Lack of structured data slows agentic AI in industrial engineering

For manufacturing companies experimenting with agentic AI in engineering work, the first obstacle is technical data, often scattered across dozens of repositories, recorded under different names from one system to another or stored in formats that software cannot interpret. This is the picture that emerges from a Massachusetts Institute of Technology (MIT) study of 28 companies and from an RSM US survey of more than a thousand North American executives.

The MIT study of 28 companies

Several manufacturing companies are experimenting with agentic AI, that is, artificial intelligence systems able to carry out tasks autonomously, in engineering activities ranging from design to searches of technical archives. A recent study by the MIT Initiative for New Manufacturing (2026), based on 33 interviews at 28 companies in the aerospace, defence, electronics and robotics sectors, collected the first assessments from the people leading these projects. The most frequent theme is the availability and quality of the technical data on which the agents would have to work.

Twenty-six different sources for a single project

The first obstacle companies face is the dispersion of technical information. An engineer at one of the companies interviewed reports having consulted 22 experts involved in the development of a single programme and having identified 26 distinct data sources. These include personal hard drives, SharePoint sites (Microsoft's document management platform), custom-built internal databases and shared folders. Added to this are multiple versions of the same document and file naming conventions that vary from one working group to the next. The founder of a components company also points out that the same item appears under different names from one system to another and that free-text fields accumulate numerous variants of the same information. Before an AI agent can work on data of this kind, the fields have to be cleaned and standardised (MIT, 2026).

A survey by the consulting firm RSM US (2026) of more than a thousand North American executives at companies already using AI, conducted in March 2026, measures the same problem across a wider sample. Among the manufacturing companies surveyed, 32% name data quality and traceability as the main obstacle to AI adoption. Many companies introduce AI through a pilot project, that is, a small-scale trial to be extended only if the outcome is positive. Among manufacturers whose pilot projects failed in the last two years, 51% attribute the failure to the data used in the trial.

In Italy the spread of AI in manufacturing remains limited and data is among the main obstacles cited. In 2025, 14.7% of manufacturing companies with at least 10 employees used AI (Istat, 2025). Among companies in all sectors that evaluated AI without adopting it, 45.2% cite the unavailability or poor quality of data as an obstacle, the third most frequent after the lack of skills and the lack of regulatory clarity.

Data confined to the company perimeter

Fragmentation is compounded by security and intellectual property constraints. The head of a division operating in the defence sector explains that most of the more advanced algorithms are available only in the cloud. The technical data handled by the division, subject to export controls, must instead remain in isolated environments, that is, in systems not connected to external networks (MIT, 2026). The same principle, on different legal grounds, applies to any company that handles process secrets or is subject to sector certifications, such as those in aerospace or medical devices. In these cases access to data is limited by contractual constraints.

Information locked in PDFs

When data exists and can be found, it is often stored in a format that an automated system cannot use. An executive at a company that develops tools for electronic engineering observes that most technical information remains locked in PDF files, a format designed to be read by a person, not queried by software. The problem also affects information that comes from outside the company. An engineer at the US space agency notes that suppliers, both the machine shops to which machining is outsourced and component manufacturers, do not offer direct access to their technical data, that is, an interface to which software can connect. For the more advanced components, moreover, prices and specifications are still obtained only through sales channels. An AI agent tasked with choosing a component or comparing several suppliers would therefore have no source from which to retrieve the information it needs (MIT, 2026).

The knowledge that stays in people's heads

Part of the technical knowledge, finally, does not appear in any document. The technical director of an additive manufacturing company observes that the most specialised expertise on the company's own technology lies largely in the experience of its employees. This is tacit knowledge, built up on the shop floor, which those who hold it tend not to formalise because they consider it obvious. With retirements, the risk is that this knowledge leaves with the people who hold it (MIT, 2026).

Europe tries to pool data

These problems were found mainly at US companies, but the fragmentation of technical data also concerns Europe. EU manufacturing comprises 2.2 million companies, mostly small and medium-sized, and generates about a quarter of the total turnover of European companies (European Commission, 2025). It was to this industrial base that the Commission addressed, in October 2025, the manufacturing, engineering and construction actions of the Apply AI Strategy, the European programme that funds and supports AI adoption. These include funding for data sharing initiatives among companies in the same sector (data pooling). Another action strengthens the testing and experimentation facilities already active in Europe, where a company can verify an AI solution on a real case before adopting it at scale. It is a different approach to the same problem, based on infrastructure shared at sector level, which spreads the cost of reorganising data across several companies.

Before choosing the AI model, map the data and skills

The difficulties with data are compounded by one with skills. Few engineers understand how the AI models they use work. An executive at a mid-sized engineering firm observes that many know how to use AI, but very few know how to develop or reconfigure it (MIT, 2026). A Deloitte survey (2026) of 3,235 executives in 24 countries, conducted between August and September 2025, adds the organisational dimension. Only one company in five has a mature governance model for autonomous AI agents, that is, rules establishing at which steps human intervention remains mandatory, how automated decisions are checked and which records of the system's behaviour must be kept.

The difficulties described by the MIT study, by the RSM US and Deloitte surveys and by the Istat data indicate that adopting an agentic AI model requires preliminary work on data and organisation. Before adopting an agent, a company must verify where the technical data required by the task is located, in what format it is stored, who holds the knowledge needed to interpret it and which governance rules regulate its use. Without these conditions, an agentic AI pilot project in engineering remains at the demonstration stage.

Sources

• Edwards, K.M., Bauer, M., Jacquillat, C., Hart, A.J., Ahmed, F. (2026): Agentic AI in Engineering and Manufacturing: Industry Perspectives on Utility, Adoption, Challenges, and Opportunities, MIT Initiative for New Manufacturing, March 2026
• European Commission (2025): AI for Manufacturing, Engineering and Construction: Commission Launches Flagship Actions under the Apply AI Strategy, 28 November 2025 (actions announced on 8 October 2025)
• RSM US (2026): Middle Market AI Survey 2026, 21 July 2026 (data collected 5-16 March 2026)
• Istat (2025): Imprese e ICT, anno 2025, Statistiche report, 15 December 2025
• Deloitte (2026): State of AI in the Enterprise, 2026 edition (survey August-September 2025)