How long does it take to launch AI agents?

The honest answer is that the technical work is rarely what sets the date. Connecting a phone number, an inbox and a CRM is measured in days. Deciding what the agent should say when a customer asks for a discount is measured in however long it takes the right people in your business to agree.

What actually takes the time

Four things, roughly in the order they become the bottleneck.

• Finding the policies. Most businesses know their refund window and their escalation rules, but the knowledge lives in people rather than in a document. Writing it down is often the first time it has been written down.
• Approving the wording. What the agent says is your voice going out to your customers. Somebody has to read it and own it, and that person is usually busy.
• Deciding the permissions. Which actions can it complete on its own, which need a person, and where is the line. This is a judgment call about risk, not a technical question.
• Connecting the systems. Genuinely the fast part, unless a system has no API, in which case we will tell you early rather than discovering it late.

A realistic shape

For one agent on one or two channels, working from documentation that already exists, two to three weeks from the first call to answering real conversations is normal. Add a week for each additional system that needs connecting, and add more if the policies have to be written from scratch.

Businesses that launch fastest have one thing in common: a single person who can approve wording without convening a committee. That matters more than the size of the company or the complexity of the stack.

What launch day looks like

Not a switch that flips everything at once. The usual sequence is to run the agent on a narrow slice first, typically one channel outside office hours, where the alternative is voicemail and the downside of an imperfect answer is small. You read the transcripts, you fix what reads wrong, and you widen.

Reading the transcripts is not optional in the first two weeks. It is the fastest way to find the questions you did not anticipate, and every business has some. The agent handles them by saying it does not know and offering a callback, which is the correct behavior, but each one is a gap worth closing.

What we need from you

• Whatever documentation exists, however messy. A help center, an internal wiki, a folder of PDFs. We would rather start from your real material than from a clean sheet.
• Access to the systems the agent needs to read, at the permission level it needs and no higher.
• One named person who can approve copy and settle policy questions.
• An escalation route that a real person actually watches.

That last one catches people out. An escalation path that goes to an inbox nobody reads is worse than no escalation path, because the customer has been told a person will follow up. Agree who owns it before launch, not after the first handoff.

If you want it faster

Narrow the scope rather than compressing the approvals. One agent, one channel, one clearly defined job, with everything else escalating to a person, can be live quickly and gives you something real to react to. Broad and rushed is the combination that produces a system nobody trusts, and rebuilding trust takes longer than building it.