Manufacturers prioritise traceability and evidence in AI-driven decisions

Manufacturers are getting better at using AI to summarise documents, surface patterns and draft reports. But far fewer are prepared to let those systems shape a decision that affects a machine, a product or a compliance outcome.

That distinction matters. In a plant, an AI tool is not just helping someone write faster or search more efficiently. It may be suggesting a process adjustment, escalating a quality issue, supporting supplier approval or triggering a workflow with real-...

world consequences such as scrap, downtime, rework or a safety incident. At IIoT World’s AI Manufacturing Day 2026, Chris Huff and Anthony Vigliotti of Adlib Software, together with Mathias Oppelt of Siemens, argued that the real question is not whether AI sounds convincing, but whether its recommendation can be traced, checked and justified before action is taken.

The panel’s central point was that manufacturing AI has to be dependable, not merely persuasive. General-purpose models are often optimised to produce fluent answers, but production environments rely on current specifications, approved procedures, tolerances, quality records and engineering context. If any of those pieces are missing, an answer can still appear credible while being wrong for the plant. In other words, the technology must understand the language of industry: drawings, technical notes, simulations, specifications and version-controlled documents, not text alone.

That concern is reflected in wider research. A recent review in MDPI’s Sensors journal describes AI’s growing role in predictive maintenance, quality control and supply chain optimisation, while also warning about the black-box nature of many models, bias and the absence of strong trust frameworks. The paper links those issues to the Industry 5.0 idea of resilient, ethical innovation and human-machine collaboration, which mirrors the manufacturing industry’s broader struggle to balance automation with accountability.

The IIoT World discussion drew a useful distinction between insight and action. Many manufacturers have already used AI to search knowledge bases, classify documents or summarise maintenance records. Those are valuable tasks, but they are not the same as allowing an AI system to move work forward. The risk rises as AI shifts from retrieving information to generating a recommendation and then into a workflow step. A system might route a supplier file for review, flag a quality exception or recommend a change to a process setting. In each case, users need to know exactly what evidence underpinned the suggestion.

That is where traceability becomes essential. A reviewer should be able to see the document used, the version consulted, the approval status, the relevant page or table and the reasoning chain that led to the recommendation. If the system cannot provide that, people are forced to reconstruct the evidence themselves, which undermines the very efficiency AI is supposed to create.

The challenge is compounded by the reality of brownfield manufacturing environments. Plants were built to make products, not to feed model-ready data into AI systems. Critical information is often scattered across standard operating procedures, batch records, engineering drawings, supplier specifications, quality reports, scanned PDFs, SharePoint folders, legacy systems and OEM binders. The answer may exist somewhere, but that does not mean an AI system can locate it, interpret it correctly and connect it to the right decision.

This is why the source material matters as much as the model. The panel argued that traceability has to start at ingestion, not as an afterthought. If AI is going to recommend a supplier change, a process adjustment or a quality action, the reviewer must be able to see the underlying evidence immediately, including the chain of custody behind the record. Treating documents as evidence, rather than simple inputs, raises the standard significantly.

Related work in industry reinforces the point. ElixirData’s discrete manufacturing material describes a context-graph approach that links machines, materials, schedules, quality specifications and supply chains to preserve full traceability. Meanwhile, a Springer review on transparent supply chains argues that AI, IoT and blockchain can improve visibility and resilience when they are combined in a disciplined way. The common thread is that context is not optional; it is what makes intelligent systems usable in operational settings.

Human review still matters, particularly when AI gets closer to production, maintenance, quality or engineering decisions. But the reviewer can only add value if the system presents enough evidence to support a proper judgement. If the output is opaque, the person must return to the source material and check everything manually. That slows adoption and limits the benefit of the technology. There is also a practical constraint: AI models can be limited by context windows and token capacity, meaning they may not process every relevant detail at once. In manufacturing, that can be enough to omit the one item that changes the decision.

For that reason, simple chat-style document tools often fall short in industrial settings. Manufacturing records are rarely neat text files. They may include diagrams, handwritten notes, embedded images, tables, scans, checkmarks and engineering annotations. If those elements are not handled properly before the AI system acts, the output may look clean while silently missing a crucial exception.

The practical answer, the panel suggested, is to start where the risk is manageable. Procurement, contract review, supplier onboarding, regulatory filing and back-office quality tasks are easier places to begin because they are usually better bounded and less immediately physical than direct shop-floor control. In those settings, AI can be tested on whether it can support action using complete, traceable and usable information. That is a more realistic path than giving a model direct influence over a live process from the outset.

TechRadar recently made a similar argument in the context of supply chain volatility, noting that AI can help firms respond to disruption only when it is embedded in disciplined workflows with human oversight, transparency and accountability. That principle applies just as strongly in manufacturing, where the cost of acting on incomplete context can be immediate and visible.

The broader lesson is that manufacturers do not need perfect AI before moving forward. They do need clear boundaries, reliable records and a defensible evidence trail. A useful answer is not necessarily a safe recommendation. Before an AI system is allowed to act, the organisation needs to know what it used, where the information came from, whether it was current and approved, and who owns the outcome.

That, ultimately, is the test. In manufacturing, the question is not whether AI can speak with confidence. It is whether the business can prove the decision was sound.

Source: Noah Wire Services