High-volume triage is one of the most common pain points in enterprise service management. Requests come in through portals, voice, emails, EmployeeWorks... Every one of them needs to land with the right workflow before any real work can begin.

 

The triage team handling that cannot keep up with the volume: requesters are waiting for hours or days, to sometimes only get asked to provide some clarifications.

 

While AI feels like the perfect solution, it's usage need to be cost-conscious by choosing the right tool for the right job. Agentic AI earns its place only where a process genuinely requires autonomous reasoning and will shine when classic automation and machine learning are not sufficient. If you have not read my previous post on choosing the right tool for the right job, that is a good place to start.

---

Layered design: exhaust cheaper options first

 

Layer 1 - Predictive Intelligence

 

PI runs on every case, consumes no Assists, and completes in milliseconds. When the request maps clearly to a known pattern, PI sets the category and routes the case automatically. No generative AI involved, no cost incurred.

 

Layer 2 - AI Agent

Invoked only when PI cannot resolve with sufficient confidence. The agent analyses the case, reasons against more context, and fills in what PI could not determine.

But its role goes further than routing: when the request is ambiguous , the AI agent immediately reaches back to the requester asking a contextualised clarifying question. Rather than a request sitting in a queue for hours or days until a human picks it up, that conversation starts within seconds of submission.

 

Layer 3 - Human Triage

Reserved for genuinely complex or out-of-scope cases. By the time a case reaches a human, the AI agent has already documented its reasoning, so the team is not starting from scratch.

---

The architecture

 

When a case is created, 2 PI models run in parallel: one for category, one for assignment group.

If both models resolve with sufficient confidence, the case routes immediately at zero AI cost.

Else, the AI Agent is invoked, handles what it can, and either routes the case or requests clarification from the requester. The whole process completes in seconds.

---

The power of doing this in ServiceNow

 

There are many ways to build triage logic, but where you build it matters as much as how.

By centralising this design in ServiceNow, there is one central place to maintain the intelligence, improve it, and measure it. A refinement to the routing logic applies instantly across every intake channel: portal, email, voice, or a messaging app. All the data collected across those interactions feeds back into the same models, making the system progressively smarter.

As interaction patterns evolve and new channels emerge, the logic behind them does not! You simply connect a new channel to the same brain.

What also makes ServiceNow the right home for this design is the tooling available in one place: flows, automation, machine learning, generative AI, and human task management are not separate tools that need to be integrated. They are native capabilities that can be orchestrated toward a single objective.

This use case is a direct expression of that: a flow handles the deterministic routing, PI handles the machine learning layer, an AI Agent handles the autonomous reasoning, and human triage handles what requires judgment.

Each capability doing what it is built for, in one platform, for one outcome.

---

Start today, improve as you go

 

A common reason organisations delay automation is data quality... This design removes that dependency.

You start with the data you have. PI performs at whatever level your current data supports. Where it is uncertain, the AI Agent steps in and handles the case. As cases are resolved and closed, they become new training material for PI automatically. Data quality improves through normal operations, not through a separate cleanup project.

There is no need to wait for a hypothetical data remediation effort. The architecture is designed to improve from the moment it goes live.

---

A design that improves over time

What makes this model particularly effective is that it is self-reinforcing. When the AI Agent routes a case that PI could not, and that case is eventually closed, it becomes training material for the PI models. Over time, PI becomes more accurate, the agent is invoked less often, optimising costs.

The system gets smarter the longer it runs, without any manual retraining effort.

---

Beyond the example

 

While this design was originally built for a university to tackle thousands of student inquiries, these patterns apply to any ServiceNow customer using the platform for Enterprise Service Management.

The investment is modest. The payoff, faster routing, immediate requester communication, and a cost profile that scales with actual complexity rather than raw volume, compounds over time.

---

In the next post, I will walk through how this was built: the flow architecture, the agent instruction design, and the decisions behind the confidence threshold logic.

---

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.