Process mining for SMEs: see how your processes really run

The process in the manual is rarely the process that actually runs.

On paper, an invoice moves from receipt to checking, then to approval, then to payment. In practice it sits in an inbox for three days, goes back because a purchase order number is missing, and gets approved twice because nobody was sure whose job it was. Nobody saw the whole thing. Everyone only knows their own part.

Process mining makes that real flow visible, using data your systems already store. For a small business it is less a major project than a question of order: see first, then improve, and only then automate.

What process mining is

Process mining analyses the traces a case leaves behind in your systems. Every time an order is created in your ERP, a delivery is booked or an invoice is paid, a record with a timestamp appears.

The core concepts were summarised by the IEEE Task Force on Process Mining in its Process Mining Manifesto (van der Aalst et al., 2012). The starting point is an event log. Each event refers to a case and an activity, and the events are ordered. Further attributes, such as a timestamp or the resource that did the work, may be added.

In practice, three columns are enough to begin with: a case number, an activity and a timestamp.

Three questions it answers

The manifesto distinguishes three types of process mining:

• Discovery: how does the process really run? The data produces a process map that also shows the unplanned variants.
• Conformance: where does practice depart from the rules? Invoices paid without approval, for example.
• Enhancement: where is time lost? The map is extended with waiting times and frequencies.

Our view: for a small business the first question is usually the most valuable. A worked-out target process rarely exists. For conformance, though, a single rule is already enough, such as "invoices above CHF 5,000 need approval".

The honest question first: is your data good enough?

This is where you find out whether process mining will work for you at all. The manifesto grades event data on five maturity levels, from one star to five.

• Four and five stars: events are recorded systematically, completely and reliably. The manifesto's example is the logs of traditional workflow systems.
• Three stars: events are recorded automatically and match reality, but are not necessarily complete. Examples are tables in ERP systems and event logs of CRM systems.
• Two stars: events are recorded automatically but incompletely, because people can bypass the system. Examples are document and product management systems and service engineers' worksheets.
• One star: traces recorded by hand that may not match reality, such as sticky notes on paper documents or paper-based medical records.

The manifesto's conclusion is clear. Process mining can be applied from three stars upwards. At one or two stars, the analysis is typically problematic and the results are not trustworthy.

Our view for Swiss SMEs: where a process runs through an ERP, accounting software or a ticketing system, the data often sits at three stars. Where it runs through email, spreadsheets and quick word at someone's desk, process mining finds little. That is no reason to give up. It is a reason to start with the right process.

Where it pays off for a small business

Good candidates share three traits: they repeat often, they pass through several people or systems, and they leave timestamps.

• Order to cash: order, check, delivery, invoice, payment received. You see where orders stall and how long it takes until the money arrives.
• Incoming invoices: capture, coding, approval, payment. Loops show up here, such as invoices sent back more than once.
• Enquiry to quote: entry in the CRM, qualification, quote, follow-up. You see how many enquiries sit untouched for days.
• Support tickets: opened, forwarded, resolved, reopened. Frequent hand-offs point to unclear ownership.

What these processes have in common: the bottleneck is often not where the people involved expect it. Everyone sees their own working time. Nobody sees the waiting time between two stations.

Understand, then improve, then automate

Process mining is often lumped together with process automation. The two do different jobs:

• Process mining shows how a process runs.
• Process improvement decides how it should run.
• Automation puts suitable steps into software.

The order matters. Automate a clumsy process and you end up with a clumsy process that runs faster. Some bottlenecks need no technology at all, just a clear owner or one approval step fewer.

Only once the better process is settled does it make sense to ask which steps a machine should take over. Which ones typically qualify is covered in 12 processes you can automate today. And if the data lives in two systems that don't talk to each other, the work starts with connecting ERP and CRM.

The Swiss angle: analyse processes, not people

An event log often records who carried out a step as well. That quickly turns process mining into a question of employment law.

Three provisions matter here:

• Art. 26(1) of Ordinance 3 to the Employment Act (ArGV 3): monitoring and control systems intended to monitor employees' behaviour at the workplace must not be used. Where such systems are needed for other reasons, para. 2 requires that they be designed and arranged in particular so that employees' health and freedom of movement are not impaired.
• Art. 328b of the Code of Obligations: an employer may only process data about employees insofar as it concerns their suitability for the job or is needed to perform the employment contract.
• Art. 6 of the Data Protection Act (FADP): personal data may only be collected for a specific purpose that the person concerned can recognise, processing must be proportionate, and data no longer needed must be destroyed or anonymised.

In practice this gives a clear line: analyse the case, not the person. Specifically:

• Drop the column with names, or replace it with a role such as "accounts" or "sales". If only one person holds a role, they remain identifiable; then drop the column altogether.
• Write down the purpose before you start and tell the team.
• Delete the export once the analysis is done.

If the log contains customer names, the same applies: leave out what the analysis does not need.

Using process mining to compare individual employees, on the other hand, comes close to the prohibition in Art. 26 ArGV 3. Where exactly the line falls in a given case is a question for a specialist before you begin. More on data protection in AI and data analysis is in AI and data protection in Switzerland.

How to start with one process

A first attempt needs no budget for a platform. It needs a question and an export.

1. One question, one process. Not "where are we inefficient?" but "why does it take so long from order to invoice?".
2. An export with three columns. Case number, activity, timestamp, as a CSV from your ERP, accounting software or ticketing system. If the system also records when an activity started, add that as a fourth column. Without names, see above.
3. A first analysis. If you know some Python, the open-source library PM4Py (licensed under AGPL-3.0) is enough for a test. Without programming skills, a commercial tool's trial version is the easier route.
4. A conversation. Show the process map to the people who work in the process. The data shows where time is lost. Why, the people involved usually know.
5. One change, then measure again. With the same export, a few weeks later.

If the data is spread across several programs, the export is the real effort. How systems exchange data is explained in the article on APIs for SMEs.

Which metrics count

Four KPIs are enough to begin with:

• Lead time: how long does a case take from start to finish?
• Processing time: how long is someone actually working on it?
• Rework rate: how often does a case go back a step?
• Number of variants: how many different paths does the same process take?

The most telling comparison is between the first two. For that, though, the export needs two points in time per activity, start and completion. With a single timestamp you can measure lead time, but not processing time.

If lead time is a multiple of processing time, the problem is waiting, not working. Faster software won't help then. A shorter hand-off will.

No metric works on its own. A shorter processing time is worth nothing if rework goes up as a result.

What process mining cannot do

Three limits worth knowing before you start:

• It only sees what gets recorded. A phone call, a walk to the next office, a quick agreement at the coffee machine are missing from the log.
• It shows where, not why. A bottleneck in the data is a pointer, not a diagnosis.
• It improves nothing by itself. Without someone who owns the change, the process map stays a nice picture.

AI changes little here. Once the process is clear, it can sort anomalies or take over follow-up tasks. What it looks like when a system acts on its own is shown in AI agents for business.

What ALPENIQ does here

ALPENIQ AI analyses processes, improves them and then automates what can be automated, with integrations between CRM, ERP and Microsoft 365. Process analysis is the first step, not an add-on. The full scope is on the process automation page.

We do not run a process mining platform of our own. Where large data volumes need to be analysed continuously, a specialised tool is the better choice.

Conclusion

Process mining answers a simple question with data: how does this process really run? For a small business it pays off where a process runs through systems with reliable timestamps. The way in is a single process, an export with three columns and a conversation with the people involved. Analyse the case and not the person, and you also stay on the line Swiss law draws.

Further reading

• How to choose and measure your first AI task: AI strategy for SMEs
• What Microsoft 365 already includes for automation: Power Automate for SMEs
• Terms without the jargon: API in the glossary

Frequently asked questions (FAQ)