Skip to main content
Procedures tell RAUM AI how your organization handles support requests. When a ticket arrives, the Planner Agent searches your procedure library using semantic similarity and selects the most relevant entries before deciding how to respond. The more accurate and complete your procedures are, the more consistently RAUM AI resolves tickets correctly.

​
Procedure types

RAUM AI supports two procedure types, each serving a different purpose:

KNOWLEDGE

Factual reference information — policies, eligibility rules, product specs, FAQs. The AI reads KNOWLEDGE procedures to understand what is true before composing a reply.

RUNBOOK

Step-by-step workflows — what to do when a customer requests a refund, how to process a replacement, when to escalate to a human. The Planner Agent follows RUNBOOK procedures to decide which Zendesk actions to take.
Use KNOWLEDGE for “what” and RUNBOOK for “how”. A procedure can contain both kinds of content, but keeping them separate makes it easier to maintain each independently.

​
Create a procedure

1

Open the procedures section

From the RAUM AI dashboard, click Procedures in the left sidebar.
2

Click New procedure

Click the New procedure button in the top-right corner.
3

Choose a type

Select KNOWLEDGE or RUNBOOK from the type dropdown. This selection affects how the AI weights and applies the procedure during ticket processing.
4

Write the title

Give the procedure a clear, descriptive title. The AI uses the title as part of its semantic matching, so a title like “Refund eligibility — digital products” will match more reliably than “Refunds 2”.
5

Write the content

Enter the procedure content in the editor. Markdown is supported. For RUNBOOK procedures, use a numbered list to define each step explicitly.
6

Save

Click Save. RAUM AI generates a vector embedding of the procedure in the background. The procedure becomes searchable within a few seconds.

​
Best practices for writing procedures

Be specific about scope. A procedure titled “Handle billing issues” is too broad. Split it into “Process subscription cancellation requests”, “Issue partial refunds for unused service”, and “Dispute credit card chargebacks” so the AI can match each ticket type precisely. Write from the AI’s perspective. The Planner Agent reads procedures in the context of a specific ticket. Include the conditions that activate the procedure: “If the customer is within 30 days of purchase and the item is unopened…” Avoid ambiguous language. Words like “usually”, “sometimes”, or “it depends” reduce the AI’s confidence. State the rule explicitly, then use a separate procedure to handle the exception. Keep RUNBOOK steps atomic. Each step in a RUNBOOK should be a single, verifiable action. “Check order status and contact warehouse if delayed” should be two steps. Include expected outcomes. End RUNBOOK procedures with the expected ticket state: “Mark ticket as solved” or “Escalate to tier-2 with priority High”.
After saving a procedure, test it by submitting a sample ticket in Copilot Mode and checking whether RAUM AI cites the correct procedure in the reasoning output.

​
Version history

Every time you save a procedure, RAUM AI creates a new version. To view the version history for a procedure, open it in the editor and click the History icon in the top toolbar. You can view any previous version and restore it with one click.
Restoring a previous version creates a new version entry — it does not overwrite the current version. Your edit history is never lost.

​
How RAUM AI finds relevant procedures

When a ticket arrives, RAUM AI converts the ticket content into a vector embedding and performs a semantic similarity search across your entire procedure library. This means the AI matches on meaning, not just keywords. For example, a ticket that says “I bought the wrong size and want to swap it” will match a procedure titled “Process product exchanges” even though none of those words appear verbatim in the title. The Planner Agent receives the top-matching procedures along with the ticket context and real-time data from any connected tools. It then selects which procedures apply and executes the corresponding actions.
The quality of semantic search depends on your embedding model. Dedicated embedding models (such as text-embedding-3-small for OpenAI or text-embedding-004 for Gemini) produce significantly better results than using a general chat model for embeddings.