01
Overview
Zendesk Custom Agents are Specialized AI Agents built around one company's own processes, policies, knowledge and connected systems. Zendesk positions Agent Builder as the no-code environment for creating, testing and deploying these agents. The useful way to approach Agent Builder is not to start with a generic instruction such as 'handle support.' Start with one bounded job that has a clear outcome, the information needed to make decisions and explicit rules for when a person must take over.
02
Define one job
Zendesk describes a Custom Agent as an agent built for a specific business job. Good starting jobs are concrete workflows such as evaluating a refund request, processing a warranty claim, updating an account, handling an employee IT request or routing a complex service case. Define what successful completion means before connecting tools.
03
Give the agent the right business context
A Custom Agent can use the organization's knowledge, policies, workflow rules and customer or employee context. For a refund agent, that might include refund eligibility rules, purchase history and order status. For an IT agent, it might include device policy, employee information and approved replacement procedures. Keep the source material authoritative and scoped to the job.
04
Connect only the systems the job requires
Zendesk says Specialized Agents can use connected systems to take actions. The launch also says agents can operate in Zendesk or service environments including Salesforce and ServiceNow. The exact integration set depends on the organization's environment. Avoid giving a narrowly scoped agent unnecessary access simply because a connector exists.
05
Define allowed actions
The agent needs an explicit action boundary. For a commerce return workflow, allowed actions might include checking an order, verifying policy eligibility, initiating a return and updating the customer. Zendesk lists workflows such as order tracking, returns and exchanges, refunds, address changes, account updates, subscription management, case routing and escalation. Actual actions depend on connected systems and permissions.
06
Decide where human approval is required
Zendesk explicitly says businesses define where human approval is required. Use this for decisions with meaningful financial, policy or customer-impact consequences. A workflow can let the agent complete routine cases automatically while routing exceptions or sensitive decisions to a human with the conversation context preserved.
07
Configure escalation
A Custom Agent should have a clear stop condition. Zendesk says human agents can step in when expertise, judgment or empathy is needed. Define what triggers escalation, what context is handed over and which actions the AI must not complete on its own.
08
Test before broad deployment
Agent Builder is described by Zendesk as an environment for creating, testing and deploying agents. Test common successful cases, missing-data cases, conflicting-policy cases and exceptions. Verify that the agent uses the correct knowledge, calls only intended systems, respects approval boundaries and escalates when it cannot safely finish the job.
09
Choose how the agent runs
Custom Agents do not have to wait for a customer message. Zendesk says they can run proactively in response to business events or on a schedule. That opens workflows such as checking a service event and initiating a follow-up process. Scheduled or event-driven automation should use the same permission and escalation discipline as conversational use.
10
Measure outcomes rather than chat volume
Zendesk's current AI-agent commercial model centers on automated resolutions and outcomes. For a Custom Agent, useful evaluation should focus on whether the defined job was completed correctly, how often a human had to intervene, where policy or knowledge gaps caused escalation, and whether the automation created downstream rework. Zendesk says its Resolution Learning Loop uses outcomes to improve service, but individual performance claims should not be assumed for a new deployment.
Example: a returns Custom Agent
A narrowly designed returns agent can receive the request, retrieve the order, compare it with the company's return policy, collect any missing information, initiate an eligible return, update the customer and send exceptions to a human. This example follows the workflow Zendesk uses in its Specialized Agents documentation. The exact backend actions depend on the integrations and permissions configured by the business.
11
What Custom Agents are not
They are not unrestricted general-purpose assistants. Zendesk describes them as specialized agents operating within a defined job and business context. They are also different from a fixed decision-tree chatbot: Zendesk says they can reason through context and business rules and take approved actions. That added autonomy makes scope and governance more important, not less.
FAQ
Is Zendesk Agent Builder no-code? Zendesk describes Agent Builder as a no-code environment for building, testing and deploying custom agents.
Can a Custom Agent use company policies? Yes. Zendesk says Custom Agents are grounded in an organization's knowledge, policies, workflows and business rules.
Can a Zendesk Custom Agent take actions? Yes, when the relevant systems and actions are connected and permitted. Zendesk lists examples including returns, refunds, account changes and routing.
Can Custom Agents run automatically? Zendesk says they can run proactively in response to business events or on a schedule.
Can humans approve actions? Yes. Zendesk says businesses define where human approval is required and can escalate to human agents with context preserved.
Does Agent Builder guarantee a particular automation rate? No. Zendesk publishes product and customer performance claims, but a particular deployment's outcome depends on its workflow, knowledge, integrations and controls.
Sources: Zendesk Specialized Agents announcement/product page, Zendesk Custom Agents explainer and Relate 2026 Agent Builder announcement. Recheck the live Agent Builder UI, plan entitlements and connector availability before publication because product access can change.
Sources
Primary and supporting sources
Use the linked documentation to check current pricing, access and product behaviour. The dates above describe this article; they do not guarantee that every vendor claim remains current.