A Customer Support Agent That Does Not Make Stuff Up

The worst support agent is not the one that says "I do not know."

The worst support agent is the one that sounds confident and is wrong.

Customer support is one of the best use cases for agents, but only if the workflow forces the agent to stay grounded. A support agent should not improvise policy, invent product behavior, or promise something the team cannot deliver.

The job#

The job is not "answer tickets."

The job is:

• Understand the customer question.
• Identify the account and plan.
• Pull the right docs, changelog, tickets, and product state.
• Draft a source-backed answer.
• Escalate when the answer touches billing, security, legal, or data loss.
• Log what sources were used.

If the agent cannot find evidence, it should say that. That is not a failure. That is honesty.

Split the workflow#

I would split support into four agents.

Intent Classifier

Decides whether the request is product help, bug, billing, account, security, or escalation.

Account Context Scout

Reads approved customer state: plan, feature flags, recent events, previous tickets, current limits.

Answer Drafter

Writes the response with source links and a confidence level.

Escalation Reviewer

Checks whether a human needs to approve or take over.

This keeps the writing agent from pretending it owns policy.

Tool access#

A support workflow usually needs:

• Help docs
• Changelog
• Ticket history
• Account metadata
• Feature flag state
• Billing read API
• Product event summary
• Internal runbooks

Notice I said billing read API. Refunds, credits, plan changes, and account deletion should require approval.

The answer format#

Use a strict response shape:

Short answer:
What I checked:
What I found:
What I cannot confirm:
Next action:
Confidence:

This is boring on purpose. It prevents the agent from writing a novel and makes human review easier.

Evals#

A support agent needs evals against real or synthetic tickets:

• Correct intent classification.
• Correct doc citation.
• No unsupported claims.
• Escalation on billing and security.
• No private data leakage.
• Tone that sounds human but not fake.

Do not wait until customers are reading the answers to discover whether the agent hallucinates.

Build it in Codelit#

Try this:

Build a customer support agent workflow for a SaaS product. It should classify tickets, read docs and account context, draft source-backed answers, escalate billing and security issues, and require approval for account changes.

Build this support workflow

The goal is not to sound smart. The goal is to be correct, useful, and safe.