Process mining is genuinely powerful and almost entirely irrelevant to a mid-market automation programme, for a simple reason: the licence costs more than the programme.
What the tools actually cost
| Tool | Indicative 2026 pricing |
|---|---|
| Celonis, single domain | $150,000–$250,000 per year list; premium connectors $10,000–$30,000 each per year; implementation $120,000–$500,000 |
| UiPath Process Mining, standalone | $50,000–$200,000 per year |
| Soroco Scout, Skan.ai, ABBYY Timeline | No public pricing; enterprise-only, guided setup required |
| Microsoft Power Automate Premium | ≈$15 per user per month, includes process and task mining; add-on $5,000 per tenant per month beyond pooled capacity |
There is a second problem beyond price. Building the event log — the extraction and transformation work — is reported to consume up to 80% of a process-mining project's effort, large deployments run twelve to twenty-four months, and research has found two in three process-intelligence deployments under-delivering. A boutique promising to mine your ERP inside a two-week audit is promising something that does not exist.
What works instead
- Silent shadowing first. One hour per role, watching without interrupting. Questions afterwards, not during — asking while watching changes what you see.
- Interviews second, and specifically about exceptions: what happened the last time this went wrong, and what did you do?
- Logs third. Ticket systems, CRM, telephony, WhatsApp and email exports carry the volumes and timings that turn anecdote into a baseline.
- An exception map per process, distinguishing the 80–85% that follows a predictable path from the 15–20% that needs judgement.
- A scored opportunity matrix — value against feasibility — so the sequencing argument is on paper rather than in the room.
Where the tool-free approach genuinely loses
Honesty about limits: shadowing does not find the process variant that happens twice a quarter in a regional office, and it does not quantify conformance across ten thousand cases. If you need statistical rigour across a high-volume, high-variance process — order-to-cash across multiple entities, say — real process mining is the right instrument and worth its price.
For everything else, which is most things at this scale, two weeks of watching and log analysis will tell you what to automate first. And it produces something a mining tool cannot: the exception map, which is what decides whether an automation is safe to switch on.
