How to Choose Process Mining Software

Most process mining evaluations start with a shortlist. It is the natural place to begin, and it is usually a step too early. The tool you can actually use is constrained before the first demo, by the event data your systems already produce and by whether your process has the shape process mining expects.

Answer those two questions first and the shortlist narrows itself. Skip them and the best platform on the market will produce a confident, expensive picture of something that is not quite your process.

This is a guide to the questions that decide the outcome, in the order they matter.

First, the category you are buying in changed this year

In May 2026, Gartner retired its Magic Quadrant for Process Mining Platforms and replaced it with a Magic Quadrant for Process Intelligence Platforms, published on 5 May 2026 by Tushar Srivastava, David Sugden and Marc Kerremans. The scope widened at the same time, to cover process mining and task mining together.

That is not a cosmetic change, and it has two practical consequences for a buyer.

The first is that the shortlist you assemble by searching for process mining software is now a subset of a larger market. Platforms are being evaluated, and priced, against a broader definition than the one most buyer's guides still use.

The second is that task mining is now inside the same category boundary. That matters for cost and for governance, and it is covered further down.

Everything below still applies. The mechanics of process mining have not changed because the label did. But if a vendor tells you they are a Leader, ask which quadrant and which year, because the answer has a different meaning after May 2026 than before it.

Start with your event data, not the shortlist

Every process mining tool needs the same raw material: an event log. The requirement is narrow.

The IEEE Process Mining Manifesto is the founding reference for the discipline and still the definition the tooling is built on. It states the requirement plainly. Process mining techniques assume it is possible to record events sequentially, where each event refers to an activity and is related to a particular case. The IEEE XES standard, which most platforms implement to this day, makes that concrete as three mandatory fields: a case identifier, an activity name, and a timestamp.

That sounds simple. In a real enterprise it is not.

A case identifier has to mean the same thing across every system the process touches. The Manifesto names the problem directly. There is a many-to-many relationship between orders and deliveries, and a one-to-many relationship between orders and order lines. Which of those is the case?

Choose the order and you lose the detail of the deliveries. Choose the delivery and one order looks like several unrelated journeys.

That decision is usually made by whoever builds the first extraction. It quietly determines what every dashboard afterwards can show. Ask a vendor how their tool handles it, and how much data engineering it expects from you to get there. Those two answers tell you more than a feature list.

Know which of the five data levels you are at

The Manifesto grades event data on five maturity levels. They are the most useful evaluation instrument in this field, precisely because they are not about software at all.

At the top, five stars, events are recorded automatically, systematically and reliably. The log is trustworthy and complete, and the recorded events have clear semantics.

At the bottom, one star, recorded events may not correspond to reality. Events may be missing. The data is recorded by hand.

The Manifesto's verdict on that bottom level is the single most useful sentence in the document for a buyer: it does not make much sense to apply process mining to logs at that level.

Most enterprises sit somewhere in between, and the honest work of a selection process is finding out where, system by system, before committing. An ERP that has recorded order events the same way for a decade is a different proposition from a ticketing system that three teams use three different ways.

Guiding Principle 1 of the Manifesto is blunt about the consequence. Event data should be trustworthy, complete, have well defined semantics, and be safe. The quality of a process mining result heavily depends on the input. No platform overrides that.

The three process mining capabilities, and which one you are buying

Vendors use process mining to describe three different things. The academic definitions are precise, and knowing them stops you paying for one and expecting another.

Discovery takes an event log and produces a model without using any prior information. This is what most people picture, and what most demos show.

Conformance checking compares an existing process model against an event log of the same process, to see whether reality conforms to the model and the model to reality. This is what audit and regulated processes actually need.

Enhancement extends or improves an existing model using what the log shows about the real process.

The distinction matters commercially. Conformance checking is often assumed to be included when it is not, and it depends on something discovery does not need: a current, governed process model to check against. If your documented process is three years old, conformance checking measures execution against a fiction.

Process mining and task mining are different purchases

Since May 2026 these sit inside the same analyst category, which makes the distinction more important rather than less.

Process mining reads system event logs and shows how work moves across systems, end to end. Task mining reads desktop activity and shows how one person completes one step inside an application.

Process mining tells you that approvals take four days. Task mining tells you that the approver re-keys the same data into two screens. Both are useful. They are not substitutes, they carry very different privacy implications, and they are frequently priced separately.

Be explicit about which one your business case needs. If a quote covers both because the category now covers both, check whether you are paying for a capability you have no immediate plan to deploy, and whether you could get approval to deploy it at all. The privacy section below explains why that second question is not hypothetical.

What happens when the work is not one case

Classical process mining forces every event into exactly one case. When the work does not fit that shape, three specific distortions appear. Wil van der Aalst named them in 2019.

Deficiency is where events belong to no case at all. Convergence is where one event has to be duplicated across several cases. Divergence is where the same activity repeats inside one case and the model cannot tell the repetitions apart.

This is not an edge case. It is what order-to-cash looks like in most enterprises, where one order becomes several deliveries and one invoice covers several orders.

Object-centric process mining, introduced in the same work, removes the single case assumption. An event can relate to several objects of different types at once. Whether a tool supports it, and how, is a real technical distinction rather than a marketing one. Ask for it by name.

When process mining is the wrong tool

No guide in this category says this, so here it is.

When the work does not happen in a system. Process mining reads what systems record. The Manifesto's one-star level explicitly includes trails left in paper documents and yellow notes. Work that happens in conversation, in an attachment, or in someone's judgement leaves no event to mine. If a meaningful share of your process starts in an inbox rather than in a transaction, process mining will give you accurate timings with unexplained gaps between them, and you will need something that reads the communication layer as process data to close them.

When variability is extreme. The dominant discovery technique builds on directly-follows relationships, and it degrades badly when concurrency is high. The Manifesto illustrates the scale: ten activities that can happen in any order produce 3,628,800 possible sequences. The resulting model is either too simple to be true or too complex to read.

When volume is low. A process that runs forty times a year will not yield a statistically interesting model, whatever the tool costs.

When nobody can act on the finding. Not a technical limit, but it ends more projects than the technical ones. Deloitte's 2025 survey found that 41% of organisations name lack of management support as a barrier to process mining adoption, up from 26% in 2021. If the analysis has no owner with the authority to change the process, you are buying a precise description of a problem you will not fix.

The European questions: what you must prove, and what you may collect

For enterprises operating in the EU these are procurement questions, not afterthoughts, and they pull in opposite directions.

What you must prove. The EU AI Act reaches full applicability on 2 August 2026, and DORA already applies to financial services. Both expect organisations to demonstrate on an ongoing basis that processes execute as approved. Conformance checking is the mechanism that produces that evidence continuously rather than assembling it before an audit. If compliance is part of your business case, this is the capability to pressure-test, and it brings the process model question back with it.

What you may collect. Event logs routinely carry a resource attribute, which is the identity of the person who performed each step. Academic work on privacy in process mining treats employee re-identification from event data as a first-order problem, requiring techniques such as k-anonymity or differential privacy before logs can safely be analysed or shared.

In Germany, works councils hold a statutory co-determination right under Section 87(1) No. 6 of the Works Constitution Act over technical equipment designed to monitor the behaviour or performance of employees. Process mining falls within reach of that provision, and task mining considerably more so. A deployment agreed with IT and not with the works council is not a deployment.

Ask early what a tool can anonymise, at what level, and whether the analysis still works once it has. Some handle this well. Find out before the contract.

What the evidence says about value, and what it does not

The most substantial survey evidence in this market is Deloitte's Global Process Mining Survey, published in February 2025 and still the most recent edition. It surveyed more than 120 senior IT and operations leaders in late 2024. Two findings are worth a buyer's attention. Only 25% of respondents were using AI together with process mining, while 74% planned to. And management support had become the most cited barrier, at 41%, up from 26% in 2021. The technology is no longer the constraint. The organisation is.

Three caveats nobody quoting it tends to mention. The sample is around 120 respondents, small for the confidence with which the percentages get repeated. Deloitte sells process mining consulting, so this is industry research rather than independent research. And the fieldwork predates the category rename entirely.

On cost the position is simpler. There is no credible independent data on what process mining software costs, how long implementation takes, or how large a team it needs. The figures in circulation come from vendors, from the consultancies that sell the implementation, and from review platforms funded by the vendors listed on them.

That does not make them wrong. It means you should treat any “live in eight weeks” claim as a statement from a party with an interest in a fast answer, and ask what it assumes about your data. Most such claims assume event data near the top of the five levels. If yours is not, the timeline is not about the software.

A short readiness test

Before you shortlist anything, answer these six questions.

1. Can you name the case identifier for the process you want to analyse, and does it hold across every system involved?
2. Do you know which of the five maturity levels your event data is at, per system?
3. Is your business case built on discovery, conformance checking or enhancement?
4. Does the process fit one case per instance, or does it fan out into deliveries, lines and invoices?
5. Does a meaningful part of the work happen outside any system?
6. Does someone own the authority to change what the analysis finds?

Answer all six and a vendor conversation will be short and useful. If the first two are unanswered, no shortlist will help you yet, and the work to do is understanding what process intelligence is and where it stops.

Where Tekst fits

This guide has been careful about its sources, so it should be equally clear about who wrote it. Tekst builds in this market. Read the rest of the article on its merits and this section as a position.

Two of the limits above are the reason Tekst exists.

The first is that process mining reads what systems record. Event logs only exist once someone has already typed the work into the system. The hours before that, the clarification email, the revised PDF, the decision taken in an inbox, never reach the log. A process mining platform can show that an order took nine days. It cannot show that four of them went on a question nobody logged.

The second is that discovery produces understanding, not action. Acting on the finding is usually a separate project, with a separate budget and often a separate vendor.

Tekst reads the communication layer as process data and executes in the systems of record, so the map and the action sit in one platform rather than two. Teams weighing that up against a process mining platform can read how Tekst compares as a Celonis alternative. Teams coming at it from the automation side can read how it compares as a UiPath alternative.

None of that changes the advice above. Whatever you buy, and whoever you buy it from, the event log decides what you get.

The short answer

Choosing process mining software is mostly not a software decision. It is a data readiness decision, a process shape decision, and a governance decision. The tool follows from those three.

Evaluate your event logs against the five maturity levels first. Settle the compliance and works council questions before procurement rather than during it. And treat every timeline you are quoted as a claim about your data rather than about the product.

The tools in this market are, broadly, good. Most disappointing projects are not caused by picking the wrong one.