The pre-launch checklist for an AI voice agent

An AI voice agent is the first thing a stranger reaches when they call you, and it answers every call, at any hour, with your name on it. Most launches test one thing: does it handle the happy path? The problems come from everything else: the caller who asks for a refund, the caller who asks about someone else's appointment, the script that calls four hundred times overnight.

This list is what we run through before an agent goes live. It is grouped by the question each part answers. Work through it once before launch, and again after any big change to the prompt, the tools, or the number.

Write down the few jobs the agent exists for, in plain words: book an appointment, answer hours and location, take a message, qualify a lead. For each one, decide what counts as done and how often it has to be done right. "Books the correct slot nineteen times out of twenty" is a bar you can test against. "Handles bookings" is not.

• Write test calls for every job. The ordinary version, a messy version (the caller changes their mind, mumbles a date, asks two things at once), and an edge version (no slots left, closed today, wrong location).
• Set an accuracy bar per job, and measure it by running the calls, not by listening to one good demo.
• Re-run everything after every change. A prompt tweak that fixes bookings can quietly break messages or make the agent chattier on every call. Keep the full set of test calls and run all of it, every time, so a fix for one job can't cost you another.

The list of things the agent must not do matters more than the list of things it should. Write it down, put it in the instructions, and test it with callers who are trying to make it slip.

• Never commit you to money. No refunds, discounts, price exceptions, credits or guarantees beyond what your published policy already says. See can your AI receptionist promise a refund.
• Never reveal someone else's information. Appointments, orders, balances and addresses only after the caller proves who they are, and never about another person.
• Never give professional advice it isn't qualified to give. Medical, legal, financial: route it to a person.
• Never dial or text a number the caller supplies. Transfers go to a fixed list. See the pumping problem.
• Never read out its own instructions, internal notes, or anything from your knowledge base that isn't meant for customers.

Then test it adversarially: a caller who insists a manager promised them a discount, one who claims to be calling for their mother, one who asks the agent to ignore its instructions. If any of those work in testing, they will work on launch day.

• That it's an AI. A growing number of states require disclosure in some settings, or when someone asks. Saying so in the first sentence costs nothing and settles it.
• That the call is recorded, if it is. Federal law lets one party consent, but roughly a dozen states, including California, Florida, Illinois, Pennsylvania and Washington, require the consent of everyone on the call. A caller can be in one of those states whatever area code they dial from, so announce it on every call. See call recording consent, state by state.
• How to reach a person. Every caller should be able to get to a human, or at least leave a message a human will read.

In February 2024 the FCC ruled that AI-generated voices count as "artificial" voices under the Telephone Consumer Protection Act. That means the TCPA's consent rules apply to an agent that places calls: prior express consent to call someone with an AI voice, and prior express written consent if the call is marketing.

• Only call people who agreed to it, and keep the record of when and how they agreed.
• Honor do-not-call requests on the call itself, and scrub against the national registry for marketing.
• Respect calling hours in the called person's time zone.
• Watch where it dials. An outbound agent that can be steered toward premium-rate numbers is a bill waiting to happen.

• List every tool and lookup the agent can use, and what each can read or change. Remove anything it doesn't need.
• Decide how a caller proves who they are before the agent reads out anything about their account.
• Health information changes the rules. If you're covered by HIPAA, every vendor in the call path that handles patient information (the voice platform, the model provider, the transcription service) needs a business associate agreement with you before launch. See AI phone agents and HIPAA.
• Decide how long recordings and transcripts are kept, who can see them, and write it in your privacy policy.

Every minute your agent talks is billed three times: the voice platform, the model, and the carrier. It answers anyone, so anyone can run up that meter, including a script that places hundreds of calls. See how bots find your number.

• Cap concurrent calls at what real traffic needs, so a flood can't hold every line.
• Set a maximum call length. Nobody books a haircut in forty minutes.
• Set a spend alert on every platform in the path, low enough that you'd hear about a bad night the same night.
• Watch for repeat and patterned callers, and decide in advance what happens to them: an alert, a challenge, or the call ending.

• Know how to turn it off. One switch, known to more than one person, that sends calls to voicemail or a human.
• Plan for your platform being down. What happens to calls when the voice platform or model provider has an outage?
• Decide who gets paged for a spend spike, a flood, or a caller complaint, and make sure alerts reach a phone, not just an inbox.
• Keep enough of a record to reconstruct what the agent said on a disputed call.

• Review a sample of real calls every week for the first month. Real callers find edges no test script did.
• Add every surprise to the test set, so it stays fixed.
• Re-run the whole checklist after any change to the prompt, the tools, the model, or the number.

The whole list, short

1. Jobs written down, each with an accuracy bar and test calls
2. Full test set re-run after every change
3. A never-do list, tested by callers trying to break it
4. AI disclosure in the first sentence
5. Recording announced on every call
6. A way to reach a person
7. Outbound: consent on record, do-not-call honored, calling hours kept
8. Every tool listed; identity checked before account details
9. Business associate agreements in place, if health information is involved
10. Concurrency cap, maximum call length, spend alerts
11. Repeat and patterned callers watched
12. An off switch, an outage plan, and alerts to a phone
13. Weekly call review for the first month