The most expensive AI project is the one that targets the wrong workflow. You spend three months building a model, provisioning infrastructure and training the team — then discover the process you automated handles 80 cases a month, not 800. The economics never worked. The project is quietly shelved, and the next budget round treats "AI" as a word that burns money.
This happens more often than anyone admits, and the root cause is always the same. The company started with the solution — "let us use AI" — instead of the problem: which workflows actually have the characteristics that make AI worth the spend? Vendor pilots, board-level FOMO and a loud competitor announcement all pull in the same direction, toward building before anyone has looked at the evidence.
Process mining is the discipline that supplies that evidence. As the academic and vendor literature defines it, process mining sits at the bridge between data science and process science: it reconstructs how a process actually runs from the digital footprint it leaves behind — event logs, transaction records, system timestamps — rather than from how managers believe it runs. The classic three tasks are discovery (distil a process model from the logs), conformance checking (compare what the data shows against the model you assumed) and enhancement (enrich the model with performance data such as cycle time, waiting time and resource utilisation). For our purposes there is a fourth, commercial question layered on top: of all the workflows this analysis surfaces, which ones have the volume, the pattern density and the measurability to justify an AI investment at Mittelstand scale?
Why intuition fails
Most AI candidate selection happens in a meeting room, not in the data. The Geschäftsführer suggests customer service. The CFO suggests invoice processing. The Head of Operations suggests quality control. Each has a defensible reason. None has a number.
The problem with intuition is that it gravitates toward the visible pain rather than the addressable one. Customer service feels painful because complaints are loud and they reach the top. But that process may be low-volume, highly variable — every ticket genuinely different — and poorly instrumented, with no agreed baseline for resolution time or first-contact accuracy. Loud and painful, yes. A strong AI candidate, no.
Meanwhile the accounts-payable team works through several thousand invoices a month against a largely fixed set of validation rules, with a stable error rate and a cost per error that finance can already put a figure on. Nobody raised it in the meeting because it does not feel urgent — it just quietly consumes two FTEs. On the three filters that actually matter, it is the better candidate by a wide margin. Process mining exists precisely to make that comparison visible before the budget is committed.
The three-filter framework
A workflow has to clear three filters before it earns an AI investment. Each is necessary; none is sufficient on its own. Skip any one of them and you are back to building on instinct.
Filter one: volume. The workflow must move enough transactions to amortise the cost of building, deploying and maintaining the system — and that maintenance cost is the part executives consistently underestimate, because an AI workflow is not a project that ends but a system that has to be monitored, retrained and governed for as long as it runs. For most Mittelstand operations the practical floor sits somewhere around several hundred transactions a month, or roughly one full-time-equivalent of human effort dedicated to the task. The exact number depends on your labour cost and your error cost, but the logic is unforgiving: a model that saves ten minutes per case across fifty cases a month saves under nine hours of work — a few hundred euros of labour that will never cover a production AI system's true total cost of ownership. Volume is also a question of trajectory, not just today's state. A process running below the threshold but growing steadily can be worth preparing for, so that the capability is ready the quarter it crosses the line rather than two quarters after.
Filter two: pattern density. AI earns its keep on repetition. A workflow where most cases fall into a manageable set of recognisable, repeatable shapes is a strong candidate; one where every case is genuinely novel is not — there is no pattern for the model to learn, and you will spend more on edge cases than you ever save on the mainstream. You can assess this without building anything. Pull a sample of recent cases, classify them by hand, and measure the concentration: what share falls into the top few types? When the bulk of volume clusters into a handful of categories, density is high. When the long tail dominates, it is low. Insurance claims triage is the textbook high-density case — water damage, storm, theft and motor account for most of the volume, each following a predictable assessment path — whereas bespoke advisory or strategic-consulting requests sit at the opposite end, where every engagement is its own animal.
Filter three: measurability. You need three things you can point to: a clear definition of a correct output, a reliable baseline of current performance, and a method for measuring both on an ongoing basis after go-live. Without a definition of correct, you cannot train or evaluate anything. Without a baseline, you cannot prove improvement to the board. Without continuous measurement, you cannot catch the slow degradation that creeps into every deployed model as the world drifts away from its training data. The most common gap we see is an organisation that knows its throughput — cases per week — but has never measured its accuracy, the share of cases handled correctly on the first attempt. That gap is dangerous, because a system that processes faster while quietly making more mistakes manufactures rework, not savings. Building those baselines before you build the model is the subject of Measuring Operational AI Impact.
Doing the work without a platform
You do not need a six-figure process-mining licence to run this exercise. A platform like Celonis pays for itself once you are continuously mining dozens of processes across the enterprise; for the narrower question of which two or three workflows deserve an AI investment this year, structured analysis of your own system exports will get you most of the way.
Start by mapping the high-volume workflows — every recurring process that crosses more than one person or system — and estimating monthly volume and FTE involvement for each. An afternoon in a spreadsheet, pulled from ERP and ticketing exports, is enough to rank candidates by raw scale. Then score pattern density on the top handful: sample fifty to a hundred real cases, classify them, and calculate the concentration ratio. Next, check measurability: for everything that survives the first two filters, ask whether you can define a correct output, whether you already know your error rate, cycle time and cost per transaction, and whether you can keep measuring them once a model is in the loop. Finally, rank and select. The workflows that score across all three dimensions are your real candidates, and in a typical mid-market company with a few hundred employees that honest list is short — usually a handful, not a portfolio. A short list is a feature. It concentrates scarce engineering and change-management effort where it compounds.
Where the filters meet the regulator
For DACH operators there is a fourth lens that overlays all three, and it is hardening fast. Under the EU AI Act, systems used for purposes such as recruitment and worker management, or creditworthiness and credit scoring, are classified as high-risk under Annex III and carry obligations for risk management, data governance, logging, human oversight and conformity assessment. That changes the economics of an AI candidate: a high-density, high-volume CV-screening or credit-decision workflow may look perfect on the three filters and still be the wrong place to start, because the compliance burden lands squarely on you as the deployer.
The timing is in motion. The obligations for standalone Annex III systems were originally due to apply from 2 August 2026, but under the Digital Omnibus agreement reached in May 2026 they are set to be postponed to 2 December 2027 — a change that only takes legal effect once formally adopted and published in the Official Journal. The practical reading for a Mittelstand board: you have more runway than the original deadline implied, but the destination has not moved. Treat regulatory exposure as a fourth filter, and where it is high, prefer a lower-risk first workflow that builds the same operational muscle without the legal weight.
The three mistakes that survive the framework
Even teams that run the filters honestly fall into the same three traps. The first is chasing complexity: complexity does not correlate with AI value, and the highest-return candidate is usually a simple, high-volume process rather than the intricate one executives find intellectually flattering. The second is ignoring the human factor: a workflow can score perfectly and still fail because the team that owns it is resistant, the process owner controls no budget, or the data is locked in a system no one will grant access to. Operational readiness is as decisive as technical fit, which is why the Automation vs. Augmentation decision belongs in the selection conversation, not after it. The third is buying the demo: vendors stage their technology on ideal data you do not have. Judge every candidate against your volumes, your patterns and your constraints — not against a polished case that was never your process.
From candidate to first workflow
Once the top candidate is clear, the next step is deliberately not building. It is validation. Can you actually access the underlying data, cleanly and lawfully? Does the process owner genuinely back the initiative, or merely tolerate it? Are the compliance requirements understood rather than assumed? Most stalled AI projects fail one of these tests, and they fail it after the money is spent, not before.
This is what Discovery is built for: a two-week engagement that takes one promising candidate and pressure-tests whether it can become a production workflow — technically, operationally and organisationally — before a line of model code is written.
A Fit Call pressure-tests your top AI candidate against volume, pattern density, measurability and regulatory exposure — before you commit a quarter of engineering to the wrong workflow.
For a self-guided pass, our AI Operating Diagnostic walks you through the framework in about ten minutes.
References: Celonis, "What is Process Mining? Definition and benefits," https://www.celonis.com/insights/topics/what-is-process-mining; European Commission, "Annex III: High-Risk AI Systems," EU Artificial Intelligence Act, https://artificialintelligenceact.eu/annex/3/; Gibson Dunn, "EU AI Act Omnibus Agreement — Postponed High-Risk Deadlines and Other Key Changes," 2026, https://www.gibsondunn.com/eu-ai-act-omnibus-agreement-postponed-high-risk-deadlines-and-other-key-changes/.