The most expensive AI project is the one that targets the wrong workflow. You spend three months building a model, deploying infrastructure, training the team — and discover that the process you automated handles 80 cases a month, not 800. The economics never worked. The project is quietly shelved.

This happens more often than anyone admits. And the root cause is always the same: the company started with the solution ("let's use AI") instead of the problem ("which workflows have the characteristics that make AI valuable?").

Process mining answers the second question. It is the discipline of analysing operational data — event logs, transaction records, system timestamps — to understand how workflows actually function, where the bottlenecks are, and which processes have the volume, pattern density, and measurability to justify AI investment.

Why intuition fails

Most AI candidate selection happens in a meeting room. The CEO suggests customer service. The CFO suggests invoice processing. The COO suggests quality control. Each has a valid reason. None has data.

The problem with intuition is that it gravitates toward visible pain points rather than addressable ones. Customer service feels painful because complaints are loud. But the process might be low-volume (200 tickets a month), highly variable (every ticket is different), and poorly measured (no baseline for resolution time or accuracy). That is a terrible AI candidate.

Meanwhile, the accounts payable team processes 3,000 invoices monthly against a fixed set of validation rules, with a known error rate of 3.8% and a clear cost per error. Nobody in the meeting room mentioned it because it does not feel urgent. But it is an excellent AI candidate.

Process mining replaces intuition with evidence.

The three-filter framework

We evaluate AI candidates through three filters. A workflow must pass all three to justify investment.

Filter 1: Volume threshold

The workflow must handle enough transactions to justify the implementation cost and operational overhead. For most Mittelstand companies, the practical threshold is around 500+ transactions per month, or the equivalent of at least one FTE dedicated to the process.

Below this threshold, the math rarely works. A model that saves 10 minutes per case but only processes 50 cases monthly saves 8 hours — roughly EUR 400 in labour. That will not cover the cost of building, deploying, and maintaining an AI workflow.

Volume is not just about current state. Consider growth trajectory. A process handling 300 cases today but growing 15% quarterly will cross the threshold within two quarters — and having AI infrastructure ready when it does creates a strategic advantage.

Filter 2: Pattern density

AI learns from patterns. A workflow with high pattern density — where 60-80% of cases follow recognisable, repeatable structures — is a strong candidate. A workflow where every case is novel is a poor one.

How to assess pattern density without building a model: take a sample of 100 recent cases and classify them manually. How many distinct types emerge? What percentage of cases fall into the top five types? If 70%+ of cases cluster into fewer than 10 types, pattern density is high. If every case requires bespoke handling, pattern density is low.

Insurance claims triage typically shows high pattern density — water damage, windstorm, theft, and vehicle accidents account for the vast majority of claims, and each type follows a predictable assessment path. Strategic consulting requests show low pattern density — each engagement is unique.

Filter 3: Measurability

You need three measurements: a clear definition of correct output, a reliable baseline of current performance, and a method for ongoing measurement after deployment.

Without a definition of correct output, you cannot train or evaluate a model. Without a baseline, you cannot calculate improvement. Without ongoing measurement, you cannot detect degradation.

The most common measurability gap we encounter: companies know their throughput (cases per week) but not their accuracy (percentage processed correctly on the first attempt). This matters because an AI system that processes cases faster but with more errors creates more work, not less. For a structured approach to building baselines and tracking impact, see Measuring Operational AI Impact.

Practical process mining: what to do

You do not need a process mining platform to identify AI candidates. You need structured data analysis.

Step 1: Map your high-volume workflows. List every recurring process that involves more than one person or system. For each, estimate monthly transaction volume and FTE involvement. This can be done in a spreadsheet in an afternoon.

Step 2: Score pattern density. For the top 10 workflows by volume, pull a sample of 50-100 cases. Classify them manually into types. Calculate the concentration ratio — what percentage of cases fall into the top five types?

Step 3: Check measurability. For workflows that pass the first two filters, ask: Can we define what a correct output looks like? Do we know our current error rate, cycle time, and cost per transaction? Can we continue measuring these after deployment?

Step 4: Rank and select. The workflows that score highest across all three dimensions are your AI candidates. In our experience, a typical Mittelstand company with 500+ employees will identify 3-5 strong candidates.

Common mistakes in candidate selection

Selecting the most complex process. Complexity does not correlate with AI value. Often, the highest-value candidate is a simple, high-volume process — not the complex one that executives find intellectually interesting.

Ignoring the human factor. A workflow might score well on volume, pattern density, and measurability — but the team that owns it is resistant to AI, the process owner has no budget authority, or compliance requirements make deployment impractical. Operational readiness matters as much as technical fit. For guidance on navigating the human dimension, see Automation vs. Augmentation.

Choosing based on vendor demos. AI vendors demonstrate their technology on ideal use cases. Your processes are not ideal use cases. Evaluate candidates based on your data, your volumes, and your constraints — not on what looked impressive in a sales demo.

From candidate to first workflow

Once you have identified your top candidate, the next step is not building — it is validation. Can you actually access the data? Does the process owner support the initiative? Are compliance requirements clear?

This is what Discovery is designed for: a 2-week engagement that takes a promising AI candidate and validates whether it can become a production workflow — technically, operationally, and organisationally.

For a self-guided assessment of your AI candidates, our AI Operating Diagnostic walks you through the three-filter framework in about 10 minutes.

If you have already identified strong candidates and want to validate them with our team, book a Fit Call. We will give you an honest assessment of whether your candidates are ready for implementation or need further development.

This article is part of the AI in Operations series. For the full operational AI methodology, see The AI Operating System by Andreas Anding.