Connecting an AI model to enterprise data isn't always enough to produce genuinely actionable recommendations. At the Tech For Industry Show, Geoffrey Gabelle, Partner Director and Antoine Cavelius, Lead Applied Value Engineer(Celonis), a German company that has specialized in process mining for 15 years, presented their vision of an operational context model, designed as the missing link between raw data and agentic AI that's genuinely relevant for industry.

Celonis was historically built around process mining: reconstructing a company's actual processes from the logs of its IT systems, without needing to interview business experts. Over 15 years and a broad customer base, the company evolved this approach toward process intelligence — no longer just understanding what happened, but acting on processes and orchestrating differently to capture the identified value. This year, Celonis now positions itself as an operational context model, a layer intended both for process performance analysis and for feeding agentic AI development.

The speakers illustrate their point with a simple example: asking an internal LLM what to do in response to a rise in product demand produces an acceptable but generic answer (replenish stock, possibly increase safety stock). By connecting that same model to precise, event-based enterprise information, the answer becomes far more actionable: identifying that the demand increase concerns a finished product using a component currently at critical stock level, knowing that this component is theoretically delivered in 10 days by its usual supplier but is actually delivered in 15 to 20 days based on observed history, and consequently suggesting express shipping to avoid a stockout. This ability to bring AI as close as possible to a real decision point — with the necessary controls to avoid hallucinations — is what Celonis considers decisive for the next wave of enterprise AI development.

Celonis's operational context model rests on several building blocks:

These are events that happened in the past — creating, modifying or deleting an invoice, validating a receipt, and so on. The speakers offer an analogy with a Formula 1 race: classic BI shows the starting grid and the finish line, while process data shows the pit stops and incidents during the race — essential for understanding how to improve future performance, including from counter-examples (processes that didn't meet a given service level).

Many clients define target procedures in BPMN format or compliance rules: if AI is ever to operate a process, it must follow the same framework imposed on human operators.

Historical analysis of events makes it possible to predict a likely trajectory "all else being equal," or to run constrained analyses (what happens if a given supply chain parameter changes?).

Celonis describes its platform as resting on three elements: high-performance infrastructure developed over 15 years, the context model itself — built as a living digital twin of operations — and the ability to create composable solutions from this model, in three steps: analyze processes, design tomorrow's process by integrating agentic AI fed by this business context, then orchestrate and industrialize this new way of working.

During a live demonstration, Celonis illustrates its approach on a shop-floor case, connecting to an MES, a warehouse management system (WMS) and an advanced planning system (APS). The context model turns the complexity of these systems into understandable business objects and events (manufacturing order, work center, purchase order).

The analysis reveals that, out of 10,489 manufacturing orders concerned, a quality check failed, significantly slowing down the process. Digging further, an 11-hour gap between planned and actual production start is identified, traced back to a machine breakdown, itself caused by a delayed delivery of maintenance parts — from an unreliable supplier. Celonis puts the associated opportunity at around $8 million, with the machines running only 65% of the time against an 80% target based on the benchmark used.

Building on this, Celonis proposes designing a predictive maintenance agent capable of identifying at-risk suppliers ahead of time, then orchestrating an automated response: when a delay is detected, the agent alerts the user and recommends, in this example, transferring a part available in excess stock from a neighboring plant rather than waiting for the failing supplier — an action which, once validated, directly triggers the transfer order in the ERP.

Celonis, a German company, cites the example of BMW, which states that every vehicle leaving its plants has been touched at some point by the platform — with a use case on synchronizing engine assembly lines to avoid delivery delays caused by desynchronization between parallel lines.

At Mercedes, the partnership began during the Covid crisis and the microchip shortage, with a prioritization challenge: which vehicle should get priority access to available electronic components? Celonis built a supply chain "control tower," with simulation scenarios to identify the best possible allocation.

Celonis positions itself as agnostic toward AI consumption layers (Copilot, custom agents) and notably works with Microsoft to feed solutions like Copilot Studio. The typical architecture deployed at clients includes a data layer, the Celonis context layer, a consumption layer (agents, copilots) and an observability layer — for example via Microsoft's Agent 365 tool, which monitors agent access, while Celonis simultaneously measures the economic value generated by each agent since its deployment.

One technical argument highlighted: preparing the necessary context for agents in advance would, according to Celonis's observations, reduce token consumption by a factor of two to three for the same use case, by sparing agents from having to reconstruct complex queries or data transformations themselves.

Celonis's central message is that the value of agentic AI in the enterprise depends directly on the quality of the business context it can access. Without a detailed, genuine understanding of processes — built here through process mining — AI remains confined to generic answers; with that context, it becomes capable of directly actionable recommendations that can be measured in business value.

Watch the full session, along with the other Tech For Industry Show replays, on our dedicated page.