To run an AI agent inside Zendesk day to day, a support ops manager needs seven habits: classify every ticket on arrival, build views that catch escalations, define what the agent hands to people, QA the agent's conversations daily, tune knowledge and workflows weekly with a test before each change, report on automated resolutions and escalation rate, and watch the bill. Zendesk's own AI agents, intelligent triage and Zendesk QA cover much of this natively; add-on agents such as Lorikeet or Fin (now part of Salesforce) work inside the same tickets.

This guide is the operating rhythm, not the installation. If you have not connected an agent yet, start with our guide on how to add an AI agent to Zendesk.

Key takeaways

• Triage is the foundation. Zendesk intelligent triage classifies tickets by topic, sentiment, language and entities, on Suite and Support Professional plans and above; using those fields in workflows needs the Copilot add-on.

• Escalations deserve their own views. Zendesk suggests views by topic, language and sentiment, and watching group transfers as a signal that a ticket needed escalating.

• QA the AI like a teammate. Zendesk QA detects Zendesk AI agents on messaging and voice and scores them on scorecards; it requires the QA add-on.

• Know your billing unit. Zendesk bills AI agents per automated resolution (a request resolved without escalation to a human), with 5 or 10 included per agent per month depending on plan and $2.00 pay as you go.

• Self-service rarely finishes the job alone. Gartner found only 14% of service issues are fully resolved in self-service, so the human queue still needs care.

How we chose and checked these steps and tools

The steps reflect the recurring work in a Zendesk instance with an AI agent live: intake, routing, review, change and reporting. Every Zendesk fact comes from zendesk.com, including the Zendesk pricing page and Zendesk's help centre at support.zendesk.com, checked on 23 October 2026. Other vendor facts come from each vendor's own pages, and market facts from Gartner, the CFPB and NIST. Lorikeet publishes this guide and is one of the tools listed. We have tried to describe every product, including our own, using only what each vendor publishes, and we link to the source for each claim.

1. Classify every ticket on arrival: topic, sentiment, language

An agent you cannot route around is an agent you cannot manage. Turn on intelligent triage so new tickets are classified by topic, sentiment, language and entities such as product names. Zendesk classifies from the ticket subject and first public comment, writes the results to standard fields with a confidence field for each, and lets agents correct values. You choose which channels it runs on, whether classifications update as customers reply, and whether to exclude agent-created tickets.

Two plan details matter for budgeting: classifications are available on Suite and Support Professional plans and above, and using them in triggers, views logic or SLAs requires the Copilot add-on. If you bought Copilot before 11 June 2026, you may still see "Intent" rather than "Topic" in fields and APIs until later in 2026, so check your trigger conditions after the rename.

Daily habit: scan low-confidence classifications each morning. A spike usually means a new product issue the agent has not been trained on, and those tickets should go straight to people. If you run a third-party agent, its own routing tags can sit alongside Zendesk's; the tools table below shows what each vendor charges for that.

2. Build escalation views your team can work from

Zendesk's guide to views for triaged tickets suggests views for specific topics (such as unassigned refund requests for the billing team), specific languages and upset requesters. It also notes that views suit small teams, while larger organisations should route on triage fields with omnichannel routing, skills-based routing or triggers.

For an AI agent, build three views at minimum: tickets the agent escalated today, negative-sentiment tickets the agent touched, and tickets on sensitive topics (complaints, disputes, health or account security) whatever their status. Zendesk's article on identifying escalations lists group transfers as a pattern worth watching, because repeated transfers of a ticket type often mean it needs a specialist from the start, and suggests triggers that raise a ticket's priority to meet SLA requirements when an issue is often urgent.

Daily habit: work the escalation view first, before the general queue. The customer has already spoken to a machine once.

3. Define the handoff contract between agent and team

Write down what every escalation must carry: a summary, the customer's verified identity status, what the agent already did, and why it stopped. Then check the agent actually does it. With Lorikeet, the Zendesk integration works through a Zendesk webhook and triggers that fire on a new ticket, a customer reply, an internal note or a status or assignment change, using a JSON template so the payload carries ticket metadata and the latest conversation. Lorikeet sees ticket context and history, drafts or sends replies on your Zendesk channels and hands off to people, and you can start conservatively before expanding automation.

Whatever agent you run, put the reason for escalation in a ticket field, not only in an internal note. Fields can drive views, SLAs and reports; notes cannot. The CFPB's warning about chatbot "doom loops" with no offramp to a human (CFPB) is the failure this step prevents.

4. QA the agent's conversations every day

Sample-based QA was built for human teams. For an AI agent, review more. Zendesk QA automatically detects Zendesk AI agents on messaging and voice, lets you include or exclude bots from reviews, score them on scorecards with autoscoring or manual review, and use a BotQA dashboard to see how often bot conversations were escalated to humans. Its spotlight insights flag tickets with events such as churn risk, escalation requests, sentiment, SLA breaches and bot issues.

If your agent is not Zendesk's own, check whether its QA covers every ticket. Lorikeet's Coach QA scores every ticket the AI handles as Good, Warning or Critical and rolls results into a Ticket Quality Score by workflow; human tickets can be scored once enabled. If you rate a resolution bad, it is not counted as automated and not charged. easykind used Lorikeet to tag every incoming email Tier 1, 2 or 3 into the right queue, and cut email response times 92%, from 24 hours to roughly two (easykind story).

Daily habit: read every Critical or low-scoring conversation from yesterday and tag the cause: missing knowledge, wrong workflow, or a policy that should have escalated.

5. Tune knowledge and workflows weekly, and test before you ship

Most AI agent misses come from content, not the model. Gartner found that in 43% of failed self-service cases customers could not find content relevant to their issue (Gartner). Turn the week's QA tags into a ranked list of knowledge gaps and workflow fixes, and change one thing at a time.

Before a change goes live, replay it. Lorikeet tests on replays of your historical tickets and synthetic scenarios in bulk before deployment (how it works). Firm360 runs a knowledge-gap analysis every two weeks across its AI agent's MCP server, Slack and recent transcripts, publishes new articles, and Auto QA keeps reviewing them (Firm360 story). On Zendesk's side, admins can monitor the reasoning steps an AI agent took and give feedback at each step, according to the pricing page feature list.

6. Report on automated resolutions, escalation rate and quality

Pick a small, stable set of metrics and report them weekly: automated resolutions, escalation rate by topic, QA score, customer satisfaction on AI-handled conversations, and reopen rate. Zendesk's AI agent analytics dashboards report performance and BSAT, can be filtered by time and other attributes, and export to CSV for your BI tool. Its intelligent triage dashboard in Explore shows volume, trends, enrichment rates and response and resolution times by topic, sentiment and language; Zendesk began collecting that dataset on 3 August 2024.

Agree definitions before you show a trend. Zendesk counts an automated resolution only when the request was resolved "without any escalation to a human agent". Other vendors define resolution differently, so do not compare raw rates across tools.

7. Watch cost and capacity every month

Each Zendesk plan includes automated resolutions: 5 per agent per month on Support Team and Suite Team, and 10 on Suite Professional, up to 10,000 a year; beyond that you commit to a volume or pay $2.00 per resolution as you go. Seats are listed at $19 (Support Team), $55 (Suite Team) and $115 (Suite Professional) per agent per month billed annually (Zendesk pricing). Copilot writing tools and ticket summaries are pooled at 5 uses per agent per month, capped at 500 per instance.

Compare that with the hours your team now spends on escalations. Gartner predicts that by 2027, 50% of organisations that planned big service headcount cuts because of AI will abandon those plans (Gartner). Budget for people and AI together.

Tools that help

Our comparison of AI agents for Zendesk and Zendesk AI vs Lorikeet go deeper on choosing between them, and AI auto QA: grading every ticket covers QA in detail.

What still needs a human

Keep people on complaints, refunds outside policy, account security, anything regulated (financial hardship, clinical questions, legal threats) and any customer who asks for a person. Customer patience for AI is limited: Gartner found 64% of customers would prefer companies did not use AI in customer service (Gartner). The ops manager's job is to make that handoff fast and well-informed, and to own change control: who approves a new workflow, who signs off a knowledge change, and how you roll back. The NIST AI Risk Management Framework is a useful, voluntary reference for writing those controls down.

Questions to ask as you run an AI agent inside Zendesk day to day

• Which tickets did the agent resolve this week that it should have escalated, and how did we find them?

• Can we see, for any ticket, what the agent did, which knowledge it used and why it stopped?

• Do our views and SLAs use triage fields, and do those fields still match the renamed "Topic" field?

• Is QA covering every AI-handled ticket or a sample, and who reads the worst ones?

• What does one more resolved ticket cost us under each vendor's billing unit?

To try this rhythm with an agent that runs inside your existing Zendesk, you can start a 30-day free trial of Lorikeet. For QA across human and AI agents together, see how to QA 100% of tickets.