counsel; a good developer builds to it.

10. What do I get at handover?

A good answer sounds like: "An export of the prompt and settings, every tool and what it writes to, the test report with the failures in it, access to the call log, the alert list and a one-page runbook for 'it stopped.'"

A bad answer sounds like: "You get the login."

Ask to see a handover pack from a past project with the client details removed. My scope of work template turns this list into contract lines.

How do you score the answers?

Give each answer 0 for a bad one, 1 for a vague one and 2 for a good one backed by a story from a real build. The maximum is 20. Two zeros among questions 6, 7 and 8 mean pass, whatever the total, because those three decide whether you hear about problems and whether you can leave.

Start the winner on one small paid task, such as a single call flow with three written test cases. The full HVAC build is in the HVAC AI receptionist case study, and the wider buyer's view is in hiring an AI automation consultant. You can also bring a developer's answers to a call on the calendar below.

About the author

Carlos Familia, CRM & AI automation expert. GoHighLevel, n8n, AI agents, servers. Full-stack developer.

I build CRM systems in GoHighLevel, n8n automations on my own servers and AI voice and chat agents for service businesses. Full-stack developer: I write the workflow, the API integration, the database and the server it runs on, and I test every build with real journeys before handoff.