AI in PLM: near-term expectations and integration challenges

AI-generated summary

AI’s near-term value in product lifecycle management depends more on organizational readiness than on the technology itself, according to CIMdata. Looking ahead six to 24 months, the article argues that companies should treat AI as an organizational capability requiring changes to workflows, responsibilities and governance—not simply another installed tool. AI could connect mechanical, electrical and software development with manufacturing, suppliers and service, bridging CAD, simulation, PLM, ERP, BOM and MES systems. Potential benefits include faster development and better first-time quality, but many projects remain pilots, and measurable returns are still emerging. CIMdata distinguishes established predictive machine learning, used for maintenance, forecasting and estimating engineering-change impacts, from generative AI, which adds text, code and image generation and access to unstructured product knowledge. These technologies complement rather than replace one another. CIMdata’s Diego Tamburini identifies four implementation mechanisms: retrieval-augmented generation for answers grounded in company data; specialized models for prediction, vision and optimization; workflows invoking AI at predefined steps; and goal-driven agentic orchestration. These can be purchased, embedded in software or developed internally. The article also clarifies that generative design generally uses optimization and predates generative AI. Its overarching emphasis is augmentation: people remain accountable while AI improves insight and decision-making. Mistral CEO Arthur Mensch’s workflow-centered framing reinforces the need to rethink how decisions and work should flow. CIMdata’s 2026 adoption study illustrates the distance between familiarity and enterprise deployment. Among industrial respondents, 85% personally used generative AI at least several times weekly, but 58% were still researching AI or conducting pilots. Only 23% were operationalizing it in some enterprise areas, 13% had scaled it across projects or standard processes, and 19% were using or piloting agentic AI. Reported benefits included less repetitive work (41%), faster information access (36%), greater design innovation (28%), shorter design cycles (26%) and lower development costs (24%); 31% reported no measurable benefit. Broader studies support the adoption-versus-value distinction: McKinsey found 88% adoption but only about 6% attributing at least 5% of EBIT to AI; BCG found roughly 5% capturing value at scale; and MIT’s smaller Project NANDA sample found 95% of enterprise generative-AI initiatives without measurable returns yet. These studies measure different things. The study also surveyed 27 software providers and 30 service providers, exposing mismatched priorities. Users and service providers both identified poor data quality at 60%, compared with 37% of software providers assessing customer barriers. Over-promising and under-delivering was users’ leading provider complaint, cited by 56%. No software providers expected customers to prioritize innovation, although 36% said customers already realized innovation-related value. CIMdata argues that the market’s efficiency narrative understates AI’s design-discovery potential. It also warns that individual productivity gains flatten when handoffs, prioritization, organizational structures and governance cannot accommodate faster engineering. Concrete applications include NVIDIA’s Omniverse Blueprint for real-time computer-aided engineering digital twins. Its AI surrogate models learn from conventional simulation results to predict effects such as vehicle airflow rapidly, enabling faster design exploration; those approximations still require high-fidelity verification. Digital twins increasingly encompass processes, facilities and change histories, converging with digital threads. Aras describes AI-enhanced threads connecting requirements, architecture, tests, configurations, manufacturing plans and field results through PLM. The article also connects AI with extracting structured information from legacy documents, enabling model-based enterprises and extending digital transformation into robotics and automated vehicles. Governance remains a significant gap: 68% of industrial respondents cited intellectual-property protection and confidentiality concerns, 46% reliability and liability, and 35% transparency and explainability. Yet no software providers ranked governance among their three leading delivery challenges. CIMdata therefore argues that users must own accountability and guardrails and assess providers accordingly. It recommends measurable success criteria, early data and skills assessments, workflow integration, explicit autonomy and data-access rules, and evidence-based gates between pilots and production. Data remediation can be scoped to initial use cases rather than the entire enterprise. Specific capabilities arriving within six to 24 months remain speculative. CIMdata expects more complex engineering applications and shorter-lived opportunities that reward agility, but concludes that production value requires operating-model change and disciplined execution. Human oversight must move from checking every step toward setting rules, monitoring exceptions and owning outcomes, supported by workforce development, independent output validation and earlier security, privacy, regulatory and legal involvement.