Process intelligence, process mining, task mining or process documentation: which do you need?
Match the approach to the evidence you hold, and choose a connected model when the work crosses several systems or has to carry governance and cost. Process mining needs event logs with a case id, activity and timestamp; task mining needs recordings of desktop work; documentation needs people to write down what they remember; a connected model, the approach this article means by process intelligence, reads documents, transcripts and recordings, plus logs where they exist.
Written for operations, transformation and technology leaders choosing a process tool or approach for the first time, who keep hearing the four terms used as if they were one. Reviewed by Mike Bentley, founder, .
What the four terms mean, and what each needs as input
Process mining is the analysis of event logs an organisation's systems already record, where each event carries a case identifier, an activity and a timestamp. That is the minimum processmining.org, maintained by the Process and Data Science Group at RWTH Aachen University, sets out: process mining "assumes the existence of an event log where each event refers to a case, an activity, and a point in time". The IEEE Task Force on Process Mining's 2011 manifesto puts it this way: "the idea of process mining is to discover, monitor and improve real processes (i.e., not assumed processes) by extracting knowledge from event logs readily available in today's (information) systems".
Task mining is the recording and analysis of desktop interaction data: clicks, keystrokes, window switches and screen content captured from a person's workstation, analysed to find repeated routines. Researchers call the method robotic process mining, and Leno and colleagues (Business & Information Systems Engineering, 2021) define its input, a user interaction log, as "a timestamped sequence of events performed by a single user in a single workstation".
Process documentation is a description of how work is meant to happen, produced by people: a workshop, an interview, someone typing a procedure, or someone prompting an AI assistant to draft one. Its input is what people say and remember.
Process intelligence is a connected model of how work actually happens, built from evidence an organisation already holds: documents, meeting transcripts, wiki pages and recordings of the work being done, plus system logs where they exist. The term has no agreed definition. The 2011 manifesto listed "(business) process intelligence" among the terms different organisations use for overlapping concepts, giving Software AG's use of "Process Intelligence" for its platform as its example, and its glossary defined process intelligence only as "a branch of Business Intelligence focusing on Business Process Management", so ask which inputs a product reads before accepting the label.
Which approach fits the data you actually have?
Start from the evidence you hold and the question you are asking, not from the term a vendor used. Five questions settle most cases.
If you have the log, the work stays inside one system and the volume is high, mine the log and stop. If the gap is in what people do at their desks, task mining. If the work crosses systems, if a change has to be proved before anyone builds, or if the model must carry governance and cost, a connected model is what the evidence supports.
- Do you have an event log at all: a case identifier, an activity and a timestamp recorded for each step of this process? If not, no tool can mine what was never recorded, whatever the brochure says.
- Is the work in one system, or across people, documents, email and several systems? One instrumented system suits mining; across systems each log holds a fragment, and the whole path needs an approach that reads documents, transcripts and recordings.
- Is your question about what people do at their desks? If the gap is the routine around a step (copying between screens, re-keying, checking a spreadsheet), task mining is built for that, with the consent work it implies.
- Do you need to prove a change before you commit? A picture of the past cannot; you need a model you can alter and check for coverage before anyone builds.
- Does the model need to carry governance and cost? If it must say who is accountable, which AI systems sit inside the process and what it costs to run, it needs more than activity sequences, which points to a connected model.
What each approach misses
Event-log process mining misses everything that never produced a log entry: the phone call that resolved the exception, the spreadsheet where the decision was really made, the judgement a person applied. It also struggles across systems: the manifesto's first challenge, finding, merging and cleaning event data, notes that data "may be distributed over a variety of sources" and that merging "tends to be problematic when different identifiers are used in the different data sources".
Task mining misses the thread that ties actions to a case. Leno and colleagues observe that "UI logs do not come with a notion of case identifier", so a recording shows what one person did at one workstation, not which claim or order it served. It also records much that is nobody's business, so consent and scope matter more here than anywhere else.
Process documentation misses the gap between how work is meant to happen and how it happens. Written from memory and intent, it is rarely checked against evidence and ages from the day it is published. A documented process is a claim; the other three approaches test it.
A connected model misses precise frequencies and durations wherever no log feeds it. A recording of a job done once shows the path, not the distribution of paths, and documents describe what was written down, not what was skipped. A connected model is only as good as the evidence read into it, and a serious one says where each element came from.
When is event-log process mining the right answer?
Event-log process mining is the right answer when a high-volume process runs inside one well-instrumented system: order handling in an ERP, tickets in a service desk, claims in a single claims platform. There the log holds the case identifier, activity and timestamp for every step, the volume makes frequencies and durations meaningful, and the manifesto's three types of process mining (discovery, conformance checking and enhancement) all apply. Nothing else shows variants, bottlenecks and rework in a mature transactional system as directly.
Event-log process mining is the wrong answer, or at least an incomplete one, when the process lives in email, documents and meetings; when it crosses several systems that identify the same case differently; when volume is low and judgement is high; or when there is no log at all. The manifesto is candid that "it still takes considerable efforts to extract event data suitable for process mining", and that effort is spent before any insight arrives.
A useful test: ask the system owner for a table with one row per step, holding a case id, an activity and a timestamp, and see whether they can produce it this week. If they can, mine it. If the honest answer involves three systems, a data engineer and months of stitching, start with the evidence you already have instead.
How the four approaches combine
The approaches are not rivals, and most organisations end up using more than one. Process mining is the sharpest instrument for the instrumented core of a process. Task mining explains what happens at the desk around a single step. Documentation records what the organisation intends, and is often what an auditor asks for first. A connected model is where the three meet, because it can take logs as one input alongside documents, transcripts and recordings, and hold the result in a form that can be queried, changed and checked.
A claims process shows the shape. The claims platform's log reveals that one variant takes far longer than the others. Task mining at two desks shows the delay is a manual lookup in a legacy system. The documented procedure says the lookup should not be needed. A connected model holds all three alongside the accountable owner, the systems involved and the cost of the step, so the proposed fix can be modelled and its coverage checked before anyone changes a system.
The order in which you adopt them matters less than the discipline of keeping every element tied to its evidence: which log, which recording, which document, and which person said so.
Where a connected model fits, and one example of the approach
A connected model is the right starting point when the work crosses systems, when little of it is logged, when a change has to be proved before it is funded, or when governance and cost have to sit in the same picture as the process. It builds from evidence an organisation already holds (documents, meeting transcripts, wiki pages) and from watching the work done, and it takes in event logs where they exist rather than depending on them.
Omni by Valstra is one product built on this approach. It assembles a connected model from documents, meeting transcripts and wiki pages, and from a consented browser recording of someone doing the job once through its Process Recorder, rather than from system event logs. The product pages linked below describe what is built today, including a direct comparison with an event-log process mining platform.
Whichever approach you choose, keep the decision and the reasons for it. The evidence you hold will change as systems are replaced and logs appear, and the approach that fits a year from now may not be the one that fits today.
Sources
- IEEE Task Force on Process Mining, Process Mining Manifesto (PDF)
- Process and Data Science Group, RWTH Aachen University, processmining.org: Event data
- Leno, Polyvyanyy, Dumas, La Rosa and Maggi, Robotic Process Mining: Vision and Challenges, Business & Information Systems Engineering (2021)
- Object Management Group, Business Process Model and Notation (BPMN) 2.0
Common questions
- Is process intelligence just a new name for process mining?
- Not quite, though the terms have overlapped from the start: the IEEE Task Force's 2011 manifesto listed "(business) process intelligence" among the terms different organisations use for overlapping concepts. Today some products carrying the label are event-log process mining platforms and others are connected models that also read documents, transcripts and recordings, so the useful question is which inputs a given product actually reads.
- Can we start with documentation and add mining later?
- Yes, and it is a common path. Write the documentation as a set of steps with owners and systems rather than as prose, so that when logs become available the documented process can be checked against them instead of being rewritten from scratch.
- Do we need notation such as BPMN to begin?
- No. Business Process Model and Notation is an Object Management Group standard (version 2.0 was published in December 2010) and is useful once a design has to be shared precisely or handed to engineers, but a plain list of steps, owners and systems is enough to start, and notation can be applied later.
- What if our processes change every month?
- Then any snapshot ages quickly, and the IEEE Task Force's 2011 Process Mining Manifesto lists dealing with concept drift (its challenge C4: the process changing while it is being analysed) among the open challenges for mining too. Favour an approach that can re-read its evidence, such as fresh logs or a new recording, over a document someone has to remember to edit, and record when each element was last confirmed.
See it against your own processes
A demo runs on your material, not a canned dataset. Bring a process you find hard to explain.