Procedures

Procedures tell AI Assistant what your organization knows and how it should handle support work. Open Sources → Procedures in the Assistant workspace.

​

Procedure types

A useful shorthand is:

• KNOWLEDGE = what RAUM needs to know.
• RUNBOOK = what RAUM should do.

​

Ways to create a procedure

The current procedure workspace supports several creation paths:

• start from a blank procedure;
• import existing material;
• generate a draft from a description; or
• learn from past tickets, including tickets surfaced by Knowledge Gaps.

AI-assisted creation is a starting point, not automatic policy approval. Review generated procedures before relying on them in production.

​

Write a good procedure

​

Keep the scope narrow

Prefer one clear support problem over a document that tries to describe your entire operation. Titles such as Refund eligibility for annual subscriptions or Replace an item damaged in transit retrieve more reliably than something as broad as Billing.

​

State conditions explicitly

Write the condition before the outcome. For example: if a subscription renewed within a defined eligibility window and no disqualifying usage exists, state the exact outcome the procedure permits. Avoid leaving important policy decisions implied.

​

Keep runbook steps atomic

A step should describe one decision or operation. Split checking an order, checking a carrier, deciding eligibility, performing an action, and changing ticket status into separate steps RAUM can reason about and verify.

​

Distinguish retrieval from mutation

When a runbook needs an external system:

• use a SOURCE connection for read-only data;
• use an ACTION connection for anything that changes state.

This classification affects approval and Simulation behavior.

​

Target procedures

Procedures can be narrowed using metadata such as:

• Zendesk tags;
• Zendesk groups;
• Zendesk organizations;
• languages; and
• the configured match behavior.

Use targeting when a rule genuinely differs by team, customer segment, locale, or workflow. Do not create several contradictory global procedures and hope semantic search develops managerial judgment.

​

Selection behavior

Assistant retrieves procedures semantically, then applies targeting and workflow rules before using them. A ticket does not need to repeat a procedure title word-for-word. Retrieval is based on meaning, which is why your active embedding model matters.

​

Use procedures with connections

RUNBOOK content can reference configured resources such as:

• API Connections;
• MCP tools;
• Zendesk fields or actions; and
• other procedure resources exposed by the editor.

Use the editor’s resource controls rather than inventing free-form tool names. That keeps the procedure linked to resources RAUM can actually resolve.

​

AI review

Procedures can participate in RAUM’s AI review workflow. Use review when you want the system to inspect a procedure for ambiguity or missing information, ask clarification questions, and propose improvements before the content is treated as complete. Human review remains important for policy and side-effecting workflows.

​

Learn from tickets

The Learn from past tickets workflow can turn recurring support patterns into a procedure draft. This is especially useful from Quality & Testing → Knowledge Gaps, where a cluster of low-confidence tickets can seed the learning workflow. Review the resulting draft for customer-specific details, conflicting outcomes, actions that should require approval, edge cases the sampled tickets did not cover, and the final Zendesk state you actually want.

​

Test a procedure

After saving or materially changing a procedure:

1. test a representative ticket in Copilot Mode;
2. confirm the expected procedure appears in Ticket Activity;
3. verify the answer follows the rule rather than merely sounding plausible;
4. run Simulation when the procedure uses external actions; and
5. monitor Knowledge Gaps and poorly rated tickets after release.