AI process mining: your ERP already knows where you lose money

By Julien Theys
Ask an operations team to draw their order-to-cash process and you will get five boxes: order received, confirmed, shipped, invoiced, paid. Clean arrows between them.
Now follow one real order. It waits two days on a credit block, for a customer who has paid on time for years. Someone emails the customer about a missing delivery detail. The delivery date moves twice. The truck leaves half empty, and the invoice goes out two weeks after the goods.
Nobody sees that whole journey, because nobody touches all of it. Your ERP does. It has recorded every step, with a timestamp, for years. AI process mining reads that record, shows how your processes really run and puts a figure on every detour. It also gives AI agents what they need to fix those detours reliably.
Why operations leaders should care
Most companies are under pressure to "do something with AI" in operations. Many pilots stall for the same reason: nobody can say where the value sits, and the AI has no idea what normal looks like. In our experience, these pilots lack a clear picture of the process they are meant to improve.
Meanwhile, the cost of process friction hides in plain sight. Blocked and incomplete orders, manual touches, late deliveries, premium freight, excess stock and late invoices each look small. Across thousands of orders, they show up in service levels, working capital and margin.
What process mining does, in plain words
Every action in an ERP leaves a trace: which order it concerns, what happened and when. Process mining collects those event logs and rebuilds the real path of every case. The IEEE Task Force on Process Mining describes the idea in its Process Mining Manifesto: extract knowledge from the event logs that information systems already hold, then use it to discover, monitor and improve processes.
Within a few weeks, that gives you what no workshop can:
The real process map, with every variant and loop, next to the one in your procedures.
Where time goes: how long each step and each wait actually takes, by site, customer or product.
Where people compensate by hand: re-keyed orders, changed dates, manually released blocks.
The first map is usually humbling. Teams call it the spaghetti diagram, and it is the most useful slide in the room, because every strand is a real case.
One order, followed end to end
Here is an illustrative order, built from patterns that show up again and again:
Day 0: the order arrives by email and is keyed in by hand.
Day 1: a credit block stops it. A colleague releases it on day 3, for a customer with a clean payment history.
Day 4: a delivery detail is missing. Two emails later, the order is complete.
Days 5 to 9: the confirmed delivery date changes twice.
Day 12: the goods leave on a partly filled truck.
Day 26: the invoice goes out, two weeks after shipment, because a billing block sat unnoticed.
On its own, this is an anecdote. Process mining tells you how many orders follow the same path, which customers, products or plants they have in common, and what the delay costs in days of cash and hours of work. That is the difference between a complaint and a business case.
What AI adds to process mining
Process mining shows what happened. AI builds on it in three ways.
Explain. Machine learning finds which combinations of customer, product, site or user drive the detours, across thousands of cases at once.
Predict. Models flag the open orders that are likely to be blocked, late or incomplete before it happens, much as we forecast lead times for a pharmaceutical company's inventory.
Act. AI agents handle the exceptions. They propose the fix, such as releasing a block, correcting master data or rebooking a carrier, and write it back into the ERP.
That last step needs discipline. In the setups we favor, people approve each agent action at first. Once an agent has proven accurate on a given type of case, it handles those cases on its own, and every decision stays logged and open to review. A central orchestrator routes the work to small agents that each do one well-defined job, which keeps running costs down and behavior predictable.
Process context is what turns an AI agent from a clever assistant into a dependable colleague. Without it, the agent cannot tell normal from abnormal.
It pays off at scale. Within the Artefact group, an open order control tower built this way released $30M of working capital at a global manufacturer. Our operations page explains how we apply the approach and what it has delivered.
Why start from process data, not from a use case
The value comes first. The event log puts a number on each problem before anyone builds anything, so priorities rest on evidence rather than opinion.
Agents get guardrails. An agent that knows the normal path can spot what deviates from it, and explain why.
You start on the ERP you have. The approach works on SAP ECC and S/4HANA, Oracle, Microsoft Dynamics and Infor. Companies on older systems do not need to wait for a migration, and the process standards they set now make that migration safer.
Three questions to ask before your next AI project in operations
Can we see how this process runs today, case by case, and not only how it was designed?
Do we know what each deviation costs, in days, euros or manual touches?
When an AI agent acts, what will it compare against to know something is wrong, and who approves its actions until it has earned trust?
If the honest answer to any of these is "not really", start with the process data.
How to get started
We usually start with one process, such as order-to-cash or procure-to-pay. In a few weeks, we connect to your ERP, map how the process really runs and estimate the value of each fix. You leave with a prioritized list of opportunities and a clear view of where AI agents can take over routine work.
Book a call to see what your own data shows. Your ERP has been keeping score for years; it is worth reading.