Weeks for a read-only agent on one channel; longer for one that takes actions in your systems. The model is rarely what slows a deployment. The time goes on cleaning up the knowledge base, connecting the systems the agent must read and write, building a test set from your own conversations, and earning depth in stages.
What takes the time?
Most teams expect the build to be the long part. It is usually the shortest. The phases that set the pace are these.
Choosing the scope. Picking which contact reasons to automate first, and which to leave with people, needs data on what customers actually ask and a decision from whoever owns support. Teams that skip this end up automating the noisy reasons rather than the valuable ones.
Knowledge cleanup. Help articles written for people are often vague, contradictory or out of date. An agent turns every flaw into a wrong answer. Rewriting the articles for the chosen contact reasons, with owners and conditions, is commonly the longest single task.
Integrations. Reading order status or account balances needs read access to the systems of record. Taking actions, such as issuing refunds or rebooking deliveries, needs write access with limits, sandbox testing and security review. Each system adds time, and a legacy system with no clean interface can add a lot.
Testing. Building a case set from real conversations, writing expected outcomes, replaying, fixing and replaying again. Then shadowing on live traffic.
Approvals. Security, data protection, and in regulated businesses compliance and risk sign-off. These run in parallel if started early and in series if started late.
How do the factors change the timeline?
| Factor | Shorter | Longer |
|---|---|---|
| What the agent does | Answers questions only | Takes actions that change records or move money |
| Channels | One, usually chat | Several, including voice |
| Knowledge base | Current, owned, one topic per article | Scattered, outdated, mixed audiences |
| Systems | Modern, with clear interfaces | Several legacy systems |
| Regulation | Unregulated product | Regulated customers and products |
| Decision-making | One owner who can say yes | Several committees |
A general view across agent types is in how long it takes to deploy an AI agent, and the finance-specific version in how long it takes to deploy an AI agent in finance.
How should you plan it?
Stage it by capability, not by agent. A sensible sequence:
- Read-only answers on one channel. Order status, opening hours, how-to questions. Low risk, quick to test, and it proves the knowledge base.
- Drafts for a person to send. The agent drafts replies for higher-risk contacts; staff review and send. This builds evidence without customer exposure.
- Routine actions within limits. Booking, rebooking, small refunds, once the drafts show the agent gets them right.
- More channels. Voice and email after chat, each with its own verification and handover tests.
Each stage goes live when its evidence is in, not on a calendar date. That makes the timeline less predictable at the start and much more predictable for each later stage, because the team learns where its own delays come from.
An illustrative example
Take a mid-sized online retailer starting with order tracking and returns in chat.
Scoping takes a short workshop and a look at a month of tickets. The knowledge base for returns turns out to have four overlapping articles and needs rewriting, which takes longer than expected. Order tracking needs read access to the warehouse system, which has a clean interface, so it is quick. The case set and replay take a couple of weeks, and shadowing another two. Order tracking goes live first; returns follow once the rewritten articles pass replay.
Refunds are planned next, at draft depth. They need write access to the payment system, security review and a limits table signed by finance, so they are scheduled as a separate stage rather than squeezed into the first.
What do people forget to plan for?
Logging. Every verdict, action and handover should be recorded from the first day of live traffic, and the retention period decided before launch. If your agent falls under the EU AI Act's logging duties, the considerations are in how long to keep AI agent logs under the EU AI Act.
Handover staffing. People must be ready to receive escalations from day one, with the agent's notes in their tools.
Ongoing ownership. Someone must own the knowledge base, the test set and the agent's limits after launch. Deployment is the start of that work, not the end.
Where Pryme fits
In Pryme, Agent Studio assembles the agent's brief from a plain-language description, and Attest certifies each agent on your own cases, so capabilities go live one at a time as the agent's certified depth reaches the depth each needs. More is in the AI support agent that knows when to say no.
Questions people ask
Can we launch everything at once to save time?
You can, but a problem in one capability then holds back all of them. Staged launches are usually faster overall.
Does voice take longer than chat?
Usually, because of verification, speech recognition, latency and transfer testing. Start with text if you can.
What is the most common cause of delay?
Knowledge that was assumed to be ready and is not. Audit it before you set dates.