The seven-day go-live plan for a new AI phone agent

Going live in a week is realistic when the work is split into small, ordered steps and the right people are in the room. This plan is the order we follow with new customers. It assumes one main use case, such as answering inbound calls and booking appointments, and a small group of people who can make decisions quickly.

A seven-day AI voice agent launch plan works when it covers one job, one language and one line. Days one and two set scope and load knowledge, days three and four connect systems and write rules, day five is for test calls, day six is a soft launch, and day seven reviews and widens traffic.

The scenario is illustrative: a clinic group, a property firm or a service desk starting with a single line and a single job.

Before Day 1: who you need

Name these people before you start. Without them, days slip.

• An owner who can make decisions about scope and wording.
• A subject expert who knows the answers callers need.
• Someone with access to your calendar, CRM or booking system.
• A tester who will make calls and write down what they hear.

Agree on one measure of success, such as the share of calls resolved without a human, or bookings made per day.

Day 1: Choose the number and the job

The first day is about decisions, not configuration.

• Decide the one job the agent will do first. Do not try to cover everything.
• Choose the number. A new native number is the quickest. If you want to keep your existing number, decide whether to forward to it, or connect it to your PBX or CCaaS.
• Agree the languages the agent should speak, and the market it should sound like.
• Decide the hours. Many teams start with after-hours and overflow, then extend.

Checklist:

• One job written in a sentence
• Number route chosen
• Languages and hours agreed
• Owner, expert and tester named

Day 2: Load the knowledge

The agent is only as good as what it knows. Gather the sources and clean them.

• Collect your price lists, policies, opening hours, locations and frequently asked questions.
• Remove anything out of date or contradictory.
• Mark what the agent may say freely, what it may say only to verified callers and what it must never say.
• Add local terms: branch names, product names, staff names and common mispronunciations.

Keep it short and accurate. A hundred precise answers beat a thousand loosely related pages.

Checklist:

• Documents gathered and checked
• Contradictions resolved
• Local names and terms added
• Sensitive topics marked

Day 3: Connect the systems

Now connect the places where work happens, such as the integrations you already use. The agent should do the work during the call, so it needs to read and write.

• Connect your calendar or booking system.
• Connect the CRM so each call is logged against the right customer.
• Decide what the agent can create: bookings, tasks, callbacks, notes.
• Test one read and one write end to end, with a dummy record.

If a system is not ready, begin with the work that does not need it, such as answering questions and taking messages, and add the integration later.

Checklist:

• Calendar connected
• CRM connected
• One test booking made and visible
• One test task created for the right person

Day 4: Write the rules

Rules decide what the agent may do alone, what needs approval and when it hands over. Day 4 is the most important day. See Writing rules your agent can keep for example wording.

• List the actions the agent can take by itself.
• Set limits: amounts, dates, number of attempts.
• Name the situations that trigger a handoff, and who receives it.
• Decide what is recorded, redacted and kept.

A short example of the shape of a rule:

Action: Move an existing appointment
Allowed: Yes, to any free slot in the next 30 days
Needs approval: Moves further out than 30 days
Never: Cancel without the caller confirming

Checklist:

• Permitted actions listed
• Limits set
• Handoff triggers and recipients named
• Recording and redaction decided

Day 5: Test calls

Now you call it. Not once, but many times, and not only with the easy questions.

• Call from your own phone, in the languages your customers use.
• Interrupt it. Talk over it. Change your mind halfway.
• Ask things it should not answer, and check it declines or hands over.
• Ask for a person and check the handoff summary your team sees.
• Try a noisy room, a speakerphone and a poor signal.

Write down every moment that felt wrong, however small. Group them as knowledge gaps, rule gaps and wording problems. Fix them and call again.

Checklist:

• At least twenty varied test calls
• Handoff tested with a real team member
• Failures logged and grouped
• Fixes made and retested

Day 6: Soft launch

Go live quietly. Route a small slice of real traffic, such as after-hours calls only, or a fraction of overflow.

• Tell the team what to expect and how to take a handoff.
• Watch the command centre during the first hour.
• Listen to the first calls the same day.
• Check that bookings and notes land in the right places.
• Have a way to pause the agent instantly, and make sure the team knows how.

Treat the soft launch as the last test, not the first celebration.

Checklist:

• Small slice of traffic live
• Team briefed
• First calls reviewed the same day
• Pause switch known to everyone

Day 7: Review and widen

Look at what happened.

• Review flagged calls and fix the knowledge or rules behind them.
• Check the measure of success you chose on Day 1, using the AI voice agent KPIs as a guide.
• Read the end-of-day summary and decide whether it tells you what you need.
• Widen the traffic, or extend the hours, in a controlled step.
• Plan the next use case, but do not start it until this one is stable.

Checklist:

• Flagged calls reviewed
• Success measure checked
• Traffic widened in one step
• Next use case chosen, not started

Common ways the week slips

Most delays come from the same few causes:

• Nobody owns decisions, so questions wait for days.
• Knowledge is scattered across inboxes and old documents.
• System access arrives late.
• The scope grows on Day 3.

The fix for all of them is the same: keep to one job, name the owner and settle questions in the room.

If you only have three days

Some teams need a faster start. Compress the plan by narrowing the job, not by skipping steps: one line, one language, one hour window. Do Days 1 and 2 together, Days 3 and 4 together, and keep a full day for test calls. Skipping test calls is the one shortcut that always costs more time later.

What to do next

Put the seven days in your calendar and name the four people. If you are not sure about the first job, start with the one that costs you the most missed calls.

See what is possible on the pricing page, review the appointment booking use case, or try a call to hear the starting point.

Frequently asked questions

How long does it take to launch an AI voice agent?

This plan takes seven days for one job, one line and one language. If you have less time, narrow the job rather than skipping steps, because skipping test calls costs more time later.

Who needs to be involved in the launch?

Name an owner who makes scope decisions, a subject expert, someone with access to your calendar or CRM, and a tester who makes the calls. Without them, days slip.

What should the test calls cover?

At least twenty varied calls from your own phone, in the languages your customers use. Include interruptions, questions the agent should decline, and a handoff tested with a real team member.

Can I start with after-hours calls only?

Yes. Many teams start with after-hours and overflow calls, then extend the hours once the agent is stable. To hear the starting point on a live line, try a call.