Wil-Bot: Celonis and the Technical University of Munich build an AI virtual expert

Why did we build Wil-Bot, an AI version of Professor Wil van der Aalst, the Godfather of Process Mining?

One of our core commitments at Celonis is education — helping people internally and externally understand what process mining, Process Intelligence, Operational Intelligence, and Enterprise AI are, what they can do and why that matters. Wil van der Aalst is our Chief Scientist, the researcher who pioneered and named the field of process mining, and one of the world's highly-cited computer scientists.

Wil is also one person. He can only be in one room at a time.

In March 2025, we began a project to create an AI Avatar based on the professor. We presented a prototype at our internal AI Forum, with Wil himself sitting in the room watching a digital copy of himself explain his own field.

It sparked a useful argument about likeness, consent, and authority. It also made it clear that the video interaction layer wasn’t the challenge. The real issue was how to give it access to expert knowledge that was machine-legible, trustworthy, and scalable at the same time.

https://delivery-p141552-e1488202.adobeaemcloud.com/adobe/assets/urn:aaid:aem:8c88a122-1299-4b98-9d45-9047103821c5/original/as/wil-bot-joseph-steele-munich-creative-business-week.jpeg
Joseph Steele presents Wil-Bot at Munich Creative Business Week, April 2026

The collaboration: Celonis & TUM

Through the Celonis Academic Alliance, we began a structured research collaboration with students from Technical University of Munich (TUM).

Two TUM teams took it on across Winter Semester 2025/2026. Sparsh Tyagi, Noah Käuflin, Fabian Zimmermann, Julius Scheuer, Jing Pu, and Safa Kaplan conducted a systematic literature review of close to 1,200 papers, then developed, designed, built and tested working prototypes.

The projects were supervised at TUM by Isabel Hertl, M.Sc., and Alina Straßer, M.Sc., and on the Celonis side by Dr. Sarah Güsken and myself.

https://delivery-p141552-e1488202.adobeaemcloud.com/adobe/assets/urn:aaid:aem:5dca7a9b-52cb-449d-b3ac-ee0b8b268eba/original/as/wil-bot-team-at-celonis.jpeg
The Wil-Bot team at the Celonis Munich office, January 2026

How Wil-Bot works

Wil-Bot takes a question through chat or voice by a video avatar, retrieves precise fragments from a closed domain of verified knowledge, and validates the answer for confidence and provenance before returning it with sources attached. Its knowledge base combines authoritative conceptual sources — including on the one side Wil’s books, papers and lectures on the other, real case studies showing how those ideas play out in actual industry projects, backed by supplementary research and verified Celonis customer stories. Everything is chunked and tagged with metadata including source authority, knowledge type, topic, version and audience relevance. It doesn’t browse. It can’t wander.

The part that differentiates it is the Bot Constitution - a trust layer that defines what Wil-Bot is allowed to say, and how. It must disclose that it’s an AI. Every answer must cite a verifiable source and proactively asks for showing the whole source. It holds no opinions. When the knowledge base doesn’t cover a question, it’s required to say “I don’t know” rather than generate something plausible.

That refusal turned out to matter. Through internal testing - including retrieval tests, adversarial prompts and iterations of the system’s refusal behaviour - we found that clear, honest refusals inspired more trust than hedged or partially relevant answers. The same lesson has echoed in other AI systems we have developed for brands.

The simple way I express this is that trust is not a linear function of scale: an AI that attempts to answer everything is not necessarily an AI you trust.

The Bot Constitution is crucial in defining the sovereignty of the system: the rules that preserve control over what an AI system can and cannot say in someone (or something) name. As agents increasingly speak on behalf of companies, a bot constitution or something similar becomes a form of brand governance.

https://delivery-p141552-e1488202.adobeaemcloud.com/adobe/assets/urn:aaid:aem:de227dfb-c870-431b-8317-53237a56f7a2/original/as/wil-bot-architecture.jpeg
The Wil-Bot architecture: presentation, orchestration, knowledge base, LLM and validation layers. Every answer is checked for confidence and provenance before it reaches the user.

What surprised us and what we learned creating Wil-Bot

Going into the project I assumed the hard part would be the AI (i.e., picking the right model, finding a stable video interaction, authentic voice mimicry, curating a sense of humour).

They were important, but the dominant factor in whether Wil-Bot gave useful answers came down to the structural quality of the knowledge base, the rigour of the retrieval rules, and the behavioural constraints we encoded before anyone ever interacted with it.

There was something familiar in that discovery. We have access to increasingly powerful AI models, and that matters, but the context we provide them with matters much more.
Reliable virtual experts are a knowledge engineering and governance problem, not a prompt engineering problem.

The unglamorous work of structuring what the system knows and deciding what the system may and may not claim turns out to be the real work.

https://delivery-p141552-e1488202.adobeaemcloud.com/adobe/assets/urn:aaid:aem:82257e91-4217-4ed9-81c3-7cbffde23912/original/as/wil-bot-team-conference-room.jpeg
Presentation of the Wil-Bot prototypes at the Celonis office in Munich

What this means for companies

During my talk at Munich Creative Business Week I claimed that brands need to become machines. What I meant is that if brands are going to operate in an agent-mediated world, identity and messaging aren't enough. Companies need to become machine-legible systems, and I think that means three things:

A verified knowledge corpus machines can retrieve from. If you don't provide one, agents will invent it for you.

A queryable interface so humans and agents can interact with the brand directly, rather than scraping inferences off your marketing.

A constitutional layer that governs what can and cannot be said in your name.

Wil-Bot gave us a small-scale way to explore what that might look like. And what we learned was that the durable asset was never the avatar — it was the governed knowledge underneath it. Once that architecture exists, it isn’t specific to one expert. The same foundation can support other virtual experts, in other domains, forked from a shared knowledge lake.

Wil-Bot gave us a small-scale way to explore what that might look like. And what we learned was that the durable asset was never the avatar — it was the governed knowledge underneath it. Once that architecture exists, it isn’t specific to one expert. The same foundation can support other virtual experts, in other domains, forked from a shared knowledge lake.

https://delivery-p141552-e1488202.adobeaemcloud.com/adobe/assets/urn:aaid:aem:73acee77-0472-4512-acd2-2f77d809d04a/original/as/wil-bot-context-window.jpeg
The Wil-Bot interface, running in Academy mode with verified sources only.

The Wil-Bot project is a collaboration between Celonis SE and the Technical University of Munich (TUM), run through the Celonis Academic Alliance. Additional interface development by MUC:DAI.