How to Migrate a Chatbot to an AI Agent: A Playbook

The chatbot answered questions from a knowledge base and handed off when it could not. Customers learned to type "agent" immediately, and satisfaction scores followed them out the door. The upgrade is not a better model; it is giving the system what it lacked: the ability to read the customer's record, act within policy, and escalate with context, under controls that keep those actions safe. This playbook makes that upgrade intent by intent. It applies chatbot vs AI agent and the operating model in the AI agents for customer operations whitepaper.

Step 1: What do the transcripts say?

Analyze the chatbot's transcript history: intents by volume, the share resolved without handoff, the point in each intent where the bot handed off, sentiment at handoff, and the resolutions people then provided. This produces the intent inventory, the failure map, and the raw material for the golden set. Intents the bot handled well and cheaply can stay on knowledge answers; intents that failed for lack of action are the agent's first targets.

Step 2: Which intents go first?

Test	Question
Clear meaning	Is the customer's goal unambiguous?
Structured resolution	Does a defined path exist?
Available means	Are the data and actions reachable through tools?
Volume	Is there enough to measure?
Reversibility	Can the actions be corrected?

Order status, billing questions within policy, contact and address changes, and returns within policy typically pass all five. Cancellation, disputes, and complaints stay human.

Step 3: What tools does the agent need?

Expose the customer record, order and billing systems, and case management through a governed tool layer with scoped permissions, following the Model Context Protocol for the enterprise whitepaper. Classify each tool: reads flow; reversible writes (notes, address updates, replacement orders within policy) are sampled; consequential writes (refunds above threshold, closures) are separate tools, withheld or gated.

Step 4: How is identity verified?

Before disclosing or changing anything, the agent verifies the customer to the same standard a representative applies, using the existing verification flow where one exists. Verification state is carried as delegated context on every tool call so a customer can only ever reach their own data. The model is in the agent identity and access control whitepaper.

Step 5: Design escalation before enabling any action

A visible path to a person at all times; explicit triggers on low confidence, negative sentiment, repeat contact on the same issue, out-of-scope requests, and consequential actions; handoff with the conversation, record, actions, and assessment attached; and a human approval queue for gated actions with evidence. The oversight pattern is in human-in-the-loop AI explained.

Step 6: Build the golden set from real conversations

Redacted transcripts with verified correct resolutions, by intent, including cases where the correct action is to escalate and cases containing adversarial content. Score intent recognition, verification behavior, action correctness (end state, not transcript), escalation correctness, and tone. Gate every change on it. The method is in how to build a golden dataset.

Step 7: Run shadow mode

On live conversations for the target intents, the agent proposes resolutions and actions while the chatbot and people continue; outcomes are compared and disagreements analyzed. Expect to revise policy limits and escalation triggers. Continue until agreement is stable.

Step 8: Roll out intent by intent

Go live on two intents at suggest or act-with-approval, with reversible actions enabled first and consequential actions gated, per-intent dashboards live, and the chatbot's handling of other intents unchanged. Measure resolution without human touch, time to resolution, satisfaction, repeat contact, and escalation quality per intent, and cost per resolution. Add intents in waves. The role definition is in Digital FTE for customer support.

How does the knowledge base carry over?

The chatbot's knowledge base is an asset with a known problem: its articles were written to be read by a bot that could only answer, so they explain policy without stating the actions a resolver may take. Carry the content over into a versioned, retrievable knowledge layer, and revise the articles for the intents the agent now handles so each states the policy, the permitted action, and its limits in one place. Retrieval stays grounded: every answer cites the article it drew on, and articles that contradict the tool-layer policy limits are corrected rather than left for the agent to reconcile. The retrieval design follows the enterprise RAG reference architecture.

What changes for the support team?

The upgrade changes the team's work as much as the customer's experience. Representatives receive escalations with full context instead of cold handoffs, so their first minute is spent on the problem rather than re-collecting facts. Team leads own the intent specifications and adjudicate the agent's misses. Quality analysts score sampled agent conversations against the same rubric used for people. Knowledge managers become central, because every gap they close removes a class of escalations. Communicate these shifts before the first intent goes live; the team's cooperation is what makes shadow mode and the miss analysis work. The role model is described in the Digital FTE workforce planning whitepaper.

Step 9: When does the chatbot retire?

When the agent's per-intent resolution and satisfaction beat the chatbot's across the covered intents, and the remaining intents are either knowledge answers the agent also handles or human-only. Keep transcripts flowing into the golden set; the migration ends but the evaluation loop does not.

What are the common mistakes?

1. Swapping the model and calling it an agent.
2. Actions before verification and escalation.
3. Deflection as the success metric.
4. All intents at once.
5. Consequential actions gated in the prompt.
6. Discarding transcripts, the best data you have.

How does FISTA Solutions help?

FISTA Solutions upgrades chatbots to governed AI agents intent by intent, through forward deployed engineers working inside the support team on tools, verification, escalation, and evaluation, on the platform the AI enablement practice establishes. FISTA has delivered 150+ projects for 50+ companies across 12+ countries with 99.9% uptime.

To plan the upgrade for your bot, message FISTA on WhatsApp, or read AI customer support automation for the process context.