Copilot Mode

Copilot Mode is AI Assistant’s agent-reviewed operating mode. RAUM researches the ticket and prepares content for the Zendesk agent, but the agent remains responsible for reviewing and sending the customer-facing reply.

​

Enable Copilot Mode

Open AI Assistant → Zendesk → Automation → AI Mode and choose Copilot Mode. Copilot is the safer starting point for a new deployment because it lets you validate:

• procedure selection;
• tool results;
• tone and terminology;
• field and enrichment behavior; and
• the final customer-facing wording.

​

Generate a reply

Inside Zendesk, Generate runs the Assistant response flow for the current ticket. The generated result can use:

• the ticket conversation;
• Zendesk fields and metadata;
• matching procedures;
• applicable response guidelines;
• Help Center and uploaded knowledge;
• live SOURCE tool output; and
• approved ACTION behavior where the workflow requires it.

The draft is inserted into the Zendesk composer for agent review. For the exact buttons exposed by the current Zendesk app, see Zendesk App overview.

​

Current email-mode controls

The current Zendesk email editor exposes: Realtime Zendesk messaging uses a different, shorter control set. See Chat and messaging mode.

​

Action approvals

A generated response can include an ACTION tool that requires an agent decision. When approval is required, the Zendesk app presents the pending choice rather than silently executing the mutation. Read-only SOURCE tools do not need the same action-confirmation path.

​

Response caching

Administrators can enable Response Caching under: AI Assistant → Zendesk → Automation → Performance The cache duration can be configured from 1 to 480 minutes. Use it when repeated requests can safely reuse a recent response under the same relevant context. Clear the organization cache after major changes when you need to force fresh results immediately.

​

Pre-cache answers before an agent clicks Generate

Copilot Mode also supports optional pre-cache behavior. When enabled, a new-ticket webhook can run the Assistant ahead of time. The result is cached so the agent can receive it immediately when they later click Generate. Administrators can configure:

• whether pre-cache is enabled;
• its duration, from 1 to 48 hours; and
• which Zendesk channels participate.

Leaving the channel selection empty applies pre-cache to all supported channels.

​

Review before sending

Before submitting the Zendesk reply, check:

1. Facts — does the answer match the ticket and retrieved evidence?
2. Policy — did the expected procedure apply?
3. Actions — were any state-changing actions actually authorized and delivered?
4. Confidence — does a lower-confidence answer need closer human review?
5. Tone — did the correct Response guideline and Brand Voice apply?

If the first draft is repeatedly wrong for the same class of ticket, fix the underlying knowledge or guidance instead of training agents to regenerate until the model happens to agree with them.

​

Improve poor outcomes

Agents can rate Assistant responses. Poorly rated tickets can enter the AI → Coaching review queue, where the team can investigate the ticket and propose changes to procedures, guidelines, or AI memory. Learn about AI Coaching →