Automate Support Ticket Triage Across Zendesk and Intercom
Turn scattered Zendesk tickets and Intercom conversations into one governed triage queue, with AI judgment for routing and deterministic app logic for repeatable work.

TL;DR: what you'll build. A shared triage queue that pulls in every new Zendesk ticket and Intercom conversation, lets an agent propose a topic, urgency level, and owning team, and routes the routine cases automatically while flagging the rest for a person. The Support Operations Manager keeps Zendesk and Intercom as the system of record for each channel, but gets one governed view of what is open, who owns it, and why it was routed that way. The result: fewer missed signals between the two tools and a routing decision you can actually audit.
Why support teams connect Zendesk and Intercom
Many support orgs run Zendesk for ticketing and Intercom for live chat and in-product messaging. The same customer problem can show up as a Zendesk ticket from an email, and as an Intercom conversation from the in-app widget, often without anyone noticing they are related.
A Support Operations Manager owns the queue health across both. That means triage: assigning topic, priority, and the right team fast enough that nothing sits untouched during a shift change or a weekend.
The failure mode is not that Zendesk and Intercom cannot connect. Both vendors document integration paths. The failure mode is that most integrations move data between the two tools without building a real operating model: a single source of truth per record, deduplication so the same issue is not routed twice, and a place for a human to step in when the automation is unsure.
This article is a zoomed-in build inside Major's broader pattern for support automation, covered at the AI customer service agent hub, which rolls up every support use case on Major. If HubSpot, Salesforce, or another CRM also touches this account, see how Zendesk ticket workflows feed renewal-risk scoring for a related pattern that reads Zendesk tickets alongside CRM data.
Major's fit here is specific: the deterministic parts of triage (normalizing records, deduplicating, applying routing policy, writing back to each system) run as a governed app, and the agent is reserved for the one judgment call that actually needs it, classifying an ambiguous or high-stakes case.
The systems involved
Zendesk: tickets, organizations, and tags
Zendesk's Ticketing API is documented at the Zendesk API reference, with the Tickets object covered in the Tickets endpoint reference.
A ticket carries the fields a triage queue needs: subject, description, priority, status, the requester, the assigned group, and tags. Zendesk also exposes Organizations (the account record) and a Tags object for the category labels a trigger or macro would otherwise set by hand.
Authentication uses OAuth 2.0 bearer tokens (the vendor's recommended method) or a legacy API token with Basic auth, as described in Zendesk's security and authentication docs. Zendesk supports native webhooks for ticket events, so a new or updated ticket can push to your app instead of your app polling for it.
One gotcha worth planning around: Zendesk rate limits are plan-based. The saved API notes put Team plans at 200 requests per minute, Suite Growth and Support Professional at 400, Enterprise at 700, and the High Volume API add-on at 2,500, per Zendesk's rate limits page. A 429 response includes a Retry-After header your workflow must respect.
A second gotcha: Zendesk's Ticketing API supports an Idempotency-Key header on ticket creation. Reuse a key and Zendesk returns the original ticket instead of creating a duplicate, until the key expires after two hours. That is the mechanism to lean on for retry-safe writes.
Intercom: conversations, tickets, and tags
Intercom's REST API is documented starting at its API reference introduction. Two objects matter here: Conversations, which hold the messaging thread with a contact, and Tickets, Intercom's own structured-issue object with a documented create-a-ticket endpoint (a POST request).
Intercom also exposes Tags for conversation and ticket labeling, and Data Connectors, a built-in mechanism for triggering actions in other systems. Intercom's own Data Connector guide documents how to create a Zendesk ticket directly from an Intercom conversation, with field mapping and authentication set up inside Intercom.
Authentication for custom integrations uses a workspace access token for private apps, or OAuth 2.0 bearer tokens for public apps. Intercom supports topic-based webhook notifications (Webhook Models) so a new conversation or ticket update can push an event rather than wait on a poll.
Intercom's documented default rate limit is 10,000 requests per minute per app or workspace, with a 25,000-per-minute workspace-wide cap. That is generous for most support volumes, but still worth building backoff for, since a burst of inbound conversations during an outage can spike request volume quickly.
A gotcha specific to Intercom: its ticket-creation guide documents a one-way creation flow (an Intercom conversation can create a Zendesk ticket). It does not document a native two-way, always-on sync of every field. Treat that distinction as a design decision, not an afterthought: decide what field changes flow back, and build that direction explicitly rather than assuming it is symmetric.
Where customer data privacy applies
Both systems store customer personal data in ticket and conversation records. Under GDPR and CCPA, a support platform holding that data must support access and deletion requests within statutory windows, generally within one month for GDPR requests. See the GDPR text for the underlying rule.
The practical requirement for this build: your triage app's queue table should reference records by source system ID rather than copying full customer PII into a third store, so a deletion request only has to reach two systems, not three.
Architecture: app, agent, and workflow
Three pieces do different jobs here, and keeping them separate is what makes the system both fast and auditable.
The workflow is deterministic. It fires on a Zendesk ticket webhook or an Intercom conversation webhook (or a short polling interval if webhooks are not configured), fetches the full record, checks for an existing linked row by source ID, and either creates a new queue row or updates the existing one. After the agent classifies a case, the workflow applies routing policy and writes the result back to Zendesk and Intercom. None of this needs a model; it is the same steps every time.
The agent handles one job: given a normalized ticket or conversation (subject, body, requester history, prior tags), it proposes a topic, an urgency level, a target team, and a short rationale, with a confidence score. That is the one step that benefits from judgment, since free-text support requests do not sort themselves into categories with simple rules.
The app is the control surface. It holds the normalized triage queue table, the routing policy rules, the audit log of every decision, and the review screen the Support Operations Manager uses to approve or override. This is where state lives: the workflow and agent are stateless between runs, but the app remembers.
Think of the workflow as the conveyor belt and the agent as the one inspector standing at the one point on the belt where judgment is actually needed. Everything else keeps moving without a person watching it.

Zendesk and Intercom support triage architecture on Major
Build it on Major, step by step
- Connect the Zendesk and Intercom connectors in Major, using each platform's official connector. You will authenticate Zendesk with an OAuth token or API token and Intercom with a workspace access token or OAuth token, and Major's credential proxy stores both so no key sits in application code. You should see both connectors listed as active with their scopes shown.
- Build the triage app with a
triage_queuetable holding: source system, source record ID, requester or customer reference, subject, topic, priority, assigned team, status, last synced time, decision rationale, approval state, and error state. You should see a table view where each row represents one customer issue, regardless of which tool it came from. - Add the agent with a prompt that gives it the normalized record and a list of valid teams and priority levels, and asks for a structured classification with a confidence score. Try a prompt like:
"Given this support record (subject, body, requester plan tier, prior ticket count), return a topic from this list, an urgency of low/medium/high/urgent, the best-fit team from this list, a one-sentence rationale, and a confidence score from 0 to 1. If confidence is below 0.7 or urgency is high or urgent, set needs_review to true."
- Set the workflow trigger. Prefer the Zendesk ticket webhook and the Intercom conversation webhook for near-real-time intake; fall back to a short polling schedule (every few minutes) only where webhooks are not configured in your environment. You should see the workflow log a new row each time a ticket or conversation arrives.
- Test with a small batch of real but low-stakes tickets and conversations in a sandbox workspace. Confirm the dedupe logic correctly links a Zendesk ticket created from an Intercom conversation back to the same queue row instead of creating two. You should see one queue row per customer issue, not two.
- Roll out with permissions scoped to the support team's role, so only Support Operations Managers and leads can approve routing overrides or edit policy rules, while front-line agents can view the queue read-only. You should see role-based access reflected in who can click approve versus who can only view.
For the agent's classification step, a second prompt worth testing is a review-assist version:
"Summarize why this case was flagged for review in two sentences, referencing the confidence score and any conflicting signals between the Zendesk ticket priority and the Intercom conversation tags."
Governance and IT control
Every piece of this build sits behind the same controls IT already expects from a production system.
- SSO: the triage app inherits the organization's single sign-on, so access follows the same identity provider as every other Major app.
- Role-based access: view, edit, and approve permissions are set per role, so a front-line agent cannot silently change routing policy.
- Scoped credentials: Zendesk and Intercom credentials live in Major's credential proxy, scoped to only the operations each needs (read tickets and conversations, update ticket group and conversation team, and create or link records), never embedded in agent prompts or workflow code.
- Audit log: every routing decision, including the agent's rationale and confidence score, is written to the app's audit table, so a disputed routing call can be reconstructed after the fact.
- Managed database and storage: the queue table and its history live in Major's managed database, not in a spreadsheet or a model's context window that disappears after a run.
- Versioned deploys: changes to the routing policy or the agent's prompt go through a version history, so a bad policy change can be rolled back.
Any case that would change who owns an urgent customer issue, or that falls below the confidence threshold, requires the human approval step before the write-back runs.
Failure handling and edge cases
Rate limits and transient errors are the most common failure a sync like this hits. The workflow should back off and retry on a Zendesk 429 using its Retry-After header, and apply the same pattern against Intercom's documented limits even though they are higher.
Duplicate events are the second most common issue. A ticket can fire multiple webhook events in one minute (status change, then tag added, then comment). The workflow should upsert by source ID rather than insert, and use Zendesk's idempotency key on any ticket creation it triggers.
Partial syncs happen when one system's write succeeds and the other fails midway. Queue each write-back as its own retryable step with its own error state in the triage_queue row, so a Zendesk write succeeding while an Intercom tag update fails does not silently lose the second half.
Stale data is a quieter risk: if the agent classifies a case using a ticket snapshot that is a few minutes old, a new comment may have already changed the picture. Re-fetch the current record right before applying the routing decision, not just at intake.
The agent must never auto-close a ticket, never silently merge two customer records, and never route a flagged case without a person's sign-off. Those are the specific actions that stay on the human side of the line.

Human approval path for uncertain or high-priority triage cases
What to measure
- Time to triage: minutes from ticket or conversation creation to a team assignment, which is the number this whole build is meant to shrink.
- Review override rate: the share of agent classifications a person changes, which tells you how well the confidence threshold is calibrated.
- Duplicate rate: how often the same customer issue would have created two separate queue rows without dedupe logic.
- Reassignment rate after routing: how often a routed case gets bounced to a different team, a sign the routing policy needs tuning.
- Exceptions caught: the count of cases the review queue flagged before they reached the wrong team or sat unassigned.
Before you go live
- [ ] Zendesk and Intercom connector scopes confirmed to the minimum needed (read tickets/conversations, update group/team, create/link records)
- [ ] Idempotency key and dedupe-by-source-ID logic tested with a duplicate webhook event
- [ ] Confidence threshold for human review tested against a sample of real historical tickets
- [ ] Role-based access set so only leads can approve or edit routing policy
- [ ] Retry and backoff behavior tested against a simulated 429 from each system
- [ ] Audit log confirmed to capture the agent's rationale and confidence score on every decision
- [ ] Owner named for the review queue during each shift
- [ ] Alerting configured for repeated write-back failures to either system
More customer support & success AI automation use cases
- AI agents for customer support & success: every use case by tool
- Automate renewal-risk alerts with Salesforce Opportunity and Zendesk ticket data
Start building on Major
Build a shared Zendesk-Intercom triage app that keeps record history, routing decisions, and human approvals in one governed workflow.
Related articles
- Salesloft Salesforce Integration: Flag Stalled Opportunities
- Automate Shopify-ShipStation Delay Alerts with AI
- Datadog PagerDuty Integration: Automate Incident Triage
About the author
Ritvij Gautam is Head of Growth at Major.
Last reviewed November 8, 2026 by Ritvij Gautam.
Frequently asked questions
- How do you integrate Zendesk and Intercom?
- Use each platform's official API or a built-in connector, such as Intercom's Data Connectors, to move ticket and conversation data between the two. Choose one system as the source of truth per record type, keep the originating ID on every synced record, and decide explicitly which field changes flow in each direction rather than assuming a full two-way sync.
- Can an Intercom conversation create a Zendesk ticket?
- Yes. Intercom's official Data Connector guide documents creating a Zendesk ticket directly from a conversation, including field mapping and authentication setup inside Intercom. Ongoing updates to that ticket need separate configuration; the documented flow is one-way creation, not a continuous two-way sync.
- Which system should be the source of truth?
- Pick based on which team owns the customer relationship for that channel: Zendesk for email and ticket-based support, Intercom for live chat and in-product messaging. Retain the originating record ID on every synced row, and avoid letting both systems write to the same field without a clear precedence rule.
- How do you prevent duplicate tickets when syncing?
- Store the source system and source record ID on every queue row, and upsert on that pair instead of inserting blindly. Zendesk's Ticketing API also supports an idempotency key header on ticket creation, which returns the original ticket if the same key is reused within its two-hour expiry window.
- Can AI route support tickets safely?
- Yes, if the model's role is bounded to classification rather than final action. Have the agent propose a topic, urgency, and team with a confidence score, apply deterministic policy rules for routine cases, and require human review for low-confidence or high-urgency cases, with every decision and its rationale logged for audit.