How AI Assistant works

AI Assistant does more than send the ticket text to a model and accept the first answer. It builds a grounded support request from Zendesk context, retrieves applicable policy and knowledge, uses approved tools when required, and evaluates the resulting answer before it is delivered to an agent or customer.

​

The customer-facing mental model

You can think of the process in four phases:

​

What happens internally

The exact execution path is optimized by RAUM and not every ticket requires the same model calls. A typical tool-backed run uses specialized stages such as:

• Thinker — determines the problem and required evidence.
• Tool execution — retrieves live data or prepares approved actions.
• Planner / reasoning path — combines policy, retrieved evidence, and the ticket state into a resolution plan.
• Answerer — writes the final response using the grounded result.

Straightforward tickets can use a lighter reasoning path instead of paying the latency cost of unnecessary planning or tools.

​

Knowledge retrieval

RAUM uses semantic retrieval rather than requiring exact wording matches. The retrieval layer can search:

• procedures;
• AI Guidelines;
• Help Center articles;
• uploaded documents;
• crawled websites; and
• other indexed resources enabled for the Assistant.

A valid embedding configuration is important because semantic resources without usable embeddings cannot participate normally in retrieval.

​

Procedures versus guidelines

Procedures define facts, decisions, steps, and references to tools or Zendesk actions. AI Guidelines shape how the answer is communicated and can also carry targeting rules. Assistant guidelines can be marked for normal Response use, Translate use, or both. This separation prevents a translation-specific terminology rule from accidentally changing ordinary answer generation, and vice versa.

​

Tool safety

Connections are classified by what they do:

• SOURCE — read-only retrieval. These can supply evidence to the answer.
• ACTION — can change state. These are gated by approval/authorization behavior rather than being treated as ordinary lookups.

Simulation follows the same distinction: SOURCE tools can run normally so the test sees authentic data, while ACTION tools are intercepted and reported as what would have run.

​

Confidence

Generate results include calibrated confidence that is used by more than the UI. Confidence can influence:

• the confidence tag added to Zendesk;
• Agent Mode public-reply versus internal-note behavior;
• ticket completion behavior; and
• quality analysis.

Because of that, confidence should not be treated as a decorative percentage.

​

Caching and pre-caching

AI Assistant has two related performance controls under AI Assistant → Zendesk → Automation → Performance:

​

Response caching

Response caching can reuse a valid recent result for an equivalent request. Administrators can enable or disable it and configure the cache duration.

​

Copilot pre-cache

In Copilot Mode, administrators can optionally generate answers when qualifying new-ticket webhook events arrive. When the agent later clicks Generate, a valid pre-cached answer can be returned immediately. Pre-cache can be limited to selected Zendesk channels.

​

Enrichment

Where enabled, Assistant can enrich Zendesk tickets with information such as:

• sentiment;
• category;
• summary;
• configured tags or field mappings.

Enrichment behavior is managed under AI Assistant → Zendesk → Enrichment.

​

What gets recorded

Assistant runs can create Ticket Activity and related system/audit records so administrators can review what happened, including the sources/tools involved and delivery outcome where available.