Skip to main content

​
Customer Support

A customer support AI Agent can help users find documented information, follow troubleshooting steps, and prepare a handoff to a person. Design it as part of a support process, not as a replacement for policy, account controls, or human judgment.

​
Define the Scope

Start with a limited set of supported tasks, such as:
  • Answering questions from approved help content
  • Guiding users through reversible troubleshooting steps
  • Collecting details needed for a support request
  • Summarizing a conversation for a human agent
  • Drafting replies for human review
List tasks the Agent must not complete without an approved tool and appropriate authorization. Examples include changing account access, issuing refunds, or making legal commitments.

​
Configure the Agent

1

Define the Role

Describe the products, users, languages, and issue types the Agent supports. State when it must say that it does not know.
2

Add Maintained Knowledge

Connect current help articles, policies, and troubleshooting guides. Assign an owner and review date to each source.
3

Set Response Rules

Require concise steps, distinguish confirmed facts from suggestions, and ask for missing context before diagnosing a problem.
4

Design Escalation

Specify when to involve a person and which details should be included in the handoff.
5

Test the Experience

Test both the Agent and its intended Distribution Channel with common questions, outdated assumptions, unsupported requests, and sensitive cases.

​
Example System Prompt

Adapt this example to your policies and available capabilities.

​
Escalation Design

Escalate when:
  • The user asks for a person
  • The request requires identity verification or restricted account access
  • A policy exception or financial decision is required
  • The available knowledge does not support an answer
  • Troubleshooting could cause data loss or another irreversible outcome
  • The conversation involves safety, legal, or compliance concerns
An escalation summary should separate user-provided facts, actions already attempted, relevant source material, and the unresolved question.

​
Evaluation

Build a repeatable test set and review:
  • Grounding: Is the answer supported by current source material?
  • Resolution path: Are the next steps clear and appropriate?
  • Boundary handling: Does the Agent avoid unsupported actions and claims?
  • Escalation: Does it hand off at the correct point with useful context?
  • Privacy: Does it avoid collecting unnecessary sensitive data?
  • Tone: Is the response respectful without hiding uncertainty?
Track results from your own workflow. Do not assume a specific automation rate, response time, or cost saving before measurement.
Review privacy, security, retention, and access requirements before connecting customer records or support systems. A connected app should receive only the permissions needed for the approved workflow.

Create an Agent

Create the initial customer support Agent.

Craft a System Prompt

Write testable behavior and escalation instructions.

Training Concepts

Choose between maintained knowledge, examples, and fine-tuning.

Distribution Channels

Configure how users reach the support experience.