Platform Migration
AI support tools
AI support tools can suggest a reply to an agent or answer a customer directly.
AI support tools can suggest a reply to an agent or answer a customer directly. The first leaves a person approving each response; the second needs stronger rules for source access, uncertainty and handoff before it goes live.
Begin with a narrow customer issue that has approved, current support material. Identify what the tool may read, what it may say, what it must never infer and what action requires a human. Zendesk's documentation illustrates the distinction: its auto assist suggests replies for agents to approve, while its AI agents generate replies from connected knowledge sources. The exact controls vary by product and configuration.
Where AI fits in customer support
Customer-facing AI support tools include chatbots that respond to common enquiries through live chat or voice, and virtual assistants that provide personalised support. AI can also support interactive voice response and call routing by directing enquiries according to customer attributes or details, while self-service portals help customers find information independently.
Other capabilities work across customer interactions rather than answering a single enquiry. Sentiment analysis uses AI to analyse customer or employee emotions, while natural language processing helps systems understand and respond to human language. For voice services, neural voices use neural networks to create synthetic speech with human-like intonation, pace and sound.
Machine learning enables systems to learn from interactions and refine responses and decisions over time. Large language models are a type of machine learning specialised in natural language processing and generating human language.
What agent-side assistance can draw on
Zendesk describes auto assist as using the contents of submitted tickets to suggest replies, actions and instructions. Suggestions appear in the ticket interface, where agents review them before sending a reply or carrying out an action. This makes the tool an agent-facing aid rather than a customer-facing chatbot.
For auto assist, Zendesk identifies procedures as the primary knowledge source. Relevant help centre articles can also be used alongside procedure content, including public articles and internal articles made available through user segments. This gives teams a way to make internal guidance available to the tool without treating every help centre article as public.
If there are no relevant procedures or help centre articles, Zendesk says auto assist uses similar solved tickets. Teams configuring the feature can manage which sources auto assist may use.
Choose and connect knowledge sources
Zendesk AI agents can use the help centre associated with a Zendesk account, including a help centre for a particular brand where multiple brands are set up. They can also use external knowledge content brought into Zendesk through a web crawler or a knowledge connector.
An AI agent can have multiple knowledge sources, including several help centres, several external sources or a combination. Zendesk recommends keeping the overall number within a reasonable limit because too many sources can, in some cases, reduce accuracy and increase latency.
A connected knowledge source is required for an AI agent to create a generative answer. Zendesk notes that adding a generative replies block or instructing a generative procedure to produce a knowledge reply without any connected sources results in a technical error. Check that the sources needed for the intended feature are connected before enabling it.
Pre-Launch Checklist for AI Agents
- ✅ At least one knowledge source is connectedRequired for generative replies
- ✅ External sources synced within last 24 hoursEnsures up-to-date information
- ✅ Restricted content access aligned with user rolesAuthenticated customers see relevant private articles
- ✅ No more than 5–6 knowledge sources usedTo maintain accuracy and reduce latency
- ✅ Human handoff triggers definedFor uncertain or sensitive queries (e.g., tax disputes)
Account for source freshness and access
The age of the information available to an AI agent depends on the source type. Zendesk says an agent searches the current help centre content when a search is performed, but searches external knowledge using information available at the last sync, which usually happens every 24 hours. An external page may therefore have changed since the content was last brought into the account.
Restricted help centre content is subject to article view permissions. For an authenticated customer, the AI agent can use relevant restricted articles; for an unauthenticated customer, it can use only public articles.
Keep the tool within its intended scope
A support tool should only handle issues that match the sources and permissions connected to it. If an issue needs a source the tool cannot read, or an action it is not allowed to take, keep that issue out of scope.
Source lists and permissions change as products, policies and help centre content change. Review the connected sources and the issues assigned to the tool as part of routine platform maintenance, not only at go-live.
In this guide
- Testing AI answers against approved support materialThe point of an AI answer test is to check whether the response follows the support material customers are meant to rely on.
- Comparing suggested replies with autonomous responsesSuggested replies and autonomous responses have different failure paths.
- Checking human handoff when an answer is uncertainAn AI support tool needs an explicit route for uncertainty.
- Recording corrections to an AI-assisted answerA correction should improve the next support interaction, not disappear as a one-off ticket note.
