Architectural Overview

The DiscoveryAIModel metadata type represents a deployed machine learning model trained in Einstein Discovery (part of CRM Analytics and Tableau CRM). In intelligent enterprise architectures, machine learning models analyze historical transactional datasets to produce predictive scores (such as customer churn probability, opportunity win likelihood, or invoice payment delays) alongside prescriptive recommendations.

By packaging model hyperparameters, scoring coefficients, and target outcome fields into DiscoveryAIModel metadata, data science and DevOps teams can train models on scrubbed analytics sandboxes and deploy approved models deterministically into production via continuous delivery pipelines.

Einstein Discovery Story Training DiscoveryGoal & DiscoveryAIModel Model Packaging (.discoveryAIModel-meta.xml) Deploy to Production Org Flow / Apex Prediction Service Live Predictive Scoring on Record Pages

SFDX Directory and File Suffix

Property Value
Metadata Type Name DiscoveryAIModel
SFDX Directory Path force-app/main/default/discoveryAIModels/
File Suffix .discoveryAIModel-meta.xml
API Version 68.0 (Winter ’27)

Key Metadata Fields

Field Type Required Description & Architectural Considerations
label string Yes Human-readable name displayed in Model Manager and Flow Builder.
description string No Architectural description of model training features, dataset, and accuracy metrics.
discoveryGoal string Yes Developer name of the parent DiscoveryGoal this model optimizes.
modelType string Yes Algorithmic prediction structure (BinaryClassification, Regression, Multiclass).
algorithmType string No Machine learning algorithm used (GLM, GBM, RandomForest, XGBoost).
predictionType string Yes Nature of model scoring output (e.g., Probability, Score, RawValue).
status string Yes Operational deployment state (Active, Draft, Retired).

Note

Deploying a DiscoveryAIModel makes the prediction definition immediately available inside Model Manager and within the Einstein Prediction Service for real-time scoring.

Declarative XML and Package.xml Examples

DiscoveryAIModel Metadata File

force-app/main/default/discoveryAIModels/Opportunity_Win_Predictor.discoveryAIModel-meta.xml
<?xml version="1.0" encoding="UTF-8"?>
<DiscoveryAIModel xmlns="http://soap.sforce.com/2006/04/metadata">
    <algorithmType>GBM</algorithmType>
    <description>Gradient Boosted Machine predicting opportunity win probability and prescriptive recommendations.</description>
    <discoveryGoal>Maximize_Opportunity_Win_Rate</discoveryGoal>
    <label>Opportunity Win Predictor</label>
    <modelType>BinaryClassification</modelType>
    <predictionType>Probability</predictionType>
    <status>Active</status>
</DiscoveryAIModel>

package.xml Manifest

manifest/package.xml
<?xml version="1.0" encoding="UTF-8"?>
<Package xmlns="http://soap.sforce.com/2006/04/metadata">
    <types>
        <members>Opportunity_Win_Predictor</members>
        <name>DiscoveryAIModel</name>
    </types>
    <version>68.0</version>
</Package>

SFDX CLI Retrieval and Deployment Commands

Retrieve and deploy predictive models across test and production orgs:

Terminal
# Retrieve all DiscoveryAIModel components
sf project retrieve start --metadata "DiscoveryAIModel" --target-org analytics-dev

# Deploy the model and its governing goal simultaneously
sf project deploy start \
  --metadata "DiscoveryGoal:Maximize_Opportunity_Win_Rate,DiscoveryAIModel:Opportunity_Win_Predictor" \
  --target-org uat-org

# Query existing models in Model Manager via Tooling API
sf data query --use-tooling-api --query "SELECT Id, DeveloperName, ModelType, Status FROM DiscoveryAIModel" --target-org uat-org

Senior Architect Best Practices and Gotchas

Warning

Parent DiscoveryGoal Dependency: A DiscoveryAIModel cannot exist without an associated DiscoveryGoal. If the goal does not exist in the target org or is omitted from the deployment manifest, the deployment will fail immediately.

Tip

Continuous Drift Monitoring: Machine learning models experience accuracy degradation as business conditions shift. In production, configure Einstein Model Manager alert thresholds to trigger retraining workflows when prediction accuracy drops below acceptable R-squared or AUC limits.

  • Batch Scoring Governor Limits: Scoring hundreds of thousands of records synchronously via Apex can exceed daily Einstein Prediction Service API call allocations. Leverage asynchronous Batch Apex or Flow bulkification for nightly scoring jobs.
  • Explainability and Compliance: When deploying models in regulated industries (financial lending, insurance underwriting), ensure feature weights exclude protected demographic attributes to avoid algorithmic bias and compliance penalties.

Frequently asked questions

What types of machine learning algorithms does DiscoveryAIModel encapsulate?

It supports Generalized Linear Models, Gradient Boosted Machines, Random Forests, and XGBoost models trained inside Einstein Discovery.

Where are DiscoveryAIModel files stored in an SFDX project?

They are stored in force-app/main/default/discoveryAIModels with the file suffix .discoveryAIModel-meta.xml.

How is a DiscoveryAIModel invoked to score records in production?

It is scored via Flow Builder prediction actions, Apex Discovery API calls, or Einstein Prediction Service REST endpoints embedded on record pages.

Comments

Questions, corrections and tips are welcome. Comments are GitHub Discussions, so you sign in with GitHub.