Use an installed Email relay to answer a public policy question, such as how returns work. The agent builder owns the answer and Knowledge; the developer owns mail transport. This guide configures the support behavior after receiving and replying already work.
Before you begin
Have a released test Environment, a working connection manual, an inbox you control and a CX reviewer. Start with Build a support agent if the agent does not exist yet. Keep AUTO_SEND=false and restrict ALLOW_SENDERS to the pilot sender.
Choose an approved public policy source and its expected answer. Do not use an email address as proof of account ownership or expose private order information through this public-support sample.
Configure a useful answer
Prepare the support source
Add the approved policy to Knowledge and make sure the intended agent can use it. Give the reviewer a test question and a source-backed expected answer. For example, use “How do I return an unopened item?” with your actual policy; do not invent a return period for the test.
Set the email response behavior
In the agent’s instructions, adapt this example to the support scope:
Remove email-sending tools from this agent: the relay owns delivery of the approved answer. Help desk examples require AUTO_SEND=false throughout their supervised pilot.
The relay passes email_subject and email_sender as turn variables. Create those variables if the agent uses them. They supply context, not authorization. The review marker is a routing convention implemented by this example, not a confidence score or native hand-off action.
Publish and check the actual reply
Publish to the reviewed Environment used by VF_ENVIRONMENT. Send the policy question through the installed email address. Inspect the relay’s candidate and compare it with the source. Use Review replies before sending to correct and approve it.
Verify the customer outcome
Read the delivered answer
In the customer test inbox, confirm the answer states the relevant policy and a useful next step in readable plain text. Compare the provider record, relay job and runtime transcript; a browser Run or queued API response is only an intermediate check.
Test a follow-up and an unsupported question
Reply with a question requiring the first answer’s context. Confirm continuity. Then ask something outside the approved source; expect an honest limitation and review rather than an invented policy. Check that a message containing [HUMAN_REVIEW] remains held even if automatic sending is enabled for the later pilot.
The agent gives a confident but unsupported answer
Compare its source and instructions before approving more mail.
Keep review-first. Check that the correct Environment was published, that its Knowledge includes the approved source and that the instructions require an explicit limitation. Correct the candidate before delivery and repeat both supported and unsupported questions. A marker cannot guarantee detection of every uncertain answer.
The reply asks for a button click
The supplied relay sends text and holds supported visual/choice/handoff traces for review. Rewrite the agent’s path to ask a plain-text question, or have a person write the reply. Adding a link to a rich-response API does not implement email buttons or asynchronous human hand-off.
Release changes
Keep the pilot reviewed until the CX owner accepts answer quality and failure handling. A later Knowledge or agent release needs another fresh email and a continuing-thread test. Automatic replies require a reviewed public-information scope and the connection manual’s production acceptance. If quality regresses, set AUTO_SEND=false, restart the service and review already queued ready jobs before they send; changing the flag does not recall accepted mail.