Customer-facing AI agent for narrow request types
An AI agent in the help centre and chat that answers customers directly, limited to a defined list of request types with current documented answers, that tells customers it is AI and hands them to a person in one step.
Supporting approachTechnique: Retrieval-Augmented Generation
The AI agent answers customers in the help centre and in-product chat, but only for request types on an approved list, such as how to configure a feature in the customer’s version, whether their plan includes a feature, or how to connect a supported integration, each backed by current articles, and it says it is an AI.
When a question is outside the list, when the customer reports an error no article explains, when a known incident is open, or when the customer asks for a person, it hands the conversation to an agent with the plan, version, and error details it has gathered.
A support lead reviews a sample of its conversations every week, and a request type comes off the list when its errors or satisfaction slip.
This is the right choice for high-volume how-to questions about stable, documented features, after the knowledge base, routing, and agent assist are in place, and the only software-company failure in the sources, an invented login policy, shows what happens when it answers outside that list.
It is the wrong choice for new defects, outages, billing disputes, workspace access, complaints, and anything where a wrong statement could leave the company liable or cost the customer, and it is not a reason to cut the team until its resolution and satisfaction have held for months.
Strengths for this problem
- Answers how-to questions about a specific feature in the customer's version, such as configuring single sign-on or connecting an integration, at any hour and in many languages.
- Removes repetitive questions from agents' queues, leaving them the problems that need investigation.
- Can collect plan, product version, workspace, and error details or log references before handing over, so the engineer or agent starts with what they need.
Limitations for this problem
- States things its sources do not say, and customers act on them, as when a software company's support bot invented a login policy and customers cancelled.
- The company is answerable for what the bot says, as a Canadian tribunal held when an airline's chatbot misstated its fare policy.
- Customers in a 2024 field experiment at an e-commerce platform rated the chats the AI handled lower, though repeat contacts did not rise, and human intervention after an AI failure worked less well for emotional escalations than for technical ones.
- In a sales field experiment, customers who were told they were talking to a bot judged it less knowledgeable and less caring and bought less, but hiding it breaks the EU AI Act's transparency rule.
Data needed
- A list of permitted request types, each with the articles the agent may answer from and their product version.
- The customer's plan and entitlements, to avoid answering with features the customer does not have.
- Every AI conversation with its outcome: resolved, handed to a person, abandoned, or reopened later as a ticket.
How it is measured
- Share of AI conversations resolved, confirmed by the customer and by no ticket on the same issue within a week, not only those that never reached a person.
- Customer satisfaction for AI-handled conversations against agent-handled ones for the same request types.
- Rate of wrong statements found in a weekly quality review, especially about pricing, policies, and product behaviour.
- Time from a customer asking for a person to reaching one.
Pitfalls
- Answering from articles written for an older version or a different edition, so the customer is told to click a menu that has moved or use a feature their plan does not include.
- Letting the agent guess at the cause of a new error or an outage instead of pointing to the status page and handing over.
- Launching across all request types at once, including billing disputes, account access, and complaints.
- Measuring deflection, so customers who gave up count as resolved.
- Letting the agent answer from its general knowledge when no article covers the question, instead of saying so and handing over.
- Giving the agent tools to issue refunds, change ownership, or alter subscriptions, which a manipulated conversation could trigger.
- Cutting agent headcount on launch figures before repeat contacts and satisfaction have been measured for months.
Human oversight
The support lead owns the list of permitted request types and removes any type whose error rate or satisfaction falls below the agreed level. Team leads review a sample of AI conversations every week. Any customer can reach a person in one step, and every request involving money, account access, data deletion, or contract terms goes to a person.