Skip to content
IT Canvass
Experience & AI · Lesson

Predictive Intelligence

Quick answer

Train classification, similarity and clustering models on your own data and use them without eroding trust.

Key takeaways

  • Precision and coverage matter more than raw accuracy
  • Start as a suggestion, automate once trust is earned
  • Set confidence thresholds explicitly
  • Retrain on a schedule or accuracy decays

Three model types

Classification predicts a field value, for example assignment group from short description. Similarity finds records like this one, which powers agent assist. Clustering groups records to reveal patterns, often used to find candidate problems or automation targets.

Training and evaluation

Models train on historical records with a defined filter and target field. Read the precision and coverage numbers before you deploy: a model that is 90 percent precise on 20 percent of records is useful, a model that guesses on everything is not.

Deploying carefully

Introduce predictions as suggestions before automation.

  • Show the prediction with its confidence and let the agent accept it
  • Only auto assign above a confidence threshold you agreed with the service desk
  • Retrain on a schedule, processes drift and accuracy decays
  • Exclude records created by the model itself from the next training set

Want to learn this properly?

Our live, instructor-led ServiceNow Training covers this hands-on, with real projects and a certification path.

Check your understanding

  1. Which model type suggests similar past incidents?

    • A. Classification
    • B. Similarity
    • C. Clustering
    • D. Regression
    Show answer

    B. Similarity

    Similarity finds comparable records for agent assist.

  2. What should gate automatic assignment?

    • A. Record age
    • B. Confidence threshold
    • C. Assignment group size
    • D. Priority
    Show answer

    B. Confidence threshold

    Only predictions above an agreed confidence should act without a human.

Frequently asked questions

How much data do we need?

Thousands of records per class is a reasonable starting point. Sparse classes will not train well and should be excluded.

Can we explain a prediction to an auditor?

You can show the model, its training filter, its evaluation scores and the confidence of the individual prediction. That is usually what is required.
CallWhatsAppEnquire