Why teams migrate
Rules-based chatbots work when people ask predictable questions in predictable words. Most real conversations are not like that.
The fallback problem. A bot that only knows its intents ends many conversations with “I didn’t understand that.” Users learn to type “agent” to get past it, and support gets the conversation anyway, without context.
Intent sprawl. Every new question needs a new intent, training phrases and a response. Over a few years the intent list grows into hundreds of overlapping entries that nobody wants to touch.
Answers, not actions. Many bots can explain a refund policy but cannot check whether an order is eligible. An agent with tools can look up the order, apply the policy and start the return. For the difference in detail, see chatbot vs AI agent.
The migration plan
- Inventory. Export the bot’s intents, flows and a few months of transcripts. Note which intents get the most traffic, where the fallback fires and where users ask for a person.
- Prioritize. Group conversations into answer questions (policy, hours, how-to) and action requests (order status, booking, account changes). Answer questions are usually the fastest win.
- Prepare content and tools. Gather the help articles and policies the agent may answer from, fix what is out of date, and define each tool with a narrow purpose and required parameters. Our AI customer support agent guide covers the architecture.
- Rebuild. Write the agent’s instructions, connect retrieval over your content, add tools, and keep scripted flows for steps that must be exact. Define clear handoff rules.
- Run in parallel. Replay a few hundred real past conversations through the agent and score the answers. Then run it in shadow mode or for staff only before customers see it.
- Cut over. Launch on one channel or a share of traffic first, with handoff to a person always available. Review flagged conversations daily for the first two weeks.
- Decommission. Once the agent handles the full volume, retire the old intents and flows, keeping any scripted flows the agent still uses.
What to watch for
- Webhook and integration changes. Webhook fulfillment endpoints built for specific intents need to become general tools. Update any channel integration that pointed at the old bot.
- Credentials and permissions. Give each tool its own API credentials with the narrowest access it needs. An agent that can look up orders should not also be able to issue refunds unless you decide it should.
- Conversations in flight. Pick a cutover time when traffic is low. Conversations started on the old bot should finish there.
- Error handling. Decide what happens when a tool fails or times out: retry once, then apologize and hand off with context. Never let the agent guess at a failed lookup.
- Prompt injection and abuse. Users will try to make the agent ignore its instructions. Keep tool permissions narrow, validate tool inputs and log every action.
- Cost per conversation. Long conversations with large retrieved context cost more. Monitor model spend weekly and cap context size.
Timeline and cost
An agent that answers from your documents and hands off cleanly fits our AI chatbot band of $5,000–$20,000 and typically takes 2–4 weeks. An agent that also takes actions in your systems through tools fits our custom AI agent band of $15,000–$60,000 over 3–8 weeks. The AI chatbot cost guide explains where the budget goes.
What we recommend
Keep the rules-based bot if it is small, accurate and rarely hits fallback. Otherwise, start with answer questions grounded in your own content, add tools for the top two or three action requests, and keep human handoff one message away.