Skip to main content
Most AI support tools fire a single prompt at a ticket and hope the response is close enough. RAUM AI works differently. Every ticket runs through a structured, multi-step pipeline where the AI reads context, searches your knowledge base, calls your connected tools, and only then writes a response — all before sending a single word to your customer.

​
The four-step pipeline

1

Understand

The AI reads the full ticket — the message, the conversation history, Zendesk tags, and any custom fields on the record. It figures out what the customer actually wants, classifies the intent, and determines what information it needs before it can resolve the issue. Nothing is guessed at from the subject line alone.
2

Plan

With intent established, the AI searches your procedures and knowledge base using semantic similarity — not keyword matching. It identifies which procedures apply, which tools need to be called, and maps a step-by-step resolution path before taking any action. If no procedure matches, it flags the gap so you can address it.
3

Act

The AI executes the plan. It calls your connected tools — order management, billing APIs, CRM systems, MCP servers, or any external API you’ve configured — fetches live data, and takes actions when you’ve enabled autonomous operation. Tool calls are cached to avoid redundant API requests and reduce cost.
4

Reply

With all the context gathered, the AI generates the final customer-facing response. The reply follows your brand voice guidelines, stays within your configured policies, and streams directly into the Zendesk composer — ready to send or edit.

​
Multi-agent architecture

RAUM AI uses a team of specialized agents behind the scenes, each with a focused role. You don’t need to manage them directly — they operate as part of the pipeline automatically.

Thinker

Analyzes the incoming ticket to identify the root problem type, customer intent, and suspected causes. Determines which tools and data sources the pipeline needs before proceeding.

Planner

Takes the Thinker’s analysis and searches your procedures and articles. Picks the right resolution path, selects the appropriate tools, and decides whether the ticket should be auto-resolved or escalated.

Tool Executor

Runs the tool calls — API lookups, MCP server calls, data fetches — that the Planner identified. Results are cached per ticket to avoid duplicate calls and reduce latency.

Answerer

Generates the final reply using everything the pipeline collected: live data, procedure content, conversation history, and your brand voice guidelines. Streams the response directly into Zendesk.

​
Fast path for simple tickets

Not every ticket needs the full pipeline. When the Thinker determines a ticket is straightforward — a common question, a known issue, a repeat query — the Reasoner Agent handles it directly without invoking the full Planner and tool execution stages. This reduces latency and cost for tickets that don’t need live data lookups.
The fast path activates automatically. You don’t configure it — RAUM detects ticket complexity and routes accordingly.

​
Response cache

When RAUM has already generated a response for a ticket type before, it can serve a cached reply instantly — skipping AI processing entirely. This brings response time to near zero for repeated queries and eliminates the token cost for those tickets. The response cache is available on all plans and works in both Copilot and Agent Mode. Cache hits are tracked separately in your analytics dashboard so you can see exactly how many tickets are being served from cache versus processed fresh.

​
Ticket enrichment

Every ticket that passes through the pipeline gets automatically enriched with metadata that helps your team prioritize and route work. You can view enrichment data in the Zendesk sidebar alongside each ticket.
RAUM assigns a sentiment classification to each ticket: urgent, negative, neutral, or positive. Tickets flagged as urgent or negative surface in your analytics and can trigger proactive alerts to your team before they escalate.

​
How Copilot and Agent Mode fit in

The same four-step pipeline powers both modes. The difference is what happens at the end of the Reply step.

Copilot Mode

The reply streams into the Zendesk composer as a draft. Your agent reviews, edits if needed, and sends. Full human control over every response.

Agent Mode

The pipeline runs end-to-end via webhook, and RAUM submits the reply and resolves the ticket automatically. Your team only sees tickets that required escalation.
You can run both modes simultaneously. A common setup is to use Agent Mode for high-confidence, routine ticket categories and Copilot Mode for everything else.