How do I combine business rules with machine learning predictions in Automation Decision Services and call the result from BAW?

Question

Underwriting today is a rule table; the data science team has a risk model. Can Automation Decision Services use the model inside the decision, and how does a BAW process call such a decision?

Answer 1

Yes. ADS decision models can contain predictive model nodes next to rule nodes:

1. The data scientists deploy the model on Watson Machine Learning (watsonx.ai) or expose it through an Open Prediction Service (OPS, an open REST contract for scoring, also usable for models on other platforms).
2. An ADS administrator registers the platform as a machine learning provider (URL, credentials) in Decision Designer.
3. In the decision service, Create a predictive model: import the deployment; ADS generates a template with the model's input and output signature; you map the decision's data to the model inputs and encapsulate the predictive model in a decision node.
4. Rules use the prediction as an input: if the risk score of the application is more than 0.7 then set decision to "MANUAL"; rule decisions remain explainable, the model contributes one input.

Decision model "Underwriting"
  inputs: application (Application), applicant (Applicant)
  [Risk score] (predictive model node, ML provider "wml-prod", deployment "risk-model-v3") <- applicant.age, applicant.income, application.amount ...
  [Decision] (rule node) <- application, Risk score
     rules: if Risk score > 0.7 then decision = "MANUAL"; if application.amount > 50000 then decision = "MANUAL"; else "AUTO_APPROVE"
  output: underwritingResult { decision, riskScore, reasons }

# BAW: the decision service is deployed and discovered in Business Automation Studio; a service flow has a "decision" step:
[Underwriting] (ADS decision "Underwriting", version 1.3) : inputs tw.local.application, tw.local.applicant -> tw.local.underwriting
# under the hood it is a REST call: POST https://<ads-runtime>/ads/runtime/api/v1/deploymentSpaces/prod/decisions/.../operations/underwriting/execute

Guidance: version the model deployment and the decision together (a new model version is a new decision version - test with the decision's test data before promoting); log the model version and the score in the process instance (audit, later model monitoring); apply thresholds in rules, not in the model, so that business can change them without a data science release; and monitor drift with watsonx.governance or the provider's monitoring - the decision keeps working when the model degrades, it just decides worse. If the model is not on WML, wrap it in the Open Prediction Service contract rather than calling it from BAW scripts - that keeps the decision explainable in one place.

References

• ADS documentation - creating predictive models (Watson Machine Learning, Open Prediction Service)
• CP4BA documentation - Automation Decision Services
• watsonx.governance