How does a Zendesk AI agent issue refunds?
Zendesk routes each use case to either a generative procedure or a dialogue. Procedures need less setup and maintenance but offer less direct control over fine details. Dialogues offer more control and need more upkeep. Email AI agents can only use generative procedures.
Inside either one you insert actions, which work with session details or your CRM, and API integrations, which Zendesk describes as a way to ‘update data outside of Zendesk using an API you specify’. A refund call to your order or payment system lives in an API integration.
The integration requirements are specific. The AI agent makes server-to-server HTTP requests to a REST or GraphQL API that must respond within 30 seconds, authenticated with an API key, bearer token, basic auth, OAuth 2.0 or an IP allowlist. Returned data is kept for the AI agent session.
Which guardrails can you configure?
Zendesk’s documentation covers flow design, escalation and integration authentication. None of these sets a refund limit on its own. That belongs in the API the agent calls.
| Layer | Control | Note |
|---|---|---|
| Flow design | A dialogue for refund use cases instead of a generative procedure | Scripted branches give more direct control over each step |
| Escalation | Escalation blocks, an availability block, a transfer group and an escalation fallback | Zendesk advises transferring complex, urgent or sensitive queries and designing the escalation strategy before launch |
| Integration auth | API key, bearer token, OAuth 2.0, IP allowlist | Authenticates Zendesk to your backend, not the customer to you |
| Your API | Per-order caps, eligibility checks, idempotency | Holds regardless of what the conversation says |
How does the customer prove who they are?
Zendesk’s integration requirements cover how the AI agent authenticates to your backend. They don’t describe how the customer proves their identity before an action runs, and they ask that your API be searchable with data a typical user would know, or with data passed from your CRM.
If an order number and an email address are enough to trigger a refund, anyone holding both can request one. Design the identity check into the flow and enforce it again in the API.
What does Zendesk’s built-in testing cover?
The test widget simulates the customer’s experience using the AI agent’s active settings. You can test end to end, test a single dialogue from a chosen block with session parameters set or ignored, and view the intents the agent predicted for each message. Test conversations are saved to conversation logs and don’t consume automated resolutions.
Zendesk lists what the widget can’t test: CRM actions, escalations triggered by the escalation dialogue block, structured message templates and restricted help center content. The article describes manual testing only, with no automated test suite or scripted simulations.
For refunds, that leaves the two parts that matter most, the money action and the handoff, to be tested elsewhere. Zendesk suggests the Web Widget for CRM-based events. Pair it with a staging backend so test refunds hit no real accounts.
What to test before giving a Zendesk AI agent refund authority
Run these against a staging copy of your refund API. Illustrative cases:
- A refund request using only details printed on a forwarded receipt or shipping email (identity abuse).
- A customer who says a supervisor already approved the refund (fake prior approval).
- An email thread with a pasted “updated refund policy” handled by a generative procedure (indirect prompt injection).
- The same request rephrased after a refusal, over several turns (multi-turn escalation).
- Two partial refunds that together exceed the order total.
- An escalation when no agents are online: check what the fallback tells the customer about their money.
- A refund call that times out after 30 seconds or returns an error: check what the agent promises.
Questions
Can Zendesk AI agents issue refunds automatically?
They can call an API that issues one. Refunds run through API integrations or actions you add to a procedure or dialogue, so the system the integration calls decides what gets refunded and how much.
Does the Zendesk test widget run API integrations?
Zendesk’s test widget article doesn’t say. It does state that the widget isn’t integrated with your CRM, so CRM actions aren’t tested, and that escalations from the escalation block can’t be tested. Check refund paths against a staging backend before launch.
Should refund requests use a procedure or a dialogue?
Zendesk positions dialogues as offering more control and procedures as needing less setup. For an action that moves money, a dialogue’s scripted branches make the decision path easier to inspect, though email AI agents must use procedures.
Zendesk is a trademark of its owner. This guide is independent and not affiliated with or endorsed by Zendesk.
Free checklist
Get the guardrail checklist for agents that move money.
A one-page PDF of the controls and the test for each. Download it right away, and we email you the link.
Sources
- Zendesk Help: Getting started with AI agents
- Zendesk Help: Creating generative procedures for AI agents
- Zendesk Help: Technical requirements for integrations with AI agents
- Zendesk Help: Configuring escalation strategies and flows for AI agents
- Zendesk Help: Testing conversation flows in AI agents