Intent routing for AI agents: send each request to the right agent, tool or model

An LLM can route requests when given your tool descriptions and rules. A separately trained classifier is another option when you have reviewed examples and want a held-out report, batch processing and confidence-based review. Both approaches still need evaluation on your routes.

On 3,080 BANKING77 test messages across 77 intents, StayCharted’s AI Classifier Model scored 92.6%, its Dedicated AI Model scored 93.7%, and Claude Sonnet 5.5 scored 84.7% with category definitions. This measures classification on that dataset, not an end-to-end agent deployment. Results on your routes may differ. Read the full method and all three datasets →

It reads the user's message, and optionally a little context:

It answers with your routes, for example:

• Billing agent
• Order-lookup tool
• Small model, general question
• Large model, complex request
• Person

Each answer comes with a confidence. Below your cutoff, fall back to the LLM router or to a person, while keeping your safety and authorization rules in force.

1. Export requests where the right route is known

From your logs, export user messages and the route that turned out to be right:

• the agent or tool that resolved it;
• the route your team corrected it to;
• or the label your support team gave it.

A few thousand rows is a good start; around 20 reviewed examples per route is a starting point, not a guaranteed minimum for useful results.

2. Train and compare

Upload the file and train an AI Classifier Model. It reads meaning and many languages, and runs on CPU. The report shows how often it chose the right route for messages it never saw, route by route, and which routes it mixes up.

Those mix-ups are worth reading: confused routes may need more examples, clearer labels or a different model.

3. Call it before the agent runs

Publish the model and call it through the Prediction API at the start of each request:

POST /api/v1/models/{model_id}/classify
{ "input": "I was charged twice for the same order" }
→ { "label": "Billing agent", "confidence": 0.97, "needsReview": false }

• needsReview false (above your review cutoff): route directly.
• needsReview true: fall back to your LLM router or to a person. On plans with the Review Queue, those requests also wait there, so your team can confirm the right route and teach the next version.

Agents built on Claude or ChatGPT can also call the model as a tool, through the StayCharted MCP connector.

4. Learn from the misroutes

Log each time a request was rerouted, add those rows to the training file, and retrain. The router keeps up as you add tools and agents.

Use the authenticated API base URL https://amt.staycharted.com and an Authorization: Bearer <key> header. Handle errors and null answers with a safe fallback. For several training columns, send an object with the same field names. Full API reference →. API access requires Essentials or Business.

• It doesn't answer the user. It picks the route; your agents do the work.
• It only knows the routes it was trained on. A new tool needs examples before the router sends anything to it.
• It doesn't replace your safety checks. Keep your guardrails before and after the route.

Use reviewed examples that reflect your current rules. Check errors by category and sample confident answers as well as uncertain ones. Confidence is a model score, not a guarantee. Keep mandatory safety and access rules in your application. Best practices for training and review →.