Process mining is the practice of reconstructing how work actually flows through your business from the timestamps your systems already record. Every time an order is created, approved, invoiced or shipped, your ERP, CRM or hospital system writes a log entry. Put those entries in order for each case and you get a factual map of the process, including every detour, loop and delay.
That map answers the question most automation programmes get wrong: what should we automate first? Instead of picking the process the loudest manager complains about, you pick the one where the data shows the most volume, rework and waiting time.
This guide explains how process mining and task mining work, what data you need, which tools are mature in 2026, and a practical method for a mid-sized business in India, the UAE or the USA to turn logs into an automation roadmap.
What process mining is, and is not
Process mining reads event logs and builds a model of the process as it really runs. It is not a workshop where people describe the process from memory, and it is not a BI dashboard of totals. Its value comes from following individual cases end to end.
Three classic uses:
- Discovery: draw the actual process map, with every variant and how often each occurs.
- Conformance: compare reality with the intended process, such as "every purchase over a threshold needs two approvals".
- Enhancement: overlay time, cost and resource data to find bottlenecks and rework.
The market has matured. Gartner's 2025 Magic Quadrant for process mining platforms assessed 16 vendors, and industry coverage of that report highlights AI infusion, consolidation and object-centric process mining as the directions the market is moving in.
Process mining vs task mining
The two are complementary and often confused.
| Process mining | Task mining | |
|---|---|---|
| Data source | System event logs | Desktop activity on user machines |
| What it sees | Steps recorded in systems | Clicks, apps, copy-paste between systems |
| Best for | End-to-end flow, bottlenecks, rework | Manual work between systems |
| Privacy impact | Lower (system data) | Higher (records employee screens) |
| Typical output | Process map, variants, cycle times | Repetitive task patterns, automation candidates |
Process mining tells you that invoices wait four days between receipt and approval. Task mining tells you that during those four days, an accounts clerk copies values from a PDF into the ERP field by field. You need both to design the fix.
The data you need: event logs
The minimum event log is simple. Microsoft's process mining documentation lists the same three core fields every tool needs:
- Case ID: the thing moving through the process, such as an order number, invoice number or patient visit ID.
- Activity: what happened, such as "PO created", "Invoice received", "Payment released".
- Timestamp: when it happened.
Useful additional fields include the user or team, the branch, the amount, the customer segment and the system. These let you slice the map, for example "approval delays are worst in the Dubai branch for orders above a certain value".
Where the data comes from:
- ERP and accounting: change logs and document flow tables in SAP, Oracle, Tally-based setups or Zoho.
- CRM and ticketing: status change history.
- Custom applications: audit tables, if your developers built them. If not, adding a proper audit log is often the first improvement.
- Hospital systems: registration, consultation, lab order, result, billing and discharge timestamps per visit.
Object-centric process mining
Classic process mining assumes one case ID. Real processes rarely work that way. One sales order can have several deliveries, each delivery several invoices, and one payment can cover several invoices. Forcing all that into one case ID creates misleading maps.
Object-centric process mining (OCPM) models events that relate to several objects at once. The OCEL 2.0 standard defines a common format for this kind of log, and the open-source PM4Py library supports it, according to the authors of the object-centric event data standard. Commercial platforms are adding object-centric features as well.
For a mid-sized business, the practical advice is: start with a single case notion for one process, and move to object-centric analysis when you need to see order-to-cash or procure-to-pay across several linked documents.
How to pick what to automate first
Process mining gives you numbers; you still need a scoring method. A simple, defensible approach:
| Criterion | What to measure | Why it matters |
|---|---|---|
| Volume | Cases per month | Savings scale with volume |
| Manual touch time | Minutes of human work per case | Direct effort saved |
| Waiting time | Time between steps | Often bigger than work time |
| Rework rate | Cases with repeated steps | Signals errors worth fixing |
| Variant count | Number of distinct paths | Fewer variants are easier to automate |
| Data quality | Structured vs unstructured input | Affects tool choice |
| Risk | Financial or compliance impact of errors | Sets how much review is needed |
Score each candidate from 1 to 5, weight volume and touch time highest, and sort. Processes with high volume, high touch time, few variants and structured input go first. High-variant, unstructured processes may need redesign or an AI-assisted approach, which our comparison of RPA vs agentic automation covers.
A worked example: procure-to-pay in a distribution company
Consider a hypothetical distributor with branches in Kochi, Bengaluru and Dubai, using an ERP for purchasing and accounts.
- Extract: export PO creation, approval, goods receipt, invoice receipt, invoice posting and payment events for twelve months.
- Discover: the map shows the "happy path" covers only a minority of invoices. Many go through a "price mismatch" loop and a manual re-approval.
- Measure: the median time from invoice receipt to posting is several days, almost all of it waiting in an inbox.
- Drill in: task mining on two clerks' machines shows most posting time is copying fields from PDF invoices.
- Decide: automate invoice capture with document AI, auto-match against PO and receipt, and route only mismatches to a person. Separately, fix the price master that causes most mismatches.
- Verify: rerun the mining three months later and compare cycle time, touch time and rework.
That last step matters. Process mining is as useful after automation as before, because it proves (or disproves) the gain.
Tools: commercial, platform-bundled and open source
- Commercial platforms: Celonis, SAP Signavio, ARIS, Apromore, IBM and others offer deep process mining, connectors and increasingly AI assistants on top. Strong for large ERP estates.
- Platform-bundled: Microsoft Power Automate includes process mining and task mining, which can be a pragmatic choice if you already run Microsoft 365 and Power Platform.
- Open source: PM4Py (Python) is a capable library for discovery, conformance and object-centric analysis. Good for a technical team, a proof of value, or embedding mining in your own product.
- Custom analytics: for a single process, a well-built event table in your data warehouse plus a few queries can answer most questions.
A newer trend is exposing process mining data to AI agents through protocols such as the Model Context Protocol, so an agent can check process KPIs or bottlenecks before acting. This is promising but early; treat it as an add-on, not the starting point.
Common pitfalls
- Bad timestamps: batch jobs that stamp everything at midnight, or time zones mixed across branches.
- Missing activities: steps done in email or WhatsApp never appear in logs. Note the gaps rather than pretending they do not exist.
- Too many variants: a raw map with hundreds of paths is useless. Filter to the top variants that cover most cases first.
- Analysis without an owner: findings die unless a process owner commits to act on them.
- Ignoring people: a map can show where work waits, not why. Always validate findings with the team that does the work.
What results to expect
A first process mining exercise rarely produces a single dramatic insight. It usually produces three kinds of findings:
- Fixes that need no automation: a broken master-data field, a redundant approval, a report nobody reads.
- Clear automation candidates: high-volume, structured, low-variant steps with measurable manual effort.
- Redesign candidates: processes with so many variants that automating them as-is would encode the mess.
Expect the first category to be the quickest win and the second to justify the investment. The third is where you need a conversation with the business before any tool is chosen.
How to start in 30 days
- Week 1: pick one process with a clear case ID and good system coverage, such as procure-to-pay, order-to-cash or patient visit flow.
- Week 1: extract twelve months of events with case ID, activity, timestamp and two or three attributes.
- Week 2: clean timestamps, map activity names to a consistent list, and load into a tool or PM4Py.
- Week 3: discover the map, measure cycle times, rework and waiting, and score candidate improvements.
- Week 4: present three ranked opportunities to the process owner, each with expected impact and a measurement plan.
Our data analysis and workflow automation teams run this kind of assessment and then build the fixes. For hospitals, the same method applies to patient flow; our guide to hospital KPIs shows which timestamps matter, and healthcare data analytics covers the data foundation. If the analysis points to orchestration work, our n8n guide is a good next read.
Conclusion
Process mining turns automation from a matter of opinion into a matter of evidence. With three fields from your existing systems, you can see how work actually flows, where it waits and where it loops. Use that to rank opportunities by volume, effort and risk, fix broken steps before automating them, and mine again afterwards to prove the result. It is one of the cheapest ways to make sure the next automation project is the right one.
Frequently asked questions
What is process mining in simple terms?
Process mining uses timestamps from your business systems to reconstruct how each case, such as an order or invoice, actually moved through your process. It reveals real paths, delays, loops and deviations from the intended process.
What is the difference between process mining and task mining?
Process mining analyses system event logs to show end-to-end flow. Task mining records user desktop activity to show manual work between systems, such as copying data. Process mining finds where time is lost; task mining shows what people do during that time.
What data do I need for process mining?
At minimum, a case ID, an activity name and a timestamp for each event. Extra attributes such as user, branch, amount or customer segment make the analysis far more useful.
Is process mining only for large enterprises?
No. Mid-sized businesses can mine a single process using open-source tools like PM4Py or bundled tools like Power Automate Process Mining. The main requirement is reasonably clean system logs.
What is object-centric process mining?
Object-centric process mining handles events linked to several objects at once, such as one payment covering several invoices. The OCEL 2.0 standard defines a log format for it, and it gives more accurate maps for order-to-cash and procure-to-pay.
Is task mining legal under India's DPDP Act?
Task mining captures employee activity, which can include personal data. You should give clear notice, limit capture to work applications, mask sensitive fields, set retention periods and document the purpose, and take legal advice for your situation.
