SoftTech AI Registry
Zendesk · AI agents

Actions in AI Agents

Zendesk documents four types of actions that a configured AI agent can perform during a customer conversation: configured actions (conversation actions and CRM actions), API integrations, action flows, and custom actions. Together these let an AI agent read and write Zendesk records such as tickets, users and organizations, and call external systems through admin-defined HTTPS APIs, action-flow steps and MCP tools.

Recorded characteristics

Function
Zendesk documents four action types available to an AI agent in the AI agents workspace. Configured actions (conversation actions and CRM actions) set or modify session parameters, label conversations and exchange conversation details with the CRM; they can be attached at AI agent level (fired on an event in every conversation), use case level, or dialogue block level, or referenced in a procedure. API integrations are the earlier integration-builder mechanism, retained for existing configurations, with custom actions recommended instead. Action flows are admin-built multi-step sequences in the action builder that run steps in Zendesk and external systems; only flows with the Run on demand trigger can be invoked by an AI agent, and the agent receives only the outputs of the flow's last step. Custom actions are single admin-defined HTTPS API calls (GET, POST, PATCH, PUT or DELETE) with declared inputs and outputs that update or fetch data outside Zendesk. Actions are inserted into dialogue blocks by the conversation designer or referenced from generative procedures, where the agent selects among them using their names and descriptions.
Data access
Configured CRM actions for Zendesk Support read user records (including id, email, name, locale, organization id, phone, role, time zone and custom fields), ticket fields including custom fields, ticket tags and organization information, and write by adding tags, adding internal notes, updating ticket fields (assignee id, assignee email, group id, type, priority and status including solved and closed), merging tickets and applying macros. Separate documented CRM action sets exist for Sunshine Conversations and for Zendesk Chat (Legacy). Custom actions send only the inputs the administrator declares to an external HTTPS endpoint and return only the mapped outputs, with a 2MB maximum response size. Action flows can read and write Zendesk ticket, user, organization and custom-object data and call external systems, and MCP tools connected to action flows may transmit data outside the account's selected region.
Actions
Can take actions
External actions
Yes
Human confirmation
Conditional
Permission basis
Admin role
Administrative control
All action types are administrator-configured. Custom actions and action flows are created and managed in Admin Center (Apps and integrations > Actions > Custom actions, and the Action flows page containing the action builder); MCP server connections are created and managed in Admin Center and their tools then appear in the action builder. Configured actions are created in the AI agents workspace and must be explicitly attached at AI agent, use case or block level. An AI agent uses an action only where a designer has placed it in a dialogue block or referenced it in a procedure; for action flows only the Run on demand trigger type is invocable. Authentication is through admin-managed connections rather than credentials embedded in the action. Documented limits: 100 custom actions per account, 100 inputs and 100 outputs each, a 10 second timeout, 100 action flows, 60 steps per flow, and rate limiting on action and flow execution. Custom-action executions can be inspected in the integration log.
Default state
Disabled
Availability
Documented for the current AI agents workspace. Zendesk announced the integration of the action builder with AI agents on 13 May 2026, with rollout from 13 May to 12 June 2026, making custom actions and action flows invocable from the AI agents workspace. The plan tables on these pages state: configured actions and actions in AI agents on all Suite plans (Team, Growth, Professional, Enterprise, Enterprise Plus) and Support Team, Professional or Enterprise; custom actions, action flows and MCP connections on Suite Growth, Professional, Enterprise or Enterprise Plus and Support Team, Professional or Enterprise.
Licensing
No separate SKU is documented for actions themselves. The plan tables cited above set the entitlement. Action flows consume action credits in other contexts, but Zendesk documents that action flows invoked from an AI agent do not consume action credits; MCP tool steps do consume action credits when executed within an action flow.
External model or provider
Not established. Zendesk documentation reviewed does not name a model or model provider for the AI agent's selection or execution of actions.
Limitations and uncertainty
Boundary: this record covers the action-execution framework available to configured Zendesk AI agents, not Agent Copilot auto assist, not custom agents, and not the separate action builder used outside AI agents. Human confirmation is recorded as conditional because Zendesk documents no confirmation gate specific to AI-agent-invoked actions, while a conversation designer can place confirmation steps in a dialogue or procedure before an action block; an explicit platform-level approval gate for these actions is not established. Permission basis is recorded as admin_role because actions execute through admin-created connections and admin-configured endpoints rather than the end customer's permissions; Zendesk warns that a connection may have higher access privileges than agents or end users. External action capability is recorded as yes based on documented custom actions writing to external APIs and action flows calling external systems and MCP tools. Not established: an action-execution audit log distinct from the integration log and Connections metadata; whether ticket creation at escalation is an agent-selected action; and the precise runtime identity used for Zendesk-side writes.

Evidence