AI voice agent launch checklist, step by step

An AI voice agent is ready to launch when you can answer “yes” to every question below: you know what it has to do, you have heard it on your own phone, you have checked the data side, you have a small pilot planned and you know who takes the cases the agent cannot resolve. The checklist follows the 9 steps of the AgentVocal setup wizard, plus the steps before and after that the form does not cover.

Key idea

Use the list as a working sheet: tick it off, write down who owns each item, and do not launch until the items marked as blockers are settled.

How does this list map onto the 9-step wizard?

The wizard in the platform covers configuration. This list also covers the decisions around it.

The details of each step are on the how it works page. The technical setup itself is quick: in a real test, configuring a complete agent took 4 minutes and 38 seconds. What takes time is preparing your own content and decisions, which is what this list is for.

1. What is the agent’s goal and which scenario does it run? (steps 1 to 3)

Without a clear goal, the agent becomes “a robot that calls people”. Start with one sentence.

Goal

• I can say in one sentence what the agent must do (for example: “confirm cash on delivery orders before dispatch”).
• I have chosen one scenario to start with, not five at once. Others can follow after the pilot.
• I checked whether a ready-made scenario fits, or whether I start from scratch.
• I know what a successful call looks like and which outcomes are acceptable (see how to measure an AI voice agent).

Company details and modules (steps 1 and 3)

• The company details are correct (for Romanian companies, the tax ID pulls the data from ANAF, the national tax authority).
• I ticked only the modules I need: initiated calls, received calls or after-call messages.
• I read the estimated cost in the form, and the billing rules are clear to me from the pricing page.

What stays with people

• I wrote the list of cases the agent does not handle (complaints, exceptions, decisions).

2. How does the agent sound and what does it know? Voice, knowledge, integrations (steps 5 to 7)

Voice, language and tone

• I chose the language (or languages) my customers speak; see the available languages.
• The voice suits the business and its audience: I listened to a few options.
• The agent introduces itself as a virtual assistant and states its purpose in the first sentence.
• The company, product and street names are pronounced correctly.
• I decided what it must NOT promise (guaranteed deadlines, discounts, exceptions).

If you use several languages, test each one separately. An additional language also means an additional setup, so make that decision deliberately.

Knowledge, data and integrations

This is where call quality is won or lost.

Knowledge base

• I prepared the frequent questions, policies, opening hours and exceptions, written for voice. The method is in the guide to the knowledge base.
• There are no contradictory documents.
• I wrote the “I don’t know” rule: what the agent does when it has no answer.

Integrations

• I connected the systems I need: CRM, ERP, calendar, email, WhatsApp (see integrations).
• I know where the call outcome is written and who reads it.
• I checked that a test note arrived in the right system.

Phone number (step 7)

• I saw the number assigned to the agent and decided how to announce it (especially if it also takes incoming calls).
• Customers who call back reach the agent or a colleague, as planned.

3. What have you checked for GDPR, lists and calling hours? (step 8)

GDPR and informing the person

Go through the full guide to call recording and GDPR; here, check only the essentials. This is not legal advice; ask your legal adviser.

• The DPA was accepted at sign-up and I know who manages it.
• The purpose of the calls and the legal basis are recorded and checked.
• The recording announcement is made at the start of the call, and I have heard it.
• The list contains only the data needed; the agent does not ask for special category data.
• The retention period for recordings is set.
• I know how to handle a request for access, erasure or objection.
• The do-not-call list is loaded and has an owner.
• No minors are included in automated campaigns.

Lists, calling hours and cadence

• The contact list is clean: no duplicates, numbers in the correct format.
• Uniqueness is clear: the same person is not called twice for the same case.
• Calling hours are realistic for my audience (not very early in the morning or late in the evening).
• I set the number of retries and the times when the agent tries again.
• I chose the days when it does not call (public holidays, weekends, if relevant).
• I know that only calls nobody answers are not charged; voicemail is charged as a normal minute.

4. What do you test in the test call? (step 9)

The agent calls your own phone before it calls anyone else. Treat it like an exam.

• I made at least one test call on the real scenario, from start to finish.
• I also tested a call where the “customer” asks questions outside the documents.
• I tested a refusal, an interruption and a request to speak with a person.
• I tested “please don’t call me again” and checked that the number went onto the restriction list.
• I read the transcript and the summary in the portal, not just listened to the call.
• The outcome reached the CRM or ERP.
• A colleague who did not build the agent also ran a test.

If something does not sound right, adjust the instructions and repeat. Do not launch on the whole list with a plan to “fix it as we go”.

5. How do you run the pilot and the first week?

The pilot on a sample

A small pilot limits the impact of mistakes and gives you real data.

• I chose a sample from the list, small and representative (you set the size based on volume and risk).
• I avoided sensitive contacts for the first calls (key accounts, delicate cases).
• I set the criteria for continuing in advance: what I need to see before expanding.
• I set the criteria for stopping: which situations pause the campaign.
• I know how to stop or pause a campaign.
• The person responsible is available on the pilot days.

Monitoring in the first week

The first week is for listening, not waiting.

• On day one, I listen to almost every call.
• I read the summaries and the escalation reasons every day.
• I log problems by category: missing information, wording, pronunciation, missing rule, list or calling hours.
• I make small fixes the same day and log them with a date.
• I check complaints and do-not-call requests immediately.
• At the end of the week I run the first full review (format in the guide on measuring).
• I decide: expand, keep fixing, or stop.

6. Who takes the cases the agent does not resolve?

Escalation is the safety net. Without it, a difficult case stays unresolved.

• Every type of case has an owner: who takes the requests noted by the agent, who takes complaints.
• I decided how quickly a colleague gets back to people (and the agent’s script says the same, if it mentions a timeframe).
• The agent notes the request, and the call appears in the portal with a summary so someone can follow up.
• There is one place where you can see what was escalated and what was resolved.
• I know how to adjust the agent when the same escalation reason keeps coming back.
• There is a backup person for holidays.

The verdict: can you launch?

You can expand beyond the pilot if:

1. The test call and the real calls you listened to are acceptable by your criteria.
2. The data and information side has been settled with your legal adviser.
3. Escalations have an owner and a route.
4. The weekly review is in the calendar.

If one item is missing, stop and fix it; that costs less than a campaign gone wrong. When you are ready, sign up and go through the 9-step wizard with this list beside you. Want to hear how it sounds first? Try a demo.