It's the figure that grabbed the room: according to an MIT study cited during the session, 95% of AI pilot projects never make it into production. At the Tech For Industry Show, Karl Fonteneau, Customer Success Manager (9altitudes, an ERP and PLM integrator for industry, a Microsoft and PTC partner) and Jean Jacques Galliano, Partner Technical Director (PTC, a software vendor for the manufactured product development lifecycle) detailed why — and proposed a three-step method: structure the data, ensure the digital thread, then only add AI.

Technology choices amid an overwhelming software landscape, supply chains that need to become more dynamic, geopolitical tensions, demand for ever-faster new products, constrained profitability, quality and compliance: it's in this context that AI promises to automate repetitive tasks and free up innovation time for engineers and designers. A well-known, well-understood promise — provided projects are approached the right way.

The MIT study cited identifies four major causes: generic AI, which accesses internet data rather than your own product data — "we're being pushed to sell AI, but not under just any conditions," the integrator admits; standard workflows that ignore each company's own internal processes; poor or incorrect data; and, as a logical consequence, POCs that validate a concept… and never go any further. "Time spent on topics that will go nowhere."

The heart of the method: AI cannot interpret raw data — the PDFs, Word files, and siloed data on servers that make up most companies' data assets. Four steps to fix this: start from the raw data; color it with semantics — history, status, version, context; link it to other data via ontology — knowledge graphs: attaching the quality record, technical requirement, and corresponding test file to a 3D CAD file; and finally, governance — deciding who and which system can access which data, with what rights, thereby securing AI's access to the ecosystem too.

Once data is structured within each system, continuity between them still needs to be ensured: ALM (requirements), CAD, PLM, ERP, MES, CPQ, CRM and SLM (maintenance). That's the "digital thread" — call it the industrial journey — which guarantees smooth product data flow between departments. With a telling field observation: these projects often reveal that "departments don't talk to each other" — finance, the recipient of every number, knows this all too well. And a reminder: Microsoft-PTC integration has been worked on "for years, even before AI" — AI only adds a complementary service if everything has been put in place upstream.

Concrete examples presented: in ALM, faced with the millions of requirements of a car, AI detects redundancies, collisions between requirements, and gaps in the test plan ("has this traffic radar light actually been tested?"); in CAD, generative design (genetic algorithms) creates lighter shapes that use less material — a direct contribution to sustainability and the circular economy; in PLM, deduplicating part references — knowing that a managed reference can cost a company up to €6,000 a year; in ERP, an agent focused on dormant stock proposes actions based on available data.

The video presented walks through end-to-end change management: a change request launched in Teams, a change request in the PTC PLM, a workflow with justifications, AI-generated proposals (supplier impacts, costs) that the user checks and validates, a quote request via the ERP, and the final part update in the Microsoft ERP — with supply immediately having the changes. The stated guiding principle: AI proposes, human validates — "maybe one day we'll let AI act with full confidence, but for now we check."

The exchange with the audience covered very concrete points: MCP (Model Context Protocol) servers provided by the vendors themselves (Microsoft, PTC) to connect agents to software components; data location — chosen at deployment time (European, American, or other servers) — and internal encryption keys that clients can manage themselves via a digital vault (citing the example of Swiss insurers who only moved to the cloud under this condition); the ability to cut off AI's web access to work only on internal data, leaving it up to the company to ingest supplier documentation; and the fight against shadow AI — "everyone uses their own AI and sends company information all over the place": the choice of a single, governed AI, with control tools (Microsoft Purview was cited) that track who's using what and flag non-compliant usage. One final organizational point, in response to a question: prefer AI experts embedded within each business function who coordinate with each other, rather than a single, disconnected central AI team.

One: AI's value doesn't lie in a single product — first map the complete flow and tools of your industrial journey. Two: performance only comes from connected contexts — otherwise, you get fragments of answers, potentially misleading for lack of upstream-downstream visibility. Three: data first, humans always — a nod to a data scientist in cross-training before the session, worried about her future: "don't worry, plenty of companies still haven't worked on their data." The human dimension, the duo concludes, "is the essence of industry."

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