Zendesk AI Agents governance guidance

Reviewed 3 October 2026. Aicura is not affiliated with Zendesk.

About Zendesk AI Agents

Zendesk AI Agents answer customers directly across messaging, email and voice, drawing on the organisation’s help centre and other knowledge sources. Their reach grows as the deployment matures. They start by answering common questions, then follow procedures written in plain language and go on to take actions in the organisation’s own systems, such as looking up an order or changing an account, before handing over to a person when needed. Because they speak to customers on the organisation’s behalf, good governance is what lets an organisation automate more of its service with confidence.

Where governance makes the difference

• Customers receiving the wrong answer or outcome. An AI agent answers at volume and without a person reviewing each reply. An outdated policy article or a loosely written procedure can reach many customers before anyone notices, affecting their access to services and the organisation’s reputation.
• Actions reaching further than intended. Once an AI agent can call the organisation’s own systems, it can read and change customer records. Without clear limits and an owner, it can end up doing more than anyone set out to allow, or acting for a customer whose identity has not been properly confirmed.
• Customers steering the agent. An AI agent follows what it reads, so a customer message written to manipulate it can push it towards revealing information or taking an action outside its purpose.

---

For the executive team

These agents speak for the organisation

Every reply a Zendesk AI Agents deployment sends is the organisation talking to its customers. That brings real gains in speed and coverage, and it also means mistakes reach customers directly. Leadership needs to decide which conversations the organisation is comfortable automating and where a person must stay involved.

Each agent and each action needs an owner

The value grows as agents are allowed to do more, such as process refunds, change bookings or update details. Each agent needs a named person accountable for what it says and does, with agreed limits on what it may change for a customer.

Oversight needs to keep pace with change

New procedures, new knowledge sources and new connections to business systems are added over time, and the AI behind the agents can change too. The deployment approved at launch will not be the one running months later, so oversight works best as an ongoing practice rather than a one-off sign-off.

Questions worth putting to management

• Who is accountable for Zendesk AI Agents, and who owns each agent and each action it can take?
• Which customer requests are handled end to end without a person, and how was that boundary chosen?
• How does the organisation confirm a customer’s identity before an agent changes anything for them?
• How would the organisation know if agents were giving wrong answers or customers were being turned away when they should not be?
• What do customers see when they want to reach a person?
• If a regulator, customer or insurer asked how these agents are governed, what would be shown to them?

What good reporting looks like

Expect regular reporting on how many conversations agents resolve and hand over, the quality of their answers and customer satisfaction, complaints and incidents linked to them, changes to what they are allowed to do and the significant decisions taken, with who made them and why.

How Aicura can help

• One register showing every Zendesk AI Agents deployment, who owns each agent and the business systems it can act on.
• Board reporting drafted from the governance record, covering automation, quality, incidents and the decisions taken.
• A public trust centre where customers can see how the organisation uses AI in its service, backed by evidence a third party can verify.

---

For risk, compliance and assurance

Set where customer-facing automation sits in the risk appetite

Decide which requests agents may resolve on their own, which need a person to approve and which must always go to a person. Requests that affect a customer’s access to a service, money or account security deserve a lower threshold for automation than general questions.

Assess each agent on its own terms

Assess each agent separately, covering the channels it serves, the knowledge it draws on, the actions it can take, how it confirms who it is dealing with and how it hands over to a person. An agent that answers product questions carries a very different risk from one that changes customer accounts.

Reassess when anything changes

A new procedure, a new action, a new channel, a new knowledge source or a change to the underlying AI all move an agent away from what was assessed. Each needs to trigger a fresh assessment rather than wait for the annual cycle.

Make the data decisions explicitly

Decide what customer information agents may collect, use and pass to connected systems, and how long conversation records are kept. Read the vendor’s published information on how service data is used for AI features, including any use of the organisation’s own data to tailor AI to the account, and decide which features to enable on that basis. Tell customers they are dealing with an AI agent and how their information is handled.

Confirm which protections apply

Check which vendor commitments on AI data use, hosting location and data deletion apply to the organisation’s plan and region, and whether they meet its own obligations. Revisit this when the vendor updates its AI data use information or the organisation adds AI features.

Set limits and name who acts

Agree the limits that matter, such as falling resolution quality, rising complaints, unusual action volumes or a jump in handovers, and name the person who acts when one is crossed. Decide in advance when an agent or an action is switched off.

Be ready to show governance to others

• Policy. Maintain an AI use policy that covers customer-facing agents specifically, including which requests may be automated, how identity is confirmed and how customers reach a person. Tie each commitment to the control that carries it out.
• Accountability. Record decisions about what agents may do with their reasons when they are made. Have agent owners confirm periodically that their agents still match what was approved.
• Operating evidence. Keep records that controls ran, such as quality reviews of conversations, approvals of new procedures and actions and interventions when limits were crossed.
• External scrutiny. Prepare for internal and external audit, regulator enquiries about automated customer treatment and insurer questions about AI controls.
• Customers and partners. Business customers increasingly ask in due diligence how AI handles their data and their end users. Have the answer ready and supported by evidence.
• Incidents. Be able to reconstruct which agent, which procedure, which knowledge and which actions were in place when a customer was affected.

How Aicura can help

• Risk and impact assessments for each agent against recognised frameworks, run again whenever a procedure, action or channel changes, feeding a managed set of risks and controls that can mirror to the GRC platform.
• Decision disclosures for customers drafted from the register and reopened when an agent changes.
• Limits on agent activity with responses the organisation chooses, from alerting the owner to re-scoping or pausing an agent.
• Approvals, owner attestations and every control action kept as evidence that can be verified without relying on the organisation’s word, with incident handling linked to the agent involved.

---

For technology and security

Discover every form Zendesk AI Agents take

Inventory every AI agent across messaging, email and voice, including agents serving other platforms beyond Zendesk. Record for each one its knowledge sources, its procedures, the systems it connects to and the credentials those connections use. Include the other AI features enabled in the Zendesk account in the same picture.

Prepare knowledge before agents rely on it

Agents answer from the help centre and connected knowledge sources, so stale or conflicting articles become wrong answers. Review the content agents draw on, remove what is out of date and keep internal-only material out of their reach. Expand coverage in stages, starting with well-maintained topics.

Put configuration changes under change control

Treat new procedures, actions, connections and knowledge sources as changes that need approval before going live. Where configuration can be changed through an API or pipeline, run a preflight check on every change so it is tested against agreed rules first. Use the account audit log to confirm that what is running matches what was approved.

Control who can build and enable agents and actions

Limit who can create agents, write procedures and add actions or connections to business systems. Give each connection the narrowest access it needs, keep its credentials in the organisation’s own secrets management and rotate them on a schedule.

Govern the agents’ calls into business systems

Calls from an agent to the organisation’s own systems arrive at endpoints the organisation provides, so they can pass through an AI gateway before reaching those systems. Use it to limit agents to approved actions, redact or block personal information returned to the agent, screen for instructions carried in from customer messages, meter and cap activity and cut off an agent’s access when needed. Require confirmed customer identity before any action that changes a record.

Rely on vendor settings and observation for the rest

The agents’ own AI processing runs in the vendor’s cloud and cannot be reached by a gateway. Govern that part through the vendor’s settings, careful procedure design, quality review of conversations and the contract. Staff traffic from managed devices can still be managed at the network edge.

Feed everything the agents expose into monitoring

Bring the AI agent conversation data export, the account audit log and event notifications into monitoring, apply the limits risk has agreed and set responses ranging from alerting the owner to restricting an action or stopping an agent.

Test agents against the live deployment

Red-team and evaluate agents before release and on a schedule afterwards, using realistic customer messages, including attempts to manipulate the agent, against the deployed agent with its real knowledge and actions. Watch for changes in behaviour between tests.

How Aicura can help

• One inventory of every agent with its knowledge sources, actions and the identities it presents to business systems, with discovery of AI in use that is not yet recorded.
• Preflight checks before a procedure, action or connection goes live, with every answer recorded.
• A gateway in front of the business systems agents call, with guardrails covering approved actions, redaction, injection screening, rate limits and the ability to cut off an agent.
• A collector inside the network for conversation exports and audit records that flags actions an agent was never declared to take, plus scheduled red-teaming and evals against the live deployment.

---

Getting lasting value from Zendesk AI Agents

Zendesk AI Agents pay off when they are governed as an ongoing practice rather than approved once at launch. Three things matter whatever the role. The first is knowing every agent, every action it can take and who is accountable for each. The second is keeping knowledge, limits and assessments in step with a deployment that keeps doing more for customers. The third is being able to show, at any time, how decisions were made and that controls held. Organisations that get these right can automate more of their service, with confidence rather than caution.

---

Sources

• Zendesk, AI for customer service
• Zendesk, AI agents
• Zendesk, Trust Center
• Zendesk, Zendesk platform
• Zendesk Help, AI data use information
• Zendesk Help, Viewing the audit log for changes to your account
• Zendesk Developer Docs, AI agents
• Zendesk Developer Docs, Using the AI Agents Data Export API
• Zendesk Developer Docs, Creating a custom CRM integration