AI and Zendesk.

What Zendesk’s built-in AI does today

Zendesk includes AI agents in every Suite and Support plan and bills them per automated resolution, a request the AI agent resolves without a human (Zendesk pricing, checked 19 October 2026). Since 18 May 2026 those resolutions are tiered: an assisted escalation and a contained resolution do not count against your allowance, while a verified resolution, where an LLM reviews the conversation and confirms the request was resolved, does (Zendesk Help).

For human agents, the Copilot add-on classifies requests by topic, sentiment and language, suggests next steps, drafts responses and runs approved actions from procedures written in your knowledge base. It lists at US$50 per agent per month, on top of plans such as Suite Team at US$55 or Suite Professional at US$115 per agent per month, billed yearly. Zendesk bills only in USD, EUR, GBP and BRL.

Integration routes and when to use each

If you want Copilot or Zendesk AI agents to act in your own order or billing system, our guide to MCP for business systems covers building a narrow server for them to call.

Automation 1: a draft on every ticket, as an internal note

Trigger

A Zendesk trigger calls your webhook when a ticket is created or the requester replies.

Inputs

Subject, the latest public comment, the requester’s organisation, tags, custom fields, and matching help centre articles.

Model step

Classify intent and urgency, then draft a reply that cites one or two help centre articles.

Checks in code

No refunds, credits or dates beyond what policy allows; every linked article exists.

Write-back

An internal note with public set to false, the intent in a custom field, and tags such as ai_draft.

Human review

The agent edits and sends the reply as their own public comment.

Measures

Drafts sent with light edits, handle time, reopen rate.

Our lesson on using AI for customer support drafts covers draft quality, and human review queues covers keeping reviewers honest.

Automation 2: escalations a lead confirms

1. Score. On each requester reply, rules check for VIP organisations and phrases such as “Disputes Tribunal” or “lawyer”; the model scores safety concerns, vulnerability and churn risk.
2. Route. Above a threshold, set the ticket to urgent, assign the Escalations group, and add an internal note explaining why.
3. Confirm. A team lead confirms or downgrades within your SLA.
4. Tune. False positives are tagged and reviewed monthly to adjust rules and prompts.

Permissions, scopes and approval design

Use an OAuth client with resource scopes, not the broad read, write or impersonate scopes. The drafting service needs tickets:read, tickets:write and hc:read. Zendesk defines tickets:write as including “adding comments” (OAuth scopes), and a comment’s visibility follows the ticket’s initial setting unless you change it (ticket comments). So the same token that writes a private note can post a public reply. Set public to false on every note in code, and test it. Run the integration as a dedicated user so notes are clearly attributed. Our guide to agent permissions and approvals covers why sending stays with people.

Costs and limits

Zendesk Support and help centre API limits are 200 requests a minute on Team, 400 on Growth and Professional, 700 on Enterprise and 2,500 on Enterprise Plus (rate limits). A draft takes three or four calls, so limits rarely bind. Illustrative example: 10 agents on Suite Professional cost US$1,150 a month, and Copilot adds US$500. A custom drafter handling 3,000 tickets a month at 5,000 input and 400 output tokens costs about US$42 on Claude Sonnet 5.5 (US$2 and US$10 per million tokens), plus build and support, at list prices checked on 19 October 2026.

Where it goes wrong

• Trigger loops. Your note updates the ticket, which fires the trigger again. Condition triggers on updates by the requester and on the ai_draft tag being absent.
• Leaked drafts. A note posted as public emails the customer. Zendesk can make a comment private afterwards, but the email has already gone.
• Quoted history. Email replies repeat the whole thread; draft from the latest comment.
• Billing disputes. LLM-verified resolutions decide what you pay. Sample them yourself each month.
• Manual recovery. Drafts are private notes, so a bad batch costs agent time, not customer trust. Keep a single switch that disables the trigger.

Compare AI and Intercom, whose Fin agent is priced per outcome. For shared help with drafting, escalation and evaluation, see our AI customer support systems.

Written by the Levy Street team in Wellington. Last updated 19 October 2026.