Which Version of Your Business Will AI Learn?

The map and the Monday morning

Consider a hypothetical industrial-parts distributor. Its documented order process runs from credit approval to warehouse release to dispatch. On Monday, a replacement part is urgently needed. A supervisor arranges what she understands to be a permitted exception by phone, but the approval is entered into the system after the warehouse has prepared the shipment.

The log appears to show work proceeding before approval. The handbook appears to prohibit it. The supervisor can explain the exception. This invented case has several possible lessons: improve the recording, clarify the rule, or change the workflow. Simply teaching an agent the most common sequence would leave the important question unanswered.

A process map is not a confession, and a repeated shortcut is not automatically company policy. The executive task is to decide what should continue, what should change and what remains uncertain. That is a management judgment in this scenario, not a finding about ARIS customers.

More connections can reveal more gaps

Object-centric process mining relates recorded activity to multiple business objects rather than forcing everything into one case. For our distributor, that could mean considering orders, deliveries and invoices together. A June 10 Process Science paper examines the data-quality problems that arise in this richer representation, including missing or incorrect relationships. Its evaluation uses limited event-log and report evidence plus practitioner interviews. One coauthor is affiliated with SAP Signavio. [2]

The practical implication I draw is to ask what a connection means before treating it as an explanation. Does an approval belong to this order, this delivery or the customer account? Which events were never captured? The paper supports investigating these questions; it does not test my proposed distributor review or establish the new ARIS offering's reliability. [2]

Timing deserves its own question. A January 16 study in SN Computer Science investigates estimating missing timestamps and the trade-off between estimation accuracy and the share of records covered. Its experiments remove timestamps artificially from three historical logs. That is evidence about a specific data-repair problem, not proof that reconstructed histories are equivalent to observed events. [3]

For the fictional Monday order, I would keep a recorded approval time distinct from an estimated one and from a supervisor's recollection. Each may help explain the case. They should not silently become the same kind of evidence. This is my application of the study's measurement problem. [3]

Give the people who know the work a real part

A separate research paper, published December 15, 2025 in Business & Information Systems Engineering, studies an LLM-based system for acquiring process knowledge. It combines domain-expert input with analyst work and validation. Its evaluation includes a small quasi-experiment with junior analysts and an illustrative organizational case. The authors explicitly discuss limitations in capturing unspoken activity, handling ambiguity and generalizing to more complex processes. [4]

That research offers a useful direction without providing a universal recipe. An interview can surface reasoning absent from a log, while a record can expose activity an interview misses. I would bring the order supervisor, a process analyst and the relevant policy owner into the same case review. Their job would be to explain disagreements and decide which ones require a business change. This proposed arrangement applies the research; it was not the intervention tested. [2][4]

The meeting should produce more than a tidier diagram. In our hypothetical distributor, it might reveal that an approved emergency route exists but is poorly recorded. Or it might reveal that people disagree about whether that route is approved at all. Those are different problems, and buying a connection between systems does not resolve the second one.

Buy a clearer operating decision

There is a broader investment reason to do this work. McKinsey's November 5, 2025 global survey associates stronger reported AI value with workflow redesign. It is an external consultancy survey, based on self-reports rather than a randomized intervention. It cannot establish that process context caused better results, and it predates this launch. It supports considering changes to work alongside technology purchases. [5]

My proposed starting point is one consequential decision within one workflow, such as whether an order may move to warehouse release. Compare the written rule, available records and employee explanations. Then make three choices before widening the assignment: [2][4][5]

• What must be repaired? Identify missing relationships, uncertain timestamps or undocumented exceptions that could change the decision. Preserve uncertainty rather than filling it with confidence. [2][3]
• What should the agent follow? Have the responsible business owner resolve policy disagreements and approve the intended process. Observed frequency alone is not authority. This is an editorial application of analyst validation, not a tested governance method. [4]
• What is worth funding? Compare repairing the process, improving its records and adding the proposed AI capability. Include the human effort needed to explain and maintain the context. This is an investment question informed by workflow-redesign evidence, not a savings forecast. [4][5]

Keep the useful disagreement

The ARIS announcement makes process context a timely buying conversation. Its claims remain vendor descriptions, and this article has not tested the product. The leadership opportunity is to use that conversation to expose a specific ambiguity in how the business operates, then decide whether resolving it warrants investment. [1]

For our distributor, the most valuable discovery might be a missing approval record or an outdated rule. Either would be a useful result before granting an agent a larger role. Give AI a map the organization can explain, including the places where the map is still unfinished.

Questions for executives

1. Where do our approved process, execution records and employee explanations disagree?
2. Which uncertainty could change the decision we want to delegate?
3. Who can resolve that disagreement, and what work should we fund first?