Artificial Intelligence
Extending Zendesk with AI: Options and Limits
There are four ways to extend Zendesk with AI. What each one is good for, where the limits sit and which metrics show whether the extension is working.
There are four ways to extend Zendesk with AI: the platform’s native features, apps from the Zendesk Marketplace, a pre-ticket AI covering the help page, chat and contact form, and custom integrations built on the Zendesk APIs. This article covers what each route is good for, where the limits sit, and how to tell whether the extension is actually working.
Key takeaways
- Four routes: Native AI features, Marketplace apps, pre-ticket AI and custom builds on the APIs. They combine rather than compete.
- The main limit: An AI inside the ticketing system only acts once a ticket exists. The larger volume arises before that.
- The second limit: Answer quality depends entirely on the maintained knowledge base. No model compensates for missing content.
- Measurement: Deflection and CSAT belong in one report, otherwise an abandoned conversation is booked as a success.
- OMQ: OMQ Assist installs as an app from the Zendesk Marketplace, while Help, Chatbot and Contact act before the ticket — all from one knowledge base.
- How can you extend Zendesk with AI?
- How does Zendesk bill its AI agents?
- Where are the limits of AI inside a ticketing system?
- Extend or replace?
- Which metrics show whether it is working?
- What should decide the selection?
- What do GDPR and the EU AI Act require?
- How do you approach the rollout?
- Extending Zendesk with OMQ
- Key terms at a glance
- Conclusion
- Frequently asked questions
How can you extend Zendesk with AI?
Four routes are available. They are not mutually exclusive; in practice most organisations combine two or three of them.
| Route | Where the AI acts | Typical effort | Suited to |
|---|---|---|---|
| Native AI features | Inside the Zendesk platform | Configuration | A quick start with no additional vendor |
| Marketplace app | In the ticket, in chat or in the help centre | Installation and connection | Adding capability without replacing anything |
| Pre-ticket AI | Before the ticket, on your help page, in chat, in the form | Building the knowledge base | Reducing volume before tickets are created |
| Custom build on the APIs | Wherever you choose | Development project | Edge cases, in-house systems, specific workflows |
Zendesk’s native AI features
Zendesk ships its own AI components, including AI agents for automated answering and Copilot as an assistant for agents. The appeal is obvious: no second contract, no second interface. Billing for the AI agents runs separately from the platform licence, which is covered below.
Apps from the Zendesk Marketplace
Zendesk explicitly provides for extension. Apps, themes and bots can be installed through the Zendesk Marketplace to add to the functionality of Support, Chat and Guide, built technically on the Zendesk Apps Framework. For most service teams this is the least disruptive route to additional AI capability, because nothing in the existing configuration changes.
Pre-ticket AI on the help page, chat and contact form
This route deliberately leaves the ticketing system. The AI answers enquiries where they arise: in the help page search field, in the chat window and in the contact form. Only what remains open is handed to Zendesk as a ticket. The leverage is greatest here, because the enquiry never becomes a case in the first place.
Custom builds on the Zendesk APIs
Zendesk provides APIs for ticketing, the help centre and AI agents, alongside a framework for building your own apps and action flows for automating processes beyond Zendesk. This route makes sense when in-house systems have to be involved and nothing off the shelf fits. It does tie up development capacity, including in ongoing operation.
How does Zendesk bill its AI agents?
Anyone weighing the native features against an extension should understand the billing model. The following reflects Zendesk’s public documentation as of September 2026.
Zendesk bills AI agents per automated resolution, meaning per enquiry the AI resolves without escalation to a human. Since 18 May 2026, Zendesk has distinguished three tiers:
| Tier | Description per Zendesk | Draws on allowance |
|---|---|---|
| Assisted escalation | The AI contributed, for example by collecting data or routing, but a human completed the resolution | No |
| Contained resolution | The AI responded without further assistance being requested, but the subsequent check does not confirm the resolution | No |
| Verified resolution | The AI responded, no further assistance was requested, and the check confirms the resolution | Yes |
The check is performed once the conversation ends by a language model that evaluates the conversation text. When a conversation counts as ended depends on the channel: email 72 hours after the last message, messaging two hours by default with the option to extend to 72 hours, and voice immediately on hang-up.
Where are the limits of AI inside a ticketing system?
The limits below apply regardless of vendor. They follow from the architecture rather than from the quality of any particular product.
The AI only acts once the ticket exists
An AI inside the ticketing system can work an enquiry but cannot prevent it. The ticket has already been created, with everything attached to it: reporting, SLA, handling steps. The larger lever sits earlier. A Gartner survey of 5,728 customers found that 73 per cent use self-service at some point in their service journey, yet only 14 per cent fully resolve their issue there.
Without a maintained knowledge base, no model helps
Answer quality follows the content, not the language model. In the same Gartner survey, 43 per cent said they abandoned because they could not find relevant content, and 45 per cent said the company did not understand their issue. Both causes sit in the knowledge base and in intent recognition, not in compute.
Scattered knowledge produces contradictory answers
When content is maintained separately per channel, it goes stale in at least one of them, and customers get one answer in chat and another on the help page. An extension is only as good as the source it draws on.
Automation needs a visible way out
Gartner found that 87 per cent of customers consider access to a human agent essential once a company uses generative AI in service. An AI with no visible route to the team lowers satisfaction rather than raising it.
The billing metric is not the success metric
How a vendor bills follows the vendor’s logic. How well your service performs follows a different one. Treating them as the same removes your ability to evaluate the vendor.
Data protection and the AI Act narrow the field
Processing location, subprocessor chains and transparency obligations are not details for regulated sectors, public bodies and universities. They are conditions to be met before selection.
Extend or replace?
The question comes up in almost every evaluation, usually framed as a search for an alternative to Zendesk AI agents. In most cases it is posed too broadly.
| Situation | Sensible response |
|---|---|
| The AI resolves too little, but ticketing works | Extend: add a Marketplace app or a pre-ticket AI |
| Enquiries arrive faster than they are cleared | Extend: act before the ticket |
| Automation cost is not predictable | Extend, and model the additional vendor’s billing against your volume |
| Processes, roles or reporting no longer fit | Evaluate a platform change |
Only the last row justifies a migration programme. The first three are solved by an extension in which no tickets are moved and reporting, SLAs and roles stay untouched.
Which metrics show whether it is working?
Four metrics are enough, and all of them can be built from data you already hold.
| Metric | What it tells you |
|---|---|
| Ticket deflection rate | Share of enquiries resolved before a ticket is created |
| CSAT on automated conversations | Whether the enquiry was resolved or the customer gave up |
| Zero-result searches | Which content is missing from the knowledge base |
| Contact volume per channel | Whether effort is moving rather than disappearing |
The first two belong together, always. Falling tickets alongside falling satisfaction is not a saving, it is displacement into channels nobody measures. The third is also the most actionable: every search term returning nothing is a missing entry.
What should decide the selection?
Seven questions any extension vendor should be able to answer:
- Does the solution act before or after the ticket? That determines how much volume it can reach at all.
- What source does it answer from? Ideally one knowledge base serving every channel equally.
- Does it understand free-text enquiries? Paraphrasing, typos and synonyms, not just keyword matching.
- How does handover to the team work? The context already captured should travel with the case.
- Can the knowledge base be exported? Ask about format, scope and whether it requires a paid engagement.
- What triggers a charge? Get the definition in writing, along with a ceiling on variable cost.
- What compliance evidence exists? Processing agreement, processing location, subprocessors, retention periods.
Question five is the one most often skipped. The cost of a future change rarely sits in the technology. It sits in the content.
What do GDPR and the EU AI Act require?
For organisations operating in the UK and EU, two regimes apply side by side. UK GDPR and EU GDPR require a processing agreement under Article 28, named subprocessors and transparency about where data is processed. Since 2 August 2026, Article 50 of the EU AI Act has applied: people must be able to tell that they are interacting with an AI system, unless that is already obvious from the circumstances. For chatbots the duty applies with no transition period.
For OMQ: OMQ is a German company based in Berlin and provides a processing agreement under Article 28 GDPR including a named external data protection officer. Personal data is stripped out of free-text fields before the text is stored. Details on processing location and retention periods for tender responses are available through the contact form.
How do you approach the rollout?
Five steps, none of which touches the existing ticketing setup.
1. Record your baseline
Contact volume per channel, average handling time, current deflection and CSAT. Without a baseline there is no defensible before-and-after.
2. Extract your most frequent enquiries
They already sit in your Zendesk reporting and form the foundation of the knowledge base.
3. Start with one channel
The help page is usually the right place, because that is where search behaviour becomes visible. Zendesk continues unchanged.
4. Add channels one at a time
Help page first, then contact form and chat, and finally agent support inside the ticket. Each step is separately measurable and separately reversible.
5. Work through the zero-result searches
They show exactly which content is missing, so the knowledge base grows along real demand rather than assumptions.
Extending Zendesk with OMQ
OMQ acts in both places: before the ticket and inside it. Every product draws on the same central knowledge base, so an answer is maintained once and served identically in every channel.
| Product | Where it acts | Effect |
|---|---|---|
| OMQ Help | Help page and FAQs | Resolves enquiries before a ticket exists |
| OMQ Contact | Contact form | Suggests the answer before the message is sent |
| OMQ Chatbot | Website and messengers | Real-time answers plus background actions |
| OMQ Assist | Inside the Zendesk ticket | Suggests replies to agents from the same knowledge base |
| OMQ Reply | Answers recurring enquiries automatically |
The connection runs through an app in the Zendesk Marketplace. OMQ Assist reads an incoming message as the ticket is opened, identifies the request and displays matching answers in the right-hand pane of Zendesk Support. Beyond Zendesk, OMQ supports Freshdesk, Salesforce, Zammad and OTRS among others.
One example from practice: DB Connect, the Deutsche Bahn mobility subsidiary behind Call a Bike and Flinkster, has used OMQ since February 2013, initially in the contact form and the ticketing system.
The real difference is the shared source. The knowledge base that serves customers in self-service is the same one that suggests the reply to your agent inside the Zendesk ticket, which removes the most common cause of contradictory answers.
Key terms at a glance
- Zendesk Marketplace: The directory used to install apps, themes and bots into Zendesk.
- Zendesk Apps Framework: The technical basis on which extensions for Support, Chat and Guide are built.
- Automated resolution: Zendesk’s billing unit for AI agents. An enquiry the AI resolves without escalation to a human.
- Verified resolution: The tier that draws on the Zendesk allowance. A language model has confirmed the resolution after the conversation ended.
- Ticket deflection rate: Share of enquiries resolved before a ticket is created. The most meaningful metric you own.
- Intent recognition: The method an AI uses to determine the purpose behind a free-text message rather than matching keywords.
- Central knowledge base: The shared source of all service answers, ideally feeding every channel.
Conclusion
Extending Zendesk with AI is usually the smaller and more effective move than changing platform. Four routes are open and they combine: native features, Marketplace apps, pre-ticket AI and custom builds on the APIs.
The limits sit less in the technology than in the architecture and the content. An AI inside the ticketing system can work an enquiry but cannot prevent it. And without a maintained knowledge base, any extension stays ineffective, however capable the model underneath.
To evidence the result, report deflection and CSAT together. Only both figures side by side show whether fewer tickets represent resolved enquiries.
Frequently asked questions
How can you extend Zendesk with AI?
Where are the limits of AI inside a ticketing system?
How does Zendesk bill its AI agents?
Do you have to replace Zendesk to get better AI?
What is a Zendesk Marketplace app?
Which metrics show whether the AI extension is working?
What does it cost to extend Zendesk with AI?
What do GDPR and the EU AI Act require?
Sources
- Zendesk Help Center: Extending your support solution with apps and integrations
- Zendesk Developer Docs: API Reference
- Zendesk Help Center: About automated resolution tiers, introduced 18 May 2026
- Zendesk Marketplace: OMQ Assist for Zendesk Support
- Gartner: Only 14% of Customer Service Issues Are Fully Resolved in Self-Service, survey of 5,728 customers
- Gartner: 87% of Customers Say Companies Using GenAI for Customer Service Must Provide Access to a Human Agent, survey of 3,566 customers, February and March 2026