Why should you care?

AIPredict shopper bot must route “where is my order” vs “report fraud” before any generative answer.

See it live — copy this example

Use a .NET console or Web API project with Microsoft.ML. Run dotnet run after pasting.

var intents = ml.Transforms.Text.FeaturizeText("Features", "Utterance")
    .Append(ml.MulticlassClassification.Trainers.SdcaMaximumEntropy());
var intentModel = intents.Fit(labeledUtterances);
var intent = ml.Model.CreatePredictionEngine<UtteranceRow, IntentPred>(intentModel)
    .Predict(new UtteranceRow { Utterance = "I was charged twice" }).PredictedLabel;

What happened?

• Train intent classifier on utterances; slot filling can be separate NER or rules.
• ML.NET handles structured routing cheaply.

Practice next

1. Label utterances with intents.
2. Train multi-class on text.
3. Wire intent → handler map.
4. Add confidence threshold fallback.
5. Log misclassified utterances for retrain.

Remember

Intent ML first. Handler per intent. LLM optional layer.

AIPredict shop bot

Bot routes fraud utterances to dispute flow.

Outcome: Cheaper, faster than LLM-only routing.