This topic walks through the AI Command Center operating model in practice, using a fraud detection initiative at a financial services company as the scenario. Each asset type appears in the order it typically enters the picture, from business intent through to production monitoring. For definitions, go to AI Command Center operating model.
The scenario
Acme Financial Services is building an AI-powered fraud detection capability. Their data science team has trained a classification model in Amazon SageMaker, and the payments application team is ready to consume it. A business steward has identified the initiative and wants it formally governed before it goes live. This walkthrough follows every asset that Collibra creates or records as the capability moves from business intent to production, then shows how an AI agent can be layered on top once the model is running.
Business layer
AI Use Case: the business intent
Before any technical asset exists in AI Command Center, a business steward registers the initiative. They open the registry, click Register AI use case, and create an asset named Fraud Detection Automation in the AI Use Cases domain.
As soon as the asset is created, the Lifecycle Tracker automatically attaches certain out-of-the-box assessment templates, so the governance team can begin documenting business value and risk before the model ever touches production data.
The AI Use Case asset is the anchor for everything that follows. Each model, deployment, and agent created to serve this use case can be linked back to it, giving governance teams a single asset that answers the question: what business outcome is this AI working toward?
Technical layer
AI Model Version: the trained model
Acme's data science team trained a gradient boosting model on 18 months of transaction data in SageMaker and registered it as version 2.1. When the SageMaker integration is connected to Collibra, Collibra automatically ingests the model as an AI Model Version asset named fraud-detector-v2.1 in the AI Models and Agents domain.
This asset is an immutable snapshot, a governed record of the model at the exact point it was approved for use. The team links the transactions_historical data table to this asset as the training data, giving auditors a clear record of what the model was trained on. The Fraud Detection Automation AI Use Case asset is linked here too, connecting business intent to technical implementation.
AI Model Deployment: the live instance
Training a model and running it in production are different events. When the ML team deploys fraud-detector-v2.1 to a SageMaker real-time inference endpoint in us-east-1 with four GPU nodes, Collibra records this as an AI Model Deployment asset.
The deployment asset captures the operational context that the model version does not: the compute environment, the region, and the data flowing through it at runtime. The team links the transactions_live table as the inference data source and the fraud_scores table ( where the model writes its output) as the output data. Any downstream report or decision that reads from fraud_scores can now be traced back to the model that produced it.
AI Endpoint: the stable access point
The payments application needs to call the fraud scorer, but it should not be tightly coupled to a specific deployment. When the ML team rotates from fraud-detector-v2.1 to fraud-detector-v2.2, the payments app should not need to be updated. An AI Endpoint asset is registered for https://api.acmefinance.com/fraud-score , the stable URL that the payments application always calls, regardless of which model version is live behind it.
The endpoint asset is linked to the current deployment. When the underlying deployment changes, only that link needs updating in AI Command Center. The endpoint itself, and the payments application's connection to it, stays the same.
AI Monitor: the watchdog
A deployed model can degrade silently. Acme's risk team configures an AI Monitor asset to watch the deployment continuously. The monitor is set to alert if the model's F1-score drops below 0.82 in a rolling 7-day window, or if the distribution of input features shifts more than 15% from the training baseline, a pattern that signals the real-world transaction mix has changed and the model's training data is no longer representative.
When the monitor fires, the alert surfaces in the AI Command Center Agents dashboard alongside pass rate trends and trust score data for the affected deployment, giving the governance team both the signal and the context in one place.
Agent layer
Six months after the fraud scorer goes live, Acme's operations team wants to automate the first stage of fraud investigation, triaging incoming alerts before they reach a human analyst. They build an AI agent and register it in AI Command Center.
AI Agent: the autonomous investigator
The operations team registers a Fraud Investigation Agent asset in the AI Models and Agents domain. Unlike the fraud-scorer model, which receives a transaction and returns a score, this agent perceives an incoming fraud alert, decides what context it needs, retrieves that context from external systems, and either resolves the case automatically or escalates it to a human with a prepared summary.
The agent asset is linked to the Fraud Detection Automation AI Use Case, keeping the chain of governance intact: from the business objective all the way down to the autonomous system built to serve it.
AI Agent Version: a point-in-time snapshot
The first production release is registered as fraud-investigation-agent-v1.0, an AI Agent Version asset that captures exactly what the agent was doing at launch: which tools it had access to, which underlying model it called for reasoning, and which prompt template it used to generate case summaries. When the team upgrades to a larger model in the next quarter and extends the agent's tool set, they register a new version rather than overwriting the old one, preserving an auditable history of what each version of the agent was capable of.
AI Agent Tool: what the agent can do
Every external capability the agent can invoke is registered as an AI Agent Tool asset and linked to the agent version. For fraud-investigation-agent-v1.0, the team registers four tools.
- CRM Lookup: Retrieves the customer's account history and recent activity from Salesforce.
- Fraud Score: Calls the
https://api.acmefinance.com/fraud-scoreAI Endpoint to get a risk score for the flagged transaction, the same endpoint registered earlier, making the agent's dependency on the fraud-scorer model visible in AI Command Center. - Case Create: Opens a new case in the fraud case management system when the agent determines an alert warrants investigation.
- Escalation: Routes the alert to a human analyst queue, with the agent's summary attached, when confidence falls below the auto-resolve threshold.
Registering each tool as a distinct asset means governance teams know exactly what an agent version can do, and can assess the risk of each capability independently.
The complete picture
By the end of this walkthrough, every asset that touches Acme's fraud detection capability is registered and linked in AI Command Center. Any stakeholder - a risk officer, an auditor, a new team member - can open the Fraud Detection Automation AI Use Case asset and navigate the full picture: the models, the deployments, the endpoints, the agents, the tools, and the governance status of each.
- The Fraud Detection Automation AI Use Case anchors the business intent and carries the compliance assessments.
- The fraud-detector-v2.1 AI Model Version is the immutable, governed snapshot of the trained model, linked to its training data.
- The AI Model Deployment records the live instance: the compute, the region, and the inference and output data assets flowing through it.
- The AI Endpoint provides the stable access point that decouples consumers (the payments app, the agent) from deployment details.
- The AI Monitor watches the deployment continuously and surfaces quality signals in the Agents dashboard.
- The Fraud Investigation Agent, versioned as fraud-investigation-agent-v1.0, operates over the fraud scorer via its registered AI Agent Tools, with full traceability back to the AI Use Case it was built to serve.