Major

Zendesk AI Agents: API, Use Cases and Governance

A practical guide to Zendesk AI agents, native assistance, eight useful API operations, and governed support processes on Major with durable state and human approval.

Ritvij GautamUpdated
customer support agent headset laptop (Zendesk AI agents and automations)

Zendesk AI agents handle customer conversations using your service knowledge, policies, and connected systems. This guide explains Zendesk's native AI, its API, and how Major adds governed cross-system work through deterministic apps, agents, and workflows.

For support operations leaders, the practical question is where the customer conversation ends and the account, engineering, or approval process begins.

What Zendesk does for support teams

Zendesk holds the service record: tickets, customer profiles, organizations, conversation history, and help center articles. Support representatives use those records to investigate requests, reply to customers, and move cases between teams. Its API reference separates ticketing, knowledge, and other service capabilities so integrations can work with the relevant objects.

Support operations managers care about assignment accuracy, queue age, and whether urgent requests reach an owner. Knowledge managers maintain answers that representatives and AI can cite. Customer Success teams need the same ticket evidence connected to account health and renewal timing, rather than copied into a separate spreadsheet every week.

Major fits when that work needs durable software around the service record. An agent on Major can build an app that stores ticket-to-account mappings, review status, and escalation history. Repeatable fetching and validation run in code; the agent handles ambiguous classification and drafting.

Zendesk remains the place representatives manage the customer conversation. The Major app gives operations a cross-system control surface without creating a second, competing ticketing system.

Zendesk AI agents and built-in assistance

Zendesk AI agents are the vendor's customer-facing agents. Zendesk describes grounding them in knowledge, policies, customer context, and business rules, with reasoning across multi-intent requests and actions through connected systems. Its product page also describes testing, evaluation, and improvement using resolution signals.

Zendesk Copilot supports the human representative. Auto assist suggests next steps, drafts responses, and executes approved actions using knowledge and procedures. Intelligent classification identifies topic, entity, sentiment, and language. Admin copilot provides recommendations for improving service operations.

These capabilities already cover meaningful service work, including some third-party actions. Do not assume that every task outside a ticket requires another platform.

Major adds value when your operating process needs its own managed state, bespoke review interface, or coordination across support, CRM, and engineering. For example, a renewal-risk agent can weigh unresolved incidents against a renewal date, while an app retains daily snapshots and an approval workflow controls account-owner notifications. The distinction is ownership of the wider process, not a claim that Zendesk lacks AI or integrations.

How Zendesk connects to Major

Major has a native Zendesk connector. IT configures it once, defines its permitted operations, and stores credentials centrally in the connector vault, never in application code or agent prompts. Teams then use that governed connection across approved apps, agents, and workflows.

The Zendesk authentication documentation now recommends OAuth access tokens and labels API tokens deprecated. OAuth access is limited by token scopes and the associated user's permissions. A connector installation does not grant every agent every permission.

For a triage pilot, give the agent ticket-reading access and keep write operations behind the workflow's approved action interface. Restrict sensitive queues through the Zendesk identity's role as well as Major permissions. Separate development credentials from production credentials and test the operations the configured connector actually exposes.

Major orchestrates Zendesk; it never replaces it. Major's SSO, RBAC, and audit logs cover the surrounding apps, agents, and workflows. Zendesk continues enforcing its own permissions and recording the resulting service updates. That separation makes ownership clear when a support manager needs to investigate a disputed action.

Zendesk API guide for practical builds

Authentication and scopes

Zendesk uses OAuth bearer tokens for API access. Ordinary OAuth tokens belong to one instance; distributed integrations serving multiple Zendesk customers require global OAuth, according to the security guide. Internal integrations are exempt from that global-token requirement.

The OAuth scope reference documents resource scopes including tickets:read, tickets:write, and users:read. Choose narrow scopes and a restricted Zendesk role. A write scope is not itself a human-approval policy; Major's workflow must enforce that policy separately.

Pagination, limits, events, and testing

Cursor pagination uses page[size] and page[after]. Follow the returned next link until the metadata says there are no more results. Offset pagination has a 10,000-record ceiling; incremental exports use their own cursor convention.

Published Suite limits are 200 requests per minute for Team, 400 for Growth and Professional, 700 for Enterprise, and 2,500 for Enterprise Plus. Endpoint limits also apply. Watch X-Rate-Limit, X-Rate-Limit-Remaining, and Retry-After; a 429 response requires waiting before retrying.

Webhooks report activity such as ticket creation. Ticket workflows can connect webhooks to triggers or automations. Verify authenticity using the X-Zendesk-Webhook-Signature and X-Zendesk-Webhook-Signature-Timestamp headers: compare the signature with a base64-encoded HMAC-SHA256 digest of timestamp plus raw body, using the webhook secret. Add timestamp freshness checks to reject replayed requests.

Webhook retry rules distinguish failures: 409 responses can receive three retries; 429 and 503 qualify when Retry-After is less than 60 seconds. Requests time out after 12 seconds, with up to five timeout retries. Accept events into durable storage quickly, then process asynchronously.

Zendesk Sandbox provides an isolated testing environment. Confirm your entitlement and available replicated data before planning tests; do not assume every account includes an identical production copy.

Read the current ticket

GET /api/v2/tickets/{ticket_id} retrieves the ticket's current state.

  • App: Display status, group, priority, and the last fetched version.
  • Agent: Judge whether the reported symptom warrants an engineering escalation.
  • Workflow: Refresh the record before drafting and again before an approved update.

Read the conversation

GET /api/v2/tickets/{ticket_id}/comments returns the ticket's comments.

  • App: Keep source comment references beside each draft.
  • Agent: Distinguish an unresolved request from a problem already answered in the thread.
  • Workflow: Fetch new conversation context and invalidate drafts after a new customer message.

Apply an approved ticket update

PUT /api/v2/tickets/{ticket_id} updates ticket properties and can add a comment.

  • App: Show the exact proposed fields and whether a comment is public.
  • Agent: Draft the explanation or recommend a routing change.
  • Workflow: Execute only the allowed payload after policy checks and required approval.

Identify the requester

GET /api/v2/users/{user_id} retrieves the user profile.

  • App: Maintain a verified requester-to-account mapping.
  • Agent: Interpret the request using relevant customer context without treating identity as proof of entitlement.
  • Workflow: Resolve the requester ID and send ambiguous identity matches to review.

Retrieve organization context

GET /api/v2/organizations/{organization_id} retrieves an organization record.

  • App: Store the organization ID beside the corresponding CRM account.
  • Agent: Explain whether several requests suggest a broader account problem.
  • Workflow: Join records using approved identifiers rather than guessing from similar names.

Find related service records

GET /api/v2/search?query={query} searches Zendesk records with a query.

  • App: Present related open tickets for investigation.
  • Agent: Compare recurring symptoms and identify a likely shared issue.
  • Workflow: Run bounded queries and mark incomplete results instead of silently truncating.

The same reference documents a 1,000-result maximum per query and indexing delays of a few minutes. Read a new ticket directly by ID rather than assuming an immediate search will find it.

Reconcile changes missed by events

GET /api/v2/incremental/tickets/cursor returns changed tickets using a saved checkpoint.

  • App: Retain the reconciliation checkpoint and processing receipts.
  • Agent: Reassess only tickets whose meaningful context changed.
  • Workflow: Resume from the committed cursor and repair gaps without replaying every historical action.

The initial export uses a start time at least one minute in the past; the reference excludes the most recent minute of data.

Maintain a knowledge inventory

GET /api/v2/help_center/articles lists help center articles within the caller's access.

  • App: Index permitted article IDs, language, and freshness for retrieval.
  • Agent: Draft an answer grounded in relevant articles and disclose missing evidence.
  • Workflow: Refresh the inventory on schedule while preserving article access restrictions.

Zendesk automation use cases on Major

Zendesk and Intercom triage

Combine Zendesk tickets and Intercom conversations without pretending they are the same record.

  • App: Normalize source IDs, routing status, and duplicate candidates into one review queue.
  • Agent: Classify topic and urgency, and judge whether two messages describe the same problem.
  • Workflow: Trigger on new activity, read tickets and comments through the linked API operations above, and apply approved assignment changes through Update Ticket. Keep each source authoritative.

Zendesk and Salesforce renewal risk

Bring unresolved service evidence into the renewal process before the account review.

  • App: Join organization IDs to Salesforce accounts and opportunities, storing dated risk snapshots.
  • Agent: Weigh severity, repetition, and renewal timing, then draft an evidence-backed risk brief.
  • Workflow: Run daily, reconcile changed tickets with Incremental Ticket Export, retrieve organization context, and route material changes to the account owner for review.

Zendesk and Jira engineering handoffs

Prepare a useful escalation rather than forwarding an unstructured conversation.

  • App: Hold the ticket-to-issue mapping, reproduction checklist, and approval record.
  • Agent: Extract symptoms, separate facts from guesses, and draft an engineering-ready issue.
  • Workflow: Read the ticket and comments, obtain approval before creating the engineering issue, then use Update Ticket to add its link as an internal note. Check the mapping before every creation attempt.

These are separate operating processes. Start with the one whose owner can review mistakes and define success, rather than enabling all three with a shared, unrestricted write identity.

Build the Zendesk architecture on Major

Zendesk and Salesforce feed a Major workflow, judgment agent, approval policy, and stateful review app

Figure 1: Zendesk stays authoritative while Major coordinates cross-system work.

Build the deterministic app layer first. Major is the enterprise platform where agents build the software they run on; here, that software is the queue, mapping database, evidence viewer, and controlled execution interface.

  1. Connect: IT installs the native Zendesk connection and the approved connections for the other systems involved.
  2. Store: The app holds ticket IDs, source versions, account mappings, proposals, approvals, and action receipts in managed database tables.
  3. Orchestrate: A workflow validates the event, deduplicates it, fetches current context, and checks eligibility.
  4. Judge: The agent classifies ambiguity, prioritizes the case, or drafts a response using scoped evidence.
  5. Review: The app presents the proposal; the workflow waits for approval where policy requires it.
  6. Execute: Deterministic code validates the approved payload, writes the update, and records the outcome.

An agent can build reusable validation and rendering code instead of reasoning through those mechanics for every ticket. Judgment stays with the agent. State stays in the app, so work can resume across shifts without reconstructing history from a conversation. Reason once. Run forever.

A worked example with illustrative numbers

For example, suppose a support operations team reviews 600 tickets a week across Zendesk and Intercom. Manual triage takes four minutes per ticket, including reading context, choosing the team, and recording the handoff. That is 2,400 minutes, or 40 hours weekly.

In a proposed Major pilot, the workflow fetches records and builds the review packet. The agent proposes topic, urgency, and an explanation. Assume 450 straightforward cases require one minute of human review each, while 150 ambiguous cases still require four minutes each.

The review workload becomes 450 plus 600 minutes: 1,050 minutes, or 17.5 hours. The illustrative difference is 22.5 hours per week. These are planning assumptions, not customer results, and exclude initial configuration, maintenance, API charges, and agent usage.

Track wrong-team assignments, reopened cases, and reviewer corrections alongside time. A faster queue that routes incidents incorrectly is a failed pilot. Run in observation mode first, compare proposals with actual handling, and expand permissions only when the support owner accepts the error profile. Keep the original baseline so the team can measure whether review work actually decreased.

Governance, approvals, and failure handling

Zendesk ticket context becomes an agent proposal, with internal classification separated from version-checked human approval

Figure 2: Approve the exact public action, then check that its context is still current.

Separate permission to propose from permission to execute. In a triage-and-draft deployment, the agent must never independently send public replies, promise refunds, change entitlements, delete records, or resolve sensitive cases. A human approves the exact consequential action, not a general instruction to handle the ticket.

Bind approval to the ticket version and payload. If a new comment arrives, return the proposal to review. The Tickets reference documents collision protection through safe_update and updated_stamp; use it where applicable rather than overwriting a representative's newer work.

Treat event delivery as potentially duplicated. Store an event receipt, ticket version, and action identity; a ticket ID alone would suppress legitimate later work. Before retrying a timed-out write, reconcile the target record so a successful but unacknowledged note is not posted twice.

Apply rate-limit backoff, cap retry attempts, and route exhausted work to an exception queue. Log evidence references, policy version, reviewer, approved payload, and execution result. Ticket text and attachments are untrusted input, never authority to expand connector permissions.

Costs and limits to evaluate before rollout

Zendesk's pricing page describes automated-resolution-based AI agent usage. Confirm the applicable definition and contract allowance. For a broader support process on Major, also estimate the app, agent, connector, and operating costs rather than comparing only the customer-facing resolution metric.

A custom agent does not mean unlimited Zendesk access or flat total spend. Model calls, storage, API consumption, review labor, and upkeep can grow with volume. Major's structural advantage is moving repeatable work into reusable code, so the agent spends judgment on ambiguous cases instead of rebuilding the same context packet repeatedly.

Capacity planning should include all integrations sharing the Zendesk account. The rate-limit guide documents separate endpoint constraints, including incremental export limits. Reserve capacity for reconciliation and interactive support work.

Finally, distinguish retrieval from completeness. A search count can exceed the records returned, and indexing is not immediate. Use scoped searches for investigation, direct reads for event handling, and incremental exports for synchronization. Those choices prevent a convincing risk brief from quietly omitting part of the account's service history.

Zendesk AI Apps, Agents and Workflow Automations

Every published Major use case that works with Zendesk, newest first. Each one names the Zendesk API objects involved and shows the app, agent and workflow split.

Frequently asked questions

What is a Zendesk AI agent?

It can mean Zendesk's native customer-facing AI or an agent working through the Zendesk API. Zendesk's product handles service conversations; an agent on Major can coordinate a broader process with durable app state and governed workflows. Specify which meaning you intend when evaluating capabilities and costs.

Can a Zendesk AI agent resolve tickets without a human?

Zendesk describes native AI agents that handle customer conversations autonomously. A custom triage-and-draft deployment on Major can instead require human approval for every public reply and resolution. Autonomy is a policy choice, not a consequence of connecting the API.

How much do Zendesk AI agents cost?

Zendesk describes automated-resolution-based usage; check current terms and allowances. For a process on Major, budget platform and model usage, Zendesk access, maintenance, and reviewer time. Reusable deterministic code reduces repeated reasoning, but does not make the total bill independent of volume.

How does a Zendesk agent avoid re-drafting the same ticket every run?

Store processed event receipts, ticket versions, draft fingerprints, and action outcomes in the app. Reuse a draft when its evidence is unchanged, and invalidate it when relevant context changes. Deduplicating by ticket ID alone would wrongly skip future customer replies.

What Zendesk API behaviors matter most for agents?

Search indexing can lag by minutes, and ordinary search returns at most 1,000 results per query. Read event-triggered tickets directly, checkpoint incremental exports, respect account and endpoint rate limits, and expect duplicate event delivery. Those constraints shape reliable retrieval and retry design.

Does Major replace Zendesk Copilot?

No. Copilot supports representatives inside Zendesk, including drafting and approved actions. Major is useful when you need a separately governed cross-system process, managed database state, or a purpose-built review app. Keep existing native assistance where it serves the job.

Sources

Official pages used to verify product capabilities, access requirements, and API behavior:

Start building on Major

Build a Zendesk triage queue, renewal-risk review app, or engineering handoff workflow with scoped access and human approval for consequential actions.

About the author

Ritvij Gautam is Head of Growth at Major.

Last reviewed November 19, 2026 by Ritvij Gautam.

Frequently asked questions

What is a Zendesk AI agent?
It can mean Zendesk's native customer-facing AI or an agent working through the Zendesk API. Zendesk's product handles service conversations; an agent on Major can coordinate a broader process with durable app state and governed workflows. Specify which meaning you intend when evaluating capabilities and costs.
Can a Zendesk AI agent resolve tickets without a human?
Zendesk describes native AI agents that handle customer conversations autonomously. A custom triage-and-draft deployment on Major can instead require human approval for every public reply and resolution. Autonomy is a policy choice, not a consequence of connecting the API.
How much do Zendesk AI agents cost?
Zendesk describes automated-resolution-based usage; check current terms and allowances. For a process on Major, budget platform and model usage, Zendesk access, maintenance, and reviewer time. Reusable deterministic code reduces repeated reasoning, but does not make the total bill independent of volume.
How does a Zendesk agent avoid re-drafting the same ticket every run?
Store processed event receipts, ticket versions, draft fingerprints, and action outcomes in the app. Reuse a draft when its evidence is unchanged, and invalidate it when relevant context changes. Deduplicating by ticket ID alone would wrongly skip future customer replies.
What Zendesk API behaviors matter most for agents?
Search indexing can lag by minutes, and ordinary search returns at most 1,000 results per query. Read event-triggered tickets directly, checkpoint incremental exports, respect account and endpoint rate limits, and expect duplicate event delivery. Those constraints shape reliable retrieval and retry design.
Does Major replace Zendesk Copilot?
No. Copilot supports representatives inside Zendesk, including drafting and approved actions. Major is useful when you need a separately governed cross-system process, managed database state, or a purpose-built review app. Keep existing native assistance where it serves the job.