Power Automate Object-Centric Process Mining: OCEL Setup, Data Mapping and Multi-Object Analysis

With the rollout of object-centric process mining (OCPM) in Power Automate Process Mining—tracked under Microsoft's AI at Work roadmap ID 576870—Microsoft is giving enterprise IT administrators, process analysts, and Power Platform makers a native way to map and analyze workflows where a single event touches multiple business objects at the same time. Rather than flattening complex relational databases into a single timeline and hoping the resulting numbers still make sense, OCPM preserves the end-to-end relationships between orders, invoices, deliveries, and payments inside one unified model.

Here is a deep dive into how object-centric process mining works in Power Automate, why traditional case-centric logs often distort complex operations, and the exact schema and ingestion rules you need to know before feeding your enterprise event logs into the engine.

What Changed and When: Roadmap 576870 in Context​

According to Microsoft's AI at Work roadmap (ID 576870) and the Power Platform release plan on Microsoft Learn, object-centric process mining in Power Automate Process Mining is now in a Launched state for Worldwide (Standard Multi-Tenant) cloud instances on the web.

The rollout followed a multi-stage enterprise schedule:

• Public Preview: First opened on November 14, 2025 (listed as November 2025 on the AI at Work roadmap).
• General Availability: Reached general availability on June 5, 2026, per Microsoft Learn's Power Platform release plan (following an earlier April 2026 Message Center communication, MC1268680, that had initially targeted May 1, 2026, before settling into the June 2026 general availability window).
• Enablement and Licensing: Microsoft classifies OCPM as a premium feature that is automatically enabled for all Power Automate Process Mining customers across administrator, maker, marketer, and analyst roles. Because regional availability across Azure geographies can vary, administrators should still verify their specific tenant region against Microsoft's Feature Geography report.

Once enabled, the feature equips analysts to ingest object-centric event logs from Azure Data Lake or Microsoft Fabric OneLake, map both object-level and event-level attributes, render customizable multi-object process maps, track throughput times and execution metrics in a Process Overview dashboard, and isolate bottlenecks using specialized multi-object filters or path selection in the Process Analysis widget.

Case-Centric vs. Object-Centric: Why Flattening Data Hurts​

To understand why data architects and process engineers care so much about OCPM, it helps to ask a fundamental question: what actually breaks when you force a complex ERP or supply-chain workflow into a traditional case-centric process mining tool?

In conventional case-centric process mining, every event row must belong to one overarching case notion—say, Order ID. When your workflow is a tightly scoped IT helpdesk ticket that opens, gets assigned, escalates, and closes, a single case ID works beautifully.

Consider, however, the retail scenario documented in Microsoft Learn's OCPM overview:

1. A customer places two separate orders (O1 and O2).
2. To fulfill them, the retailer raises two supplier restocking orders (SO1 and SO2). One supplier shipment arrives and is unpacked; the other requires an update.
3. The store issues two invoices (I1 and I2), updates one of them, and enforces a strict governance policy: do not ship the second customer order until the invoices are paid.
4. A single consolidated payment (P1) arrives covering the billing, and only then does O2 ship.

If you force that everyday scenario into a traditional case-centric model, you immediately run into three painful trade-offs:

• The Duplication Trap (Convergence): If you pick Order as your case ID, what happens when a single invoice or a single payment covers both O1 and O2? You have to duplicate or flatten the invoice and payment events across both order timelines. Suddenly, a single $1,000 payment looks like two $1,000 payments in your aggregate metrics, inflating event counts and skewing financial analysis.
• Lost Context Across Lifecycles: If you pivot your case ID to Invoice or Shipment instead, you lose the upstream story of how customer orders interacted with supplier restocking delays.
• Hidden Governance Dependencies: When compliance auditors ask whether your warehouse is actually obeying the "ship only after payment clears" rule across split orders and consolidated invoices, a case-centric view forces you to stitch separate logs together offline.

In Power Automate's OCPM canvas, the map visualizes object lifecycles with explicit creation (start) and disposal (end) markers for each object type. Activity nodes can belong to one or several object types at once, and color-coded object-flow edges show exactly where the Invoice and Payment streams intersect with the Order stream to gate the final Ship Order activity.

Section Summary: Case-centric mining remains the right tool for straightforward, single-track workflows, and Microsoft notes many teams will continue using both approaches side by side. However, whenever business events touch multiple entities simultaneously, OCPM prevents the event duplication and context loss caused by flattening relational data.

Building an Object-Centric Event Log (OCEL): Schema and Rules​

Before you can admire a multi-colored object graph in Power Automate, you have to feed it properly structured data. Microsoft's documentation for creating an object-centric event log (OCEL) lays out strict schema rules that data engineers should build directly into their ETL pipelines.

Target Schema Requirements​

According to Microsoft Learn, Power Automate Process Mining currently ingests OCPM data from a CSV file stored in Azure Data Lake Storage Gen2 or Microsoft Fabric OneLake, or from a Delta table stored in a Fabric Lakehouse (note that while the data-preparation guide supports both CSV and Fabric Lakehouse Delta tables, the step-by-step ingestion walkthrough specifically emphasizes storing the prepared file as CSV in Azure Data Lake Gen2 or OneLake, so testing CSV first is a smart baseline).

Every OCEL dataset requires the following core columns:

• Activity: A concise, human-readable event label such as Create Order, Issue Invoice, or Receive Payment. Microsoft advises keeping activity names tightly controlled and placing high-cardinality details into separate attribute columns rather than stuffing them into the activity string.
• Timestamp: The exact event time formatted either in 24-hour yyyy-MM-dd HH:mm:ss notation or standard ISO-8601, normalized to UTC or a single consistent time zone so lifecycle ordering stays logical.
• Actor: The user, role, or automated bot (such as ProcBot or Warehouse) executing the event.
• One Column per Object Type: Dedicated columns for each entity class in your scope—for example, Order, Invoice, Payment, and SupplierOrder.
• Optional Object-Type Attributes: Columns capturing financial or operational properties tied to an object, such as PaymentAmount.

The Golden Modeling Rules​

When transforming raw SQL tables or ERP audit trails into an OCEL file, four formatting rules govern whether your model succeeds or turns into spaghetti:

1. One Row Equals One Event: Even if a single warehouse shipment packs items from two orders (O1 and O2), that shipment is recorded as a single row in the log.
2. Use the Pipe (|) Delimiter for Same-Type Multi-Objects: When one event touches multiple instances of the same object type—for instance, a single payment P1 settling two invoices I1 and I2, or one shipment fulfilling O1 and O2—you list both IDs inside that object column separated by a pipe character: I1|I2 or O1|O2.
3. Embrace Sparsity and Opaque IDs: If an event only updates invoice I2 and doesn't directly touch an order or payment at that moment, leave the Order, Payment, and SupplierOrder columns blank for that row. Never reuse identifiers across different entities; keep IDs opaque and stable (O1, I1, P1).
4. Split Events Only When Same-Type Attribute Values Differ: Avoid duplicating object attributes across rows unnecessarily. If an event touches two objects of the same type that require different attribute values on that row, Microsoft instructs makers to split them into separate events so attribute mapping stays unambiguous.

For CSV exports, ensure the file uses UTF-8 encoding, maintains a stable column header order, and quotes any fields that contain commas.

Section Summary: Preparing an OCEL log requires shifting from a flat case table to a sparse matrix where each object type gets its own column and multi-instance references use the pipe (|) delimiter.

Step-by-Step Guide: Ingesting and Mapping OCPM Data in Power Automate​

Once your CSV or Lakehouse table is staged in Azure Data Lake Gen2 or Fabric OneLake, setting up the model inside Power Automate Process Mining follows a straightforward workflow:

1. Launch the OCPM Wizard: In the left navigation pane of Power Automate, select Process mining > Start here, and then choose Object-centric process mining.
2. Name and Connect: Enter a descriptive Process name, select either Azure Data Lake or OneLake under the Data source heading, and click Continue.
3. Configure or Skip Fabric Semantic Model Export (Preview): At the Choose where to export screen, Power Automate gives you the option to export the resulting OCPM semantic model directly into a Microsoft Fabric workspace (requiring you to pick a Fabric workspace, a Fabric Lakehouse, and a default report name) or simply select Skip export. Because Microsoft Learn marks the Fabric semantic model export capability as preview, production teams evaluating core OCPM analytics can safely skip this step unless they specifically want to build custom downstream Fabric reports.
4. Authenticate to Storage: Complete the connection setup prompts for your chosen Azure Data Lake Gen2 container or Fabric OneLake path until you arrive at the Map your data screen.
5. Map Mandatory and Optional Attributes: On the mapping screen, assign your dataset columns to Power Automate's schema types:
   • Required: Exactly one Activity attribute, one Start event timestamp attribute, and at least two Object type attributes.
   • Optional: End event timestamp and Resource (Actor) attributes.
   • Recommended: Event level and Object level attributes (including financial metrics). Note the architectural distinction here: an event-level attribute binds to the event across every object referenced in that row, whereas an object-level attribute binds to a specific object type and pulls its value from either the first event or last event referencing that object.
6. Run the Engine: Select Save and Analyze. Once ingestion finishes, Power Automate opens the Process Intelligence experience for OCPM directly on the Process overview tab.

Navigating the Graph: How Object Proximity Filtering Works​

Having five or six intersecting object lifecycles on a single screen is a massive analytical upgrade, right up until the map looks like a plate of multicolored linguine. To slice through the visual noise without breaking cross-object links, Power Automate Process Mining includes four OCPM filter types in the top toolbar of the active tab:

• Object Type Filter: Temporarily checks or unchecks entire entity classes (for example, hiding Items to focus strictly on Orders and Customers).
• Object Attribute Filter: Segments the graph by cohort properties such as customer region, product category, or high-value order thresholds.
• Timeframe Filter: Restricts the analysis to objects active during a predefined window or custom date range.
• Object Proximity Filter: This is the secret weapon of OCPM. Because objects connect to one another through shared events, the proximity filter lets you choose a specific object type and instance, and then define a Proximity level (degree of separation) to determine how far out along the relational web the filter should reach:
  • Level 1: Isolates only the selected object itself.
  • Level 2: Includes all directly connected objects linked via common events—including other objects of the same object type (for instance, finding a second customer order that shared a consolidated shipment or invoice with your target order).
  • Level 3 and higher: Recursively adds indirectly connected objects tied to the entities discovered in the previous level, exposing ripple effects across your supply chain or billing network.

Critical Analysis: Separating Marketing Hype from Enterprise Reality​

If you read Microsoft's official release plan entry for Roadmap ID 576870, the business value section reads like an executive victory lap: it promises "transformative insights," immediate visibility into cascading delays, and a reduction in root-cause investigation time "from weeks to hours".

Let's apply a healthy dose of veteran IT skepticism here. Will OCPM automatically turn weeks of operational firefighting into a two-hour coffee break? Not until you do the data engineering legwork.

In practice, the hardest part of object-centric process mining isn't clicking Save and Analyze inside Power Automate; it is extracting clean foreign-key relationships out of legacy ERP systems (like SAP, Oracle, or Dynamics 365) where order tables, warehouse management logs, and accounts-receivable ledgers often live in different schemas with mismatched timestamps. If your upstream data warehouse loses the link between SupplierOrder SO1 and CustomerOrder O1, or if system clocks across departments drift by a few minutes, OCPM will faithfully visualize that disorder.

That said, once an organization builds a repeatable ETL pipeline into Azure Data Lake Gen2 or Fabric OneLake, the analytical payoff is genuinely substantial. In academic and enterprise process mining circles—where Object-Centric Process Mining was pioneered by researchers like Wil van der Aalst to solve the notorious convergence and divergence bugs of XES/case-based logs—the shift from single-case timelines to multi-object graphs is widely recognized as the biggest methodological leap in the field in over a decade. By baking OCPM natively into Power Automate Process Mining alongside OneLake and Fabric Lakehouse storage, Microsoft is sparing enterprise teams from having to export sensitive operational logs into specialized third-party mining suites just to track a split shipment.

Practical Next Steps for Windows and Power Platform Admins​

If your organization already holds Power Automate Process Mining licenses, you can put OCPM to the test today:

• Start Small with Two or Three Object Types: Resist the urge to model fifteen ERP tables on day one. Follow Microsoft's construction guidance and pick a tight triad—such as Order, Invoice, and Payment—where a known compliance rule or bottleneck exists.
• Validate Pipe-Delimited Joins: Check that your staging scripts properly group multi-object events onto a single row using the | separator rather than emitting duplicate timestamp rows.
• Use Level 2 Proximity Filters for Root-Cause Hunts: When investigating delayed orders, use a Level 2 Object Proximity filter to immediately see whether a stalled order is waiting on a shared batch invoice or a consolidated supplier delivery.

For discussions on automating your OneLake data pipelines, optimizing Power Platform licensing, or troubleshooting Dataverse and Fabric integrations, check out our dedicated Power Platform and Enterprise IT Administration sections right here on WindowsForum.com.