How to Build a Self-Optimising AI Triage Architecture on ServiceNow

The architecture behind this build was covered in Smarter Triage: a layered AI architecture that keeps costs in check at scale. If you haven't read it, start there. If you have, this is where the build begins.

 

Predictive Intelligence combined with an AI Agent is faster, cheaper, and more scalable than running everything through an AI Agent alone. PI runs server-side, consumes no tokens, and completes in less than a quarter of a second! PI handles what it knows. The agent handles what requires judgment. Over time, the gap between what PI handles and what the agent handles widens in PI's favour.

---

Predictive Intelligence

PI is the first layer. When a case arrives, it predicts 2 fields: assignment group and category. Each is a separate classification model trained on historical case data.

 

Setup

You define a solution that specifies the table, the field to predict, and the input fields the model reads from (often short description and description). Training runs against existing records, scheduled to run regularly as new resolved cases accumulate, keeping the model aligned with how your team actually routes work. Once trained, predictions complete in around 200ms with no tokens consumed.

 

 

Confidence and class thresholds

For each prediction, PI returns a predicted class, a confidence score, and a class threshold — the minimum confidence required to accept that prediction for a specific outcome. Each class can have a different threshold, and ServiceNow sets defaults based on training results. Administrators can tune them through the Class Confidence records.

In Flow Designer, a simple condition triggers the field update if the confidence is above the class threshold, dynamically, per class. If it passes, the predicted value is applied. If not, the flow routes to the AI Agent.

 

Training data quality

The quality of your training records defines the quality of your predictions, but you do not need perfect data to start. Start with what you have today, review the precision/coverage trade-off in the Class Confidence records, and refine from there.

 

 

The self-improving loop, the main advantage of this architecture.

Every time the AI Agent classifies a case that PI couldn't, that case becomes new training data for PI. The next time a similar inquiry arrives, PI handles it directly: no agent invocation, no tokens consumed, response time improves.

The more cases the system processes, the less the agent is needed. The agent is not just a fallback, it is actively teaching PI to handle cases it previously could not.

---

AI Agents

2 agents are in use. 1 handles case triage. The other manages the voice conversation.

 

Triage agent

Fires when PI is not confident enough. Its job is narrow: classify what PI could not, or request clarification when the case is genuinely ambiguous.

 

What the agent does:

1. Reads the case content and reasons directly from it
2. Uses AI reasoning to determine the most appropriate category thanks to more context awareness and smartness than PI
3. Same for assignment group
4. If the case is genuinely ambiguous, posts a clarifying comment on the case to request clarifications.

 

Voice agent

An AI Agent manages the conversation for the voice channel, taking a KB-first approach:

1. Clarify the inquiry if required
2. Search the knowledge base
3. Guide the caller with clear resolution steps if available
4. If the caller confirms resolution, automatically create a CSM Case in a resolved state, in order to track successful deflections
5. If not resolved, create an open CSM Case and route it through the triage flow

---

Voice Channel

Configured as a Now Assist Voice deployment using the KB-First Agent as its backbone. Before the agent searches or routes anything, it asks a clarifying question to better understand the caller's need. Users call in, the agent handles the conversation, and either creates a resolved case (tracked deflection) or an open case that flows into triage.

---

Moveworks

Moveworks (EmployeeWorks) runs as the AI Chat conversational channel, with 2 processes configured:

The main conversational process:

1. When a caller raises an issue:
2. Moveworks searches the knowledge base
3. Surfaces a self-service solution if one exists
4. Creates a CSM Case in ServiceNow if the case is not resolved

And a case update listener: Monitors outbound case updates from ServiceNow. When a status change, agent comment, or resolution fires, Moveworks sends a notification to the caller in their original channel. No portal check required.

 

---

Inbound Email

When a caller emails the support address, an inbound email action creates a CSM Case automatically. It enters the same triage flow as every other channel.

---

Flow: The ServiceNow Superpower, Where It All Comes Together

ServiceNow is uniquely positioned to bundle PI, AI Agents, Voice, email, AI Chat, and more into a single orchestration layer. All components are coming together into one flow that ties them into a triage superpower. 

The flow triggers on 2 conditions: when a Case is created, and when a caller clarifies the request (re-running with the additional context).

 

Parallel PI predictions

Two PI predictions run simultaneously in a parallel block, one for assignment group and one for category. Each branch calls the PI solution, compares confidence against the class threshold, and writes the predicted value if the check passes.

 

Post-parallel consolidation

After both branches complete, a Lookup Records step reads the latest case updates before any further logic runs.

 

AI Agent fallback

If either field is unresolved, the flow invokes the triage agent for those fields only.

Human fallback

If the agent cannot resolve both fields, the case routes to a human triage group.

If a caller responds with some clarifications, the flow re-triggers. PI and the agent get another attempt with the additional context, without any manual re-routing.

---

End-to-End

The stack optimises itself over time. PI handles more. The agent handles less.

 

Once you think in layers, exhausting the cheap option first, you start seeing this pattern everywhere in agentic AI design.

---

Sylvain Hauser is an AI Architect at ServiceNow, based in Australia. He holds Certified Master Architect (CMA) and Certified Technical Architect (CTA) credentials with 18+ years of ServiceNow expertise.