Process mining in Power Automate

Power Automate Process Mining (originally acquired as Minit, now an integrated capability) analyses event logs from business systems to reconstruct what processes actually do — not what the documentation says they do, not what stakeholders believe they do, but what happens in practice. The output is a quantified map of process variants, durations, and exception paths, with automation recommendations.

The core idea. Every business process leaves a trail of events in source systems — "order created", "approval granted", "invoice posted", "payment received". Stitching these events together by a common case ID (order number, invoice number) reveals the actual flow: how long each step takes, which variants are common, where rework happens, where SLAs miss.

Input data — the event log. The minimum:

• Case ID — the entity being tracked (order, ticket, invoice).
• Activity name — what happened.
• Timestamp — when.

Optional but valuable:

• Resource — who did it (user or team).
• Cost — financial impact.
• Outcome — final status.
• Attributes — customer type, region, amount, anything segmentable.

The event log is the artefact process mining consumes. Building it cleanly from source systems is the bulk of the work.

Data sources.

• Dynamics 365 / Dataverse — out-of-the-box connectors pull activity records.
• SQL Server / databases — query historical event tables.
• SAP, Salesforce — via connectors or extracted exports.
• Custom systems — CSV upload or custom connector.
• Process desktop activity — process advisor recording desktop UI events.

What process mining produces.

• Process map — a directed graph of activities with frequencies and durations on edges.
• Variants — every unique path taken; ranked by frequency.
• Bottlenecks — activities or transitions with long wait times.
• Rework loops — activities that occur multiple times per case.
• Compliance deviations — paths violating expected process rules.
• Filtering and drill-down — slice by attribute (region, customer segment, time period).

Variants. Real processes diverge into dozens or hundreds of variants:

• Variant 1 (60% of cases) — happy path, all steps in order.
• Variant 2 (15%) — manual exception handling.
• Variant 3 (5%) — rework loop.
• Long tail — hundreds of low-frequency variants with idiosyncratic paths.

Quantifying variants reveals where standardisation could yield large gains.

Automation recommendations. Process Mining identifies activities suitable for automation:

• High-frequency manual activities — automation impact is volume × per-instance time saving.
• Repetitive rework loops — pre-validation could eliminate.
• Approval bottlenecks — automated routing or AI-based pre-decisioning.

Recommendations link directly to Power Automate templates — the AI suggests a flow to start with.

Process Advisor — desktop process recording. A complementary capability: a recording agent on a user's desktop captures clicks, keystrokes, and screen events. Multiple recordings cluster into a "as-is" desktop process, ready for RPA automation. Useful for processes whose digital footprint isn't in events but in user actions.

Use cases where process mining wins.

• AP processes — invoice arrival to payment, where rework and exceptions cost real money.
• Order to cash — quote to invoice cycle time analysis.
• Customer service tickets — resolution time, escalation patterns.
• Loan applications — approval bottleneck analysis.
• Manufacturing — order to ship cycle decomposition.

Use cases where it doesn't fit well.

• Processes with poor event capture — manual processes without timestamps in systems.
• Highly variable creative work — no repeating pattern to mine.
• Strategic processes — board decisions, M&A; the value is in judgement, not flow.

The hard part — data preparation. Getting clean event data is 70%+ of a process mining project:

• Identifying the case ID across systems.
• Reconciling activity names — "Invoice Received" in one system, "Invoice Captured" in another, same event.
• Filtering noise — system events, not business events.
• Handling missing timestamps — gaps in event capture.
• Volume management — millions of events per process.

Underestimating this is the most common project failure mode.

Compliance and conformance checking. Beyond exploration, process mining can check whether actual paths conform to a defined reference model. Deviations are quantified — useful for regulatory processes where prescribed sequence matters.

Common pitfalls.

• Boil-the-ocean scope. Trying to mine every process at once. Start with one high-impact, well-instrumented process; demonstrate value; expand.
• Ignoring the data preparation cost. Sponsors expect insights in days; reality is weeks of data work first.
• Action follows insight, but slowly. Mining reveals problems; fixing them requires process redesign and change management.
• Variants over-aggregated or under-aggregated. Too aggregated, you miss patterns; too disaggregated, the long tail is noise. Tune.
• Privacy. Event logs include user IDs — process mining requires DPIA in regulated industries.

Operational guidance. Process mining is a powerful diagnostic tool; treat it as a continuous capability, not a one-time project. Mine, fix, mine again. The teams that get value embed mining into their continuous improvement cadence and tie insights to specific automation, process redesign, or training actions.