You've invested in general-purpose AI tools — ChatGPT, Claude, Mistral — and their answers to your engineering questions remain… generic. The diagnosis from Anthony Ponceot, Technical Director (Aras), a vendor of low-code PLM solutions for 25 years, is unequivocal: industrial AI has a context problem, and that context is called digital continuity.

Applied to industrial problems, general-purpose chatbots run into the wall of data source access: without a connection to the PLM and technical data, AI has only its intrinsic knowledge and web search — enough to produce answers you could have found yourself, not relevant answers about your specific product.

Follow a product's lifecycle: the requirement is born in an ALM tool, the design in CAD or electronic design software, the simulation in a third tool — whose results will often never be linked back to the design or the requirement. Next come manufacturing routings and supplier data, then in-service life, tracked… in the CRM. All this data lives in silos. Aras's position isn't to manage everything in a single tool, but to connect all this data via the PLM platform — linking a supplier or service data point back to the original requirement, so a field issue can be traced all the way back to the requirement and trigger a design change without ever deviating from it.

An engineer who has to solve a customer issue without digital continuity queries systems one by one — CRM, requirements, design — running into access rights and information that "lives in the heads of people" who retire. The digital thread brings two advantages: secure, traceable data across systems, and role-specific paths to truth — the designer looks at the next version of the product, the field service technician looks at the product as delivered to the customer. Each actor connects to the data layer they need.

The digital flow isn't open to everyone: confidential purchasing or cost data, permissions by product line, by department (quality doesn't access the same data as engineering), even by geography (Chinese teams vs. European teams) — a point that turns out to be crucial for AI, as we'll see. And the real challenge remains opening up to the ecosystem: connecting CAD, electronic design, ERP, MES — sometimes built in-house — to the product information thread. With, in regulated industries such as medical, a legal requirement for fine-grained traceability: who modified the product, how, and how the changes were implemented.

Without digital continuity, every department checks in its own systems for the impact of a new customer requirement — sourcing, purchasing, homologation, design — taking weeks to resolve. With the digital thread: immediate navigation from the requirement to system design, to the 3D definition, all the way to test procedures and manufacturing instructions.

A component — notably an electronic one — will no longer be available past a given date: buy up remaining stock, substitute, or redesign? Without a global view, resolving this can take several months, department by department. With the digital thread: affected variants, supplier implications, alternative parts already proposed in the system, compliance impacts — everything is immediately investigable.

A customer reports a defect with a serial number. With connected, configuration-managed data: which bill of materials this serialized part came from, how it was instantiated, what similar failures have been logged — and if the design has already been modified to address this defect, the repair solution may already exist.

In an industrial setting, trust in an AI agent comes from one thing: the context provided at the moment of the request. With versioned, governed data, the agent can be pointed to the right version, the right variant — and a requirement is no longer just an identifier, but its content ("the system must operate up to 50°C under normal operating conditions") and all the relationships that derive from it.

Where AI architectures talk about GraphRAG, PLM offers "thread RAG": based on the user's question, query the PLM system, unfold the relevant relationships, and augment the agent's context with that data — including the semantics of the relationships (what the link between two objects actually means), which enables answering natural-language questions. Applications: structured ingestion of requirements from unstructured documents (PDFs), impact analysis enriched by similar past engineering changes ("I have an issue with this screw, find me solutions"), variant resolution to get the exact composition of the product in question.

An unbreakable rule: no data the user doesn't have access to should be exposed to the agent. AI respects the permissions model, configuration management, and lifecycles. And a strong warning about training: data used to train a model "then belongs to the model" — anyone using it can make use of it. For a lot of data, access at the time of the request (context augmentation) is preferable to training.

Build the backbone (the configuration management platform, starting "where it hurts" — often the requirements-design connection, two worlds with different lifecycles); integrate the ecosystem (connect the data you need, by reference or by import, without operating everything in a single tool); adaptability (systems and AI evolve in months, not years — today's standards will disappear); and access governance, which the arrival of AI must not break.

Treat lifecycle data as a strategic asset, not a byproduct — it's potentially worth more than the source code, which AI will be able to regenerate anyway, whereas the data that led to its design is irreplaceable; focus on the costliest processes (the real difficulty of a PLM project is human and organizational, not technical); prioritize governance — intentionally decide which data AI can access, and for what purpose; target high-risk, high-value use cases; and build for the future, adaptable to changes in models and technologies.

Aras's presentation is methodical: without digital continuity, industrial AI remains confined to generic answers, unable to link a requirement, a design, supplier data, or field feedback. The digital thread — and its "thread RAG" variant for AI — is the essential foundation for turning general-purpose chatbots into assistants genuinely relevant to your product, without ever sacrificing access governance.

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