In my previous post, I wrote about the pre-conference workshop where we discussed the future of PLM with various PLM experts. AI was not mentioned much in this meeting, although its shadow was over us.
If you did not read my review here, it is: The week after PLM Roadmap/PDT Europe – AI! The AI here means a Dutch Ouch.
I will share a more in-depth review of the conference soon, once presentations become available, so that I can go into more detail than my first impression.
Meanwhile, in this post, I want to share the findings of the AI panel discussion during the conference—a big thank you to Jonathan Scott and Michael Kirsch for sharing their notes with me.
The panel, moderated by Alan Mendel, included, from left to right: Mikkel Haggren Brynildsen, Jørgen Dahl, Martin Eigner, Torbjörn Holm, and Markus Zimmerman.

When you see this image on the left, it is my personal opinion or comment, not part of the panel discussion.
(Did anyone notice Martin Eigner had the Hartmut Fischer name tag in front of him? Everyone knows Martin 😉)
Data first, technology second
One theme the panel kept returning to: The biggest blocker for AI in PLM is not AI; it is (qualified) data. Martin even named data completeness as the single biggest blocker.
Torbjörn added that building a strong data foundation is hard, and companies are not ready to pick it up.
A point I understand from an academic point of view, but I wonder how and when companies will realize this is a missing point – where does it start?
Jørgen‘s story could be a warning sign. In his company, GE Aerospace, they are consolidating their more than 20-year-old PLM landscape with approx. 20.000 users. Initially planned as a two-year technical program, the migration was not the issue.
They spent twelve months fixing inconsistent data. With jet engines having a thirty-year lifecycle, they never had a quality incident, illustrating that the rules and guardrails on data hold despite inconsistencies.
A standard we should keep even with AI.
Ontologies and semantics?
The panel agreed that we need a common semantic layer or ontology at the source of the data, which works far better than ontologies applied later on top of the existing infrastructure.
Martin repeated his message from the day before: semantics are created in the authoring systems, while data is administered in PLM, including simulation output and validation results checked against regulations.
The vision exists, but tools don’t support it yet. There is potential with open interfaces; MCP and APIs are essential.
Unfortunately, a CAD topology is still missing, which blocks AI use there altogether.
The missing CAD openness is an opinion I also support. If we can open up the proprietary CAD data models from the major CAD vendors, we would have a big jump forward (but they suffer from their legacy and benefit from their customer lock-in)
Trust is earned, not switched on.
Another recurring theme was traceability, repeatability and trust. With AI, when you ask the same question ten times, you might get ten different answers as LLMs are stochastic.
Markus mentioned the fundamental mismatch in timescales: AI evolves very fast, in months, but configuration and variant data must last for decades. Engineering wants the same answer to the same question every time, even as the AI models change.
Mikkel pushed back, and I think rightly: we don’t need perfectly repeatable answers; humans don’t give them either. We need answers that are good enough to be consistently true — the same standard we apply to trustworthy colleagues.
Jørgen framed it differently. We should treat AI agents like we treat our employees. Apply the same security, decision rights and accountability as with a human. That is a framework everyone understands.
Torbjörn warned that the problem is not that management distrusts AI; they trust it too much because it looks right in demos and small tests!
Interesting, Mikkel raised the human cost: reviewing in minutes what AI has produced in seconds can still take a person a day. Review becomes the bottleneck. The mental load on knowledge workers is real, and the EU AI Act’s emphasis on human oversight makes it a compliance issue too.
Here I have a different point of view – humans are not there to review AI-produced content, but they have to use AI content as additional information or context.
Still, I agree that this approach leads to a higher mental load, and organizations should adapt working hours to that, like air traffic controllers working in a broad context, who usually have shifts of a maximum of two hours.
Agents, MCP and the vendors
There was no great enthusiasm about autonomous agents, and the question was when to trust them-“human in the loop” is a vague statement here; can humans overlook the decisions and actions?
When agents call other agents, you must know what happened—which source, which process, where the data came from—but it might be too time-consuming.
Torbjörn mentioned that the Arrowhead project’s microservices approach with connectors showed this traceability is achievable.
When discussing MCP and the role of PLM vendors, Jørgen was blunt: he has essentially no expectations of them and believes vibe coding will cover much of the needs. Vendors’ ontologies are often not published.
Anyone can have Claude build an MCP server for a PLM system in minutes—and get a poor solution that overloads the system, ignores data-access security and leaves no audit trail.
Request to the vendors: proper integrations and ontologies for standard concepts like First Article Inspection, defined at the source, not bolted on at the end for “insight.”
Mikkel‘s call to action here resonated with me: jump into standards work and share the data-modeling patterns that actually work. “This is how we will survive this together.”
Markus agreed—participating in standardization builds more trust than going it alone.
Where the value actually is
The value examples were telling: better requirements and linking across design, simulation and production; operational insight at pipe and paper mills comparing real-world performance with design; specification and resource planning in construction.
And as Martin said, the value never sits in one silo; it is the broader context that provides the value. Information flows across the V-model, and the gaps between requirements, BOM and system architecture are situations where the money is lost.
His statement connects to what Mikkel called the “Turing Trap“: using AI to automate existing human-shaped processes the old way, and missing the solutions that operate across the old silos.
My favorite message and later book: Reshuffle.
Markus added nuance to the “do more with the same team” question: yes, but mainly because AI provides context, so teams can answer more complex questions—not just produce more output faster.
Guidance to take home
Mikkel: Build a knowledge warehouse and a catalog of MCP servers mapped to your data governance framework.
Jørgen: Establish guardrails, then start. The knowledge to begin already exists—stop waiting.
Martin: Look for quick, pragmatic wins — AI-driven data cleansing, variant knowledge in a graph. Not everything needs to be GenAI.
Torbjörn: Build one common semantic layer instead of repeating one in every system, with AI on top.
Markus: AI can be scary. Strong change management is needed to bring people along.
Michael Kirsch from his notes: There is no data quality without well-defined semantics. As long as information sits scattered across inaccessible silos, every AI answer is built on sand. Start connecting data across the lifecycle with shared meaning.
Conclusion
This time, a temporary conclusion. The discussion, combined with Laurent Finck’s session, presenting the results from CIMdata’s Global Study of Industrial Companies and Solution Providers, assured me that most people in the field are pragmatic, perhaps the ones from the survey a little more behind than the panelists. But the real hype comes from the vendors, not from the businesses yet.
There is work to do, for sure, discovering how and where AI can bring “sustainable” value for the business, as individual efficiency usually does not materialize as promised. Learning and curious about where we will be another half year from now.



AI in Dutch is an expression similar to the English expression “ouch“, an expression you can use when you feel mild dismay or realize a problem.
Therefore, this time we aimed for a less publicly announced event, more of a coffee machine discussion with a group of selected opinions coming to the conference or living close to Gothenburg – to keep it informal and with an open agenda.




The big differentiator, of course, is the core business model of the company. Traditional industrial manufacturing companies own the core process from the beginning to (their) end – be it service or disposal,





A few years ago, the Chief Sustainability Officer (CSO) or the Sustainability team sat in a separate silo within the organization. Just as we had the CDO (Chief Digital Officer) ten years ago, the separate CFO role is now less needed.
If you have time to study and read, look at this development documented over time. Our interviews are on the 

During my summer break, I have been thinking about why companies struggle to start a digital transformation and how businesses keep going, struggling to change.



This approach can lead to ludicrous situations, where an engineer has to request a part number when designing a new product, or, in another case, use a temporary part number until the part is mature enough to be registered in ERP, according to the ERP standard. These are real examples I encountered.
For me, the discussion is closed; many companies already know the theory and best practices from the past or for their current operations. But will they work on a better future?
In my network, many PLM experts, like 

I wrote about this topic in June, addressing the mood in the field at the time: 


In a recent PGGA discussion, Pierre David shared his perspective on how Design for Sustainability (DfS) is evolving from an environmental assessment activity into a core part of product development. Central to this evolution is a shift from traditional “eco-design” towards Design to Value, where sustainability is considered alongside cost, technical performance and other engineering objectives rather than being managed as a separate workstream.
Back on track after a side step (see: 


Markus Dietrich (CEO, team neusta) then illustrated the complementary value they bring: modern front-end development, data-driven capabilities and dashboards on top of the CONTACT platform – the modern connected PLM in my terminology.





However, I believe in a different approach: define the desired To-Be status from the company’s business strategy for the coming years, then derive the enterprise processes and the underlying data model needed to get there.


What surprised me first was the roadmap’s level of detail. Most vendors talk about roadmaps but rarely share specifics or commit to capabilities before a release is actually available. Here, I saw the opposite.
Comparing the agendas of the 2022 and 2024 editions with this year’s, I noticed sustainability has become less prominent — less hype, fewer dedicated topics. Understandable, as we’re overwhelmed by other immediate global challenges, despite the extreme climate warnings we’re facing.


As a reminder: Have a look at our last PGGA interview to get ready for this episode: 

In nearly twenty years of coaching PLM implementations, I’ve noticed something striking: these projects often mirror politics—not just in complexity, but in the blame game that follows when things go wrong.


In the past, debates focused on who was to blame for project problems, often blaming the stakeholder who was not at the table.
But one issue always remained:
For decades, we have heard horror stories about failed PLM implementations, each supposedly explained by one simple cause.
That only works if the company owns its vision and strategy. An MVP approach also demands end-to-end stakeholder involvement, because everyone contributes to the solution. At the same time, our limbic brain works against us: it pushes us to protect what we know and react strongly to change.




The posts I referred to highlight real problems from the past and suggest that new insights and AI might help us build better businesses. The question is whether that promise will be fulfilled.
AI could help businesses break through organizational silos by pulling together information across functions.
That is why I use the term human thread: the network of relationships that connects people across the business. 



Despite political headwinds, businesses have been implementing more sustainability initiatives, and we were curious to hear from PLM vendors and implementers about what they are currently observing and offering to the field.
I think that it doesn't take a lot of imagination today to at least create some non-vendor-specific tooling. Build it…
Effective collaboration in product lifecycle management can help teams connect different functions and reduce gaps between development stages. I find…
Hi Jos, Knowing your background in methodology and education, I wanted to share a longer article with you: “What is…
Interesting reflection, Jos. In my experience, the situation you describe is very recognizable. At the company where I work, sustainability…
[…] (The following post from PLM Green Global Alliance cofounder Jos Voskuil first appeared in his European PLM-focused blog HERE.) […]