Process Mining: What It Is and How AI Changes It

2026-11-07 · Tommaso Maria Ricci

Most companies have never seen their own processes. They have seen the flowchart someone drew for an ISO audit, the swimlane in a consulting deck, the "happy path" a vendor configured during an ERP rollout. Process mining shows something different: what actually happened, case by case, reconstructed from the timestamps your systems already record. And the gap between the drawing and the data is almost always larger than anyone in the room expects.

That gap is now a budget line. According to Gartner, worldwide spending on process mining software grew by more than 30 percent in 2024, after 39.5 percent growth the year before, and is expected to pass $2 billion by 2028. Deloitte's 2025 survey found that 48 percent of organizations using process mining have rolled it out company wide. Yet only a quarter use it together with AI, which is exactly where the next wave of value sits.

This guide explains what process mining is, how it works, where it pays back, how AI changes it, and how to run a first project without buying a platform you do not need yet. I have spent more than twenty years building companies and fixing operations from the inside, and the pattern is consistent: the teams that automate after they measure win, and the ones that automate first spend the next year undoing it.

What is process mining, in plain terms

Process mining is a family of techniques that reads event data from your IT systems and rebuilds the real flow of a business process from it. Every time someone creates a purchase order, approves an invoice, changes a delivery date or closes a ticket, a system writes a record with a timestamp. Put enough of those records together and you can see the sequence of steps each case followed.

The minimum ingredients are three fields per event:

Case ID: the thing moving through the process, such as an order number, an invoice, a claim or a patient visit.
Activity: what happened, for example "invoice received", "price changed", "goods receipt posted".
Timestamp: when it happened.

With only those three columns, an algorithm can draw a map of every path cases took, count how often each path occurs, and measure how long cases waited between steps. Add more attributes (supplier, region, amount, user, product line) and you can slice the map to find out why some paths are slow or expensive.

The field has an academic backbone. The Process Mining Manifesto, written by the IEEE Task Force on Process Mining with more than 75 contributors from over 50 organizations and chaired by Wil van der Aalst, set out the shared vocabulary in 2011 and was published in 2012. The task force also produced XES, the standard format for event logs, revised as IEEE 1849-2023 in September 2023.

The three things process mining does

The manifesto describes three core uses, and they still map well to business value:

Discovery: build a process model from the event log, with no prior model. This is the "show me what really happens" moment.
Conformance checking: compare the real flow against the intended one. Which invoices were paid without a goods receipt? Which orders skipped credit checks?
Enhancement: enrich the model with performance data, such as waiting times, rework loops, costs, and the attributes that predict delays.

Most first projects start with discovery and end up spending most of their value in conformance and enhancement. Discovery tells you the process is messy. Conformance and enhancement tell you which mess costs money.

What process mining is not

It is worth clearing up three common confusions, because they shape what you buy and who you staff.

It is not business intelligence. A BI dashboard tells you that average invoice cycle time is 23 days. Process mining tells you that 31 percent of invoices go through a price correction loop that adds nine days, and that the loop is concentrated in four suppliers.

It is not process modeling. Modeling is drawing how a process should work. Mining is measuring how it does work. You need both, but they answer different questions.

It is not automation. Process mining finds where automation would help. It does not by itself remove a single click. Confusing the two is why so many robotic process automation programs automated broken steps at speed.

Why process mining matters now: the numbers

Process mining has existed in research since the early 2000s. Three things changed in the last few years and pushed it from a niche analytics technique into a mainstream enterprise purchase.

The market is growing fast

Gartner's 2025 Magic Quadrant for Process Mining Platforms, summarized by Process Excellence Network, puts 2023 market revenue at $871.6 million and expects spending to pass $2 billion by 2028, a compound annual growth rate of about 18 percent. The report evaluated 16 vendors. It also flagged consolidation: smaller vendors that cannot scale are likely acquisition targets, and large enterprise software companies are moving in.

For a buyer, that has two consequences. First, prices and packaging will keep shifting, so avoid multi-year lock-in on a first project. Second, process mining is increasingly a feature inside platforms you may already pay for, which changes the build or buy calculation.

Adoption is maturing, but value is uneven

The Deloitte Global Process Mining Survey 2025, which covered more than 120 respondents, gives a useful picture of where organizations stand:

48 percent report company-wide adoption, up from 45 percent in 2021.
61 percent expect process improvement as the main benefit, and 59 percent expect cost savings, up from 46 percent.
80 percent agree or strongly agree that process mining delivers added value.
41 percent cite lack of management attention as a challenge, up from 26 percent in 2021.
55 percent say the C-suite actively supports their efforts.

Read those together and the story is clear. The technology works for most people who deploy it seriously. What fails is sponsorship. Process mining surfaces uncomfortable facts about how work really gets done, and without an executive who wants to hear them, the insights die in a slide.

AI raised the stakes

The same Deloitte survey found that only 25 percent of respondents actively combine AI with process mining, while 67 percent do not use AI in their process mining initiatives at all. That is a gap, not a verdict. Every serious AI program in operations needs to know where work flows, where it waits and where decisions are made. Process mining is the cheapest way to get that map before you put agents, copilots or automation on top of it.

Gartner names object-centric process mining as one of the main trends. Instead of following a single case type, such as an order, it follows several linked objects at once: orders, deliveries, invoices and payments, which in real life do not line up one to one. One order can trigger three deliveries and two invoices. Classic process mining flattens that and distorts the picture. Object-centric methods keep it intact, which matters a lot when you want an AI system to act across functions.

Process mining applications: where it pays back

Process mining applies to any process that leaves a digital trail. In practice, value concentrates in a handful of areas, and the Deloitte survey confirms that procurement and accounting remain the core, with supply chain, IT operations, claims and customer service growing.

Procure to pay

This is the classic first use case because the data is clean (it lives in the ERP) and the money is visible. Typical findings:

Maverick buying: purchase orders created after the invoice arrives.
Price and quantity changes after approval, which create rework and leakage.
Lost discounts because invoices sit in approval queues past early payment windows.
Duplicate payments and invoices paid without a matching goods receipt.

If you want a process baseline before mining, the controls in our guide to procure to pay process controls are a good checklist to compare the real flow against.

Order to cash

On the revenue side, process mining shows where orders stall: credit blocks, manual pricing overrides, partial deliveries, billing errors that trigger disputes. The metric that usually moves is days sales outstanding, because every rework loop before the invoice delays the cash. The stages and owners we describe in order to cash process steps and owners give you the target model for conformance checking.

Finance close and reconciliation

Month end close is a process with a deadline and a lot of handoffs. Mining the journal entry and reconciliation logs shows which accounts get reopened, which adjustments arrive late and who waits for whom. It pairs naturally with automation: once you know which reconciliations are repetitive and rule based, reconciliation automation becomes a targeted investment instead of a guess.

Customer service and claims

Ticketing systems and claims platforms log every status change. Mining them reveals ping-pong between teams, tickets reopened after closure, and claims that wait for documents nobody requested. Here the payback is often customer experience as much as cost.

IT operations

Incident and change management tools produce rich event data. Process mining shows how incidents really move between support levels, how often changes bypass approval, and where escalations cluster. Deloitte lists IT operations among the growing application areas.

Healthcare and service operations

Any service business that schedules people through steps can benefit. In a medical center I worked with, capacity grew by 20 percent. The work started from the question process mining asks by design: where does a case actually wait, and why? You do not always need a platform to ask it, but you always need the data.

How process mining works, step by step

A process mining project has five technical stages. Each one has a typical failure mode, and knowing them in advance saves weeks.

Step 1: Choose the process and the question

Do not start with "let's mine everything". Start with a process that has a clear owner, a measurable pain and data in one or two systems. Then write the business question in one sentence, for example: "Why do 30 percent of our invoices miss the payment term?"

Failure mode: picking a process because the data is easy, not because the pain is real. You get a beautiful map nobody acts on.

Step 2: Extract the event log

Identify the tables or logs that record the steps, map them to case ID, activity and timestamp, and extract a sample. For ERP systems this usually means joining several tables. For SaaS tools it often means API exports or change history.

Failure mode: missing activities. If approvals happen in email, the log will show a mysterious gap. Document what the data cannot see before you interpret anything.

Step 3: Clean and validate

Expect duplicate events, timestamps in different time zones, batch jobs that write hundreds of events at the same second, and activities named three different ways. Spend time here. Then validate a handful of cases end to end with someone who does the work every day.

Failure mode: skipping validation. A process map built on bad data is worse than no map, because it looks authoritative.

Step 4: Discover and analyze

Run discovery, look at the variants (the distinct paths cases follow) and sort them by frequency and duration. Typically a few variants cover most cases, and a long tail of rare variants carries a disproportionate share of delay and cost. Then use conformance checking against the intended process and slice by attributes to find root causes.

Failure mode: falling in love with the spaghetti diagram. The full map of a real process is unreadable. Filter aggressively and focus on the variants that matter for your question.

Step 5: Act and monitor

Turn findings into changes: a policy fix, a system configuration, a training gap, an automation. Then keep the log refreshing and track whether the change moved the metric. This is where process mining becomes process intelligence: a continuous view, not a one-off audit.

Failure mode: treating the project as a report. If nobody owns the actions and nobody watches the metric, the process drifts back within a quarter.

Process mining before automation: where to start

This is the section most readers came for, because it is where money is won or lost. Automation without process data is a bet. Automation after process mining is an investment with a baseline.

Here is the order I recommend, and the reasoning behind it.

1. Measure the variant spread before you automate anything

If 80 percent of cases follow two or three variants, automation is straightforward: build for those paths and route exceptions to people. If cases spread across dozens of variants with no dominant path, automating now locks the chaos in code. Standardize first.

2. Kill rework loops before you speed up steps

The biggest waste is rarely a slow step. It is a step that happens twice: a price corrected after approval, a ticket reopened, a document requested again. Removing a loop saves the whole loop. Speeding up one step saves seconds.

3. Automate handoffs and waiting, not just tasks

Process mining usually shows that cases spend far more time waiting between steps than being worked on. That waiting is where workflow automation and notifications pay back fastest. Our workflow automation guide covers how to design those handoffs.

4. Only then automate the tasks inside the steps

Once the flow is clean, task-level automation (data entry, matching, classification, drafting) has a stable target. This is where AI models are most useful, because they deal with the unstructured inputs that rules cannot handle. For the broader picture of what to automate first, see our business process automation guide.

5. Keep mining after go-live

Automation changes the process. New exceptions appear, bots fail silently, people find workarounds. Keep the event log flowing and compare the post-automation process against the baseline. If you cannot show the before and after, you cannot defend the next investment.

If you are about to sign an automation contract and you do not yet have this baseline, it is worth pausing for two or three weeks to build it. A short conversation with someone who has done it before often saves months; you can request a consultation from the contact page of this site and we will look at your process data together.

How AI changes process mining

AI and process mining reinforce each other in both directions. AI makes process mining easier to use and more predictive. Process mining makes AI safer and better targeted.

AI inside process mining

Natural language queries. Instead of building filters, an analyst asks "which suppliers cause the most price changes after approval?" and gets an answer with the supporting cases. This lowers the skill barrier, which matters because process mining expertise is scarce.
Predictive monitoring. Machine learning models trained on historical cases predict which open cases will be late or will need rework, so teams can act before the deadline instead of after.
Root cause suggestions. Algorithms rank the attributes most associated with a bad outcome, such as long duration or a compliance violation. A human still decides whether the correlation is causal.
Event log preparation. Generative AI helps map messy activity names, suggest table joins and document data quality issues. This is the most underrated use, because data preparation is where most project time goes.

Process mining for AI

This direction matters more for most companies right now.

Targeting. Process mining shows where AI would remove the most waiting, rework or manual effort. Without it, AI use cases are chosen by enthusiasm, not by impact.
Grounding agents. AI agents that act across systems need to know the real process, including the exceptions. A process model and its variants are the best map you can give them.
Guardrails and monitoring. Once AI systems take actions in a process, conformance checking tells you whether they follow the rules. It is one of the few practical ways to audit what an agent did at scale.
Measuring value. The event log before and after deployment is the cleanest evidence of whether an AI initiative paid back.

This is why the 67 percent of organizations that do not combine AI with process mining are leaving value on the table. Our enterprise AI adoption framework treats process visibility as a prerequisite for exactly this reason.

Task mining: the missing piece

Process mining sees what systems record. It does not see the copy and paste between a spreadsheet and the ERP, the lookup in a browser tab, or the email that triggered the step. Task mining captures user interaction data on desktops (with consent and privacy controls) to fill that gap. It is useful when a lot of work happens outside the core systems, but it raises real privacy and works council questions. Treat it as a second phase, with legal and HR involved from the start.

Build, buy or borrow: choosing your process mining approach

You have three realistic options, and the right one depends on how many processes you plan to mine and how often.

Option 1: A dedicated process mining platform

Best when you plan to mine several processes continuously, have large data volumes and want enterprise features like object-centric analysis, connectors and governance. The trade-off is cost and implementation time. Expect a meaningful license and an integration effort measured in weeks or months.

Option 2: Process mining inside a platform you already use

Several ERP, automation and low-code vendors now include process mining capabilities. Best when your process lives mostly in one ecosystem and you want quick wins. The trade-off is depth and portability: analysis across systems can be harder.

Option 3: Open source and a project team

Academic and open source tools can discover process models from an exported event log. Best for a first proof of concept, a one-off analysis, or a team with data skills that wants to learn before buying. The trade-off is scalability and maintenance.

For a broader decision framework, see our analysis of build vs buy for AI software. My bias for a first project is clear: prove value on one process with the lightest tool that answers your question, then buy for scale once you know what scale means for you.

Questions to ask any vendor

Which of our systems do you connect to natively, and what is custom?
Do you support object-centric analysis, or do we have to flatten our data to a single case?
How do you handle data that is spread across several systems?
What does the event log refresh look like after go-live, and who maintains it?
How is access controlled, and can we mask personal data?
What does pricing look like when we add the second and third process?

The process mining metrics that actually matter

A process map is a picture. Decisions need numbers. These are the metrics I ask for in every process mining review, because each one points to a specific type of fix.

Throughput time, split into work and wait

Total cycle time hides the real story. Split it into time when someone is working on the case and time when the case is waiting. In most administrative processes waiting dominates by a wide margin. That tells you to fix queues, approvals and handoffs before you invest in making individual steps faster.

Variant concentration

Count how many distinct paths cases follow and what share of cases the top five variants cover. High concentration means the process is ready for automation. Low concentration means you need standard work, clearer rules or better master data first.

Rework rate

Measure the share of cases where an activity repeats: a price changed twice, an approval requested again, a ticket reopened. Rework is the most direct measure of waste, and it usually has a small number of root causes you can name.

Conformance rate

Measure the share of cases that follow the intended process, and list the most frequent deviations. Not every deviation is bad. Some are smart workarounds that should become the new standard. But deviations that bypass controls, such as payment before goods receipt, are compliance risk with a number attached.

Automation rate

For each activity, measure the share performed by a system user or bot versus a human. This shows where automation already exists, where it fails and falls back to people, and where the next automation would remove the most manual effort.

First time right

Measure the share of cases that go from start to end on the shortest path without exceptions, changes or loops. It is the single number that best summarizes process health, and it is easy for executives to understand. Track it monthly and it becomes the headline metric for your whole program.

Pick two or three of these for your first project, link each one to money (hours, discounts, penalties, working capital), and report them every month. That discipline is what turns a process mining tool into a management habit.

Process mining readiness scorecard

Score each statement from 0 (not true) to 2 (fully true). Be honest; the point is to find the weak spot before a vendor does.

We have one process with a clear owner and a measurable pain point.
That process runs mostly in one or two systems that log timestamps.
We can extract the data without a six-month IT project.
Someone who does the work every day can validate the data with us.
We have an executive sponsor who will act on uncomfortable findings.
We know the intended process well enough to check conformance against it.
We have a baseline metric (cycle time, cost per case, error rate) today.
We can involve legal and data protection early if personal data is involved.
We have, or can borrow, someone with data preparation skills.
We have budget and authority to change the process after the analysis.

How to read your score:

16 to 20: you are ready. Pick the tool that fits your systems and start.
10 to 15: run a proof of concept on one process with a light tool. Fix the gaps it reveals before scaling.
Below 10: do the groundwork first. Usually the missing piece is sponsorship (statements 5 and 10) or data access (statement 3), not technology.

A 30, 60 and 90 day roadmap for your first process mining project

Days 1 to 30: scope and data

Pick the process and write the business question in one sentence.
Name the process owner and the executive sponsor.
Map the systems involved and identify the tables or logs that record each step.
Extract a sample event log covering at least three months of cases.
Agree on data protection measures: pseudonymize user and customer fields where you can.
Define the baseline metric you want to move.

Days 31 to 60: discovery and diagnosis

Clean the log and validate ten to twenty cases end to end with people who do the work.
Run discovery and identify the top variants by frequency and duration.
Run conformance checking against the intended process.
Slice by attributes (supplier, region, product, team) to locate the root causes.
Size the opportunity in money: hours of rework, lost discounts, late payments, penalties.
Present three to five findings with numbers to the sponsor, not a map with 200 boxes.

Days 61 to 90: action and monitoring

Choose two or three changes with the best value to effort ratio.
Implement them: policy, configuration, training or targeted automation.
Set up a regular refresh of the event log and a simple monitoring view.
Compare the metric against the baseline after four to six weeks.
Decide whether to extend to the next process, and with which tool.

At the end of 90 days you should have one process measurably improved and a clear view of whether a platform investment makes sense. If you want help sizing that second step, the consultation request on this site is the fastest way to start; bring your event log sample and your baseline metric.

The mistakes I see most often

Starting with the tool instead of the question

The most expensive mistake is buying a platform before knowing which decision it will inform. The license is bought, the connector is built, the map is drawn, and then nobody knows what to do with it.

Mining a process nobody owns

If no single person is accountable for the end to end process, findings fall between departments. Procure to pay crosses procurement, operations and finance. Someone has to own the whole thing, or at least own the project.

Ignoring what the data cannot see

Approvals in email, decisions in meetings, work in spreadsheets. If you do not document these blind spots, you will misread gaps as delays and blame the wrong team.

Using process mining to police people

When people learn the tool is used to rank individuals, they change how they log work, and the data quality collapses. Focus on process design and system configuration, not personal performance. In Europe, involving the works council early is not optional in many cases.

Stopping after the first report

The first analysis is the cheapest one. The value comes from monitoring changes over time. A process mining project that ends with a PDF has spent most of the cost and captured a fraction of the benefit.

Automating before standardizing

I have said it already, and I will repeat it because it is the costliest error of all: automating a process with dozens of variants freezes the chaos. Our month-end close guide shows the same principle on a finance process: define the steps and owners, then automate.

What results look like in practice

Process mining results are rarely one dramatic number. They are a set of specific, connected fixes. In the businesses I have worked with, the pattern behind large outcomes was always the same: first see the real flow, then remove what does not add value, then automate what remains. A hotel that grew revenue from 9 million to 10 million did not get there with one tool. A sports distribution company that grew sales by 30 percent with AI marketing first had to understand where its leads and orders actually stalled. The method is the same whether you call it process mining or not: data first, opinions second.

Realistic expectations for a first project:

A clear diagnosis of where time and money leak in one process, quantified.
Two or three changes implemented, with measurable effect on cycle time, rework or compliance.
A baseline that makes every future automation or AI investment easier to justify.
An honest answer to whether a platform is worth it for your company.

FAQ

What is process mining in simple terms?

Process mining is a way to rebuild how a business process really runs by reading the timestamps that IT systems already record. Each event needs a case ID, such as an order number, an activity name and a timestamp. Algorithms then draw every path cases actually followed, count how often each path occurs and measure the waiting time between steps. The result shows the real process, including rework loops, skipped controls and delays, rather than the version on a flowchart.

Should we use process mining before automation, and where do we start?

Yes. Using process mining before automation tells you which paths are stable enough to automate and which need standardizing first. Start with one process that has a clear owner, a measurable pain point and data in one or two systems, such as procure to pay. Measure how many variants cases follow, remove rework loops, automate handoffs and waiting time, and only then automate individual tasks. Keep mining after go-live to prove the result against the baseline.

How long does a first process mining project take?

A focused first project usually takes about 90 days. The first month goes to choosing the process, extracting and cleaning the event log and agreeing on data protection. The second month covers discovery, conformance checking and root cause analysis. The third month is for implementing two or three changes and checking the effect against the baseline. Data extraction is the most common source of delay, especially when the process spans several systems.

What data do we need for process mining?

You need an event log with at least three fields for each event: a case ID that identifies the item moving through the process, an activity name describing what happened and a timestamp. Most ERP, CRM, ticketing and workflow systems store this data in tables or change histories. Extra attributes such as supplier, region, amount or team make the analysis far more useful, because they let you explain why some cases are slow or costly.

How does AI change process mining?

AI makes process mining easier to use and more predictive. Natural language queries let business users ask questions without building filters, machine learning predicts which open cases will be late, and generative AI speeds up event log preparation. In the other direction, process mining helps AI by showing where automation will have the most impact, giving AI agents an accurate map of the real process and checking afterwards whether automated actions followed the rules.

Is process mining worth it for a mid-sized company?

It can be, if you start small. A mid-sized company does not need an enterprise platform on day one. One process, a sample event log and a light tool are enough to find out whether there is value. If the first project quantifies savings in rework, lost discounts or late payments that clearly exceed the cost of a platform, scale up. If it does not, you have still gained a baseline for future automation decisions at low cost.

What is the difference between process mining and task mining?

Process mining reads event data from business systems such as an ERP or ticketing tool, so it sees the steps those systems record. Task mining captures user interactions on desktops, such as clicks, copy and paste and application switches, so it sees work that happens between systems. Task mining fills gaps but raises stronger privacy and employment questions, so most companies run it as a second phase with legal, data protection and employee representatives involved.