When process mining cannot see the work
Process mining is a good technology pointed at the wrong half of a regulated operation. The processes with the largest prize are precisely the ones it cannot observe.
How process mining actually works
Process mining reconstructs a process from event logs, the timestamped records that transactional systems produce as work moves through them. Given a case identifier, an activity name and a timestamp, it can rebuild the real path work took, show variants, and quantify where time is lost.
Where a process runs end to end inside one or two systems, this is genuinely excellent. Order-to-cash in a single ERP. A claims workflow in one claims platform. The data is there, it is complete, and the analysis is close to objective.
The structural problem in regulated operations
Now consider the processes that actually cost a fund administrator or a specialist insurer money.
Investor onboarding runs across a transfer agency platform, a document store, three email threads and a counterparty portal. AML periodic refresh is tracked in a spreadsheet one person maintains. Dealing exceptions are worked from a shared mailbox with no queue and no ageing. Client reporting is assembled from four extracts and finished by hand.
None of that produces a usable event log. There is no case identifier that spans the whole journey, no activity names, and no timestamps for the parts that happen in someone's inbox.
Dark work
The term worth having is dark work: activity that consumes real capacity but leaves no trace in any system of record.
- The shared mailbox that three people work without a queue, an ageing report or a record of who resolved what.
- The spreadsheet that tracks a rolling population, the one that would be difficult to explain to a regulator.
- The chase-up call, the workaround that exists because two systems do not talk, the check somebody does because of something that went wrong in 2019.
- The exception path, which is where the cost concentrates and which almost never appears in the documented procedure.
Dark work is where rework, risk and capacity loss cluster. It is also where the AI opportunity is largest, because the work is manual precisely because no system was ever built for it. This is why an AI opportunity assessment that only reads system data will miss most of the prize.
Why the alternatives do not close the gap
| Approach | Where it falls short |
|---|---|
| Process mining | Needs event logs. The highest-cost processes generate none. Long, expensive data-engineering phase before any insight appears. |
| Task mining | Records desktop activity, so it captures screens saturated with client data. Captures the happy path far more reliably than the exceptions that break automations. Tells you what happened, not why. |
| Procedure documents | Describe how work is supposed to be done. The divergence between that and actual practice is usually the finding, not the input. |
| A workshop and a whiteboard | Captures the real path, but without a consistent scoring method, a baseline or an evidence trail, so it cannot be defended when the board asks how you chose. |
What works instead: structured capture with provenance
If the data does not exist, create it deliberately rather than inferring it from a partial signal.
- Facilitated capture with the people who do the work, not the people who own it. Activities, actors, systems, handoffs, controls, volumes, cycle times, exception rates and, critically, where the work waits.
- Validation by a named person for each process, recorded. That signature is what makes the map stand up to scrutiny later.
- Evidence carries provenance. Where a procedure document is used as a starting draft, it is marked as such, with a confidence flag, and where observed practice contradicts it the divergence is recorded rather than smoothed over. That divergence is frequently the most valuable finding in the engagement.
- Structure, not a diagram. The output has to be queryable, scoreable, re-runnable, comparable, rather than a picture. A diagram is a photograph of a moving object. Once it is structured, every step can be scored consistently.
If your highest-cost processes genuinely run end to end inside one or two systems, and you have the data engineering capacity to support it, process mining will beat structured capture on precision and on ongoing cost. We will say so. If you are already midway through an enterprise implementation, a second tool will not help you, and that is one of the cases where we decline the work.
Common questions.
Why does process mining not work for fund administration or insurance operations?
Process mining reconstructs a process from event logs produced by systems of record. It works well where the whole process runs inside one or two transactional systems. In fund administration, insurance operations and specialist banking, the most expensive work runs across email, spreadsheets, shared mailboxes and counterparty portals, which generate no usable event log. The processes with the largest prize are therefore the ones process mining cannot see.
What is dark work?
Dark work is activity that consumes real capacity but leaves no trace in any system of record: the shared mailbox, the spreadsheet that tracks a cycle, the chase-up call, the workaround that exists because two systems do not talk. It is typically where cost, risk and rework concentrate, and it is invisible to any tool that relies on system data.
Can task mining solve this?
Task mining captures desktop activity and closes part of the gap, but it introduces its own problems in regulated firms: it records screens saturated with client data, it captures the happy path more reliably than the exceptions that break automations, and it tells you what happened without telling you why. It is a useful supplement to structured capture, not a replacement for it.
Is process mining a competitor to PinpointProof?
Not usually. A firm with a mature process mining implementation already has a tool, a team and event-log coverage of its core systems. The firms we work with have the opposite problem. Their highest-cost processes never touch a system that produces a usable log. If you are midway through an enterprise mining programme, a second tool will not help you.
Related.
See it on your own process.
Bring one process you would describe as expensive. Thirty minutes, and you will see what its map and scores look like.
Thirty minutes. Bring one process.