Many companies have invested heavily in AI – only to discover, once it's up and running, that something fundamental is missing. It isn't the technology.
Imagine an experienced maintenance technician leaving the company. He knows which maintenance instruction actually applies – not the one in the system, but the one that was updated after the last machine modification. He knows who to call when a test approval is missing. He knows where to find the current drawing, even though it was never entered into the PLM system. None of this knowledge is written down anywhere. It's simply gone.
This is exactly the problem many companies are now trying to solve with AI. And it's exactly where they fail – not because AI is bad, but because AI is only ever as good as the knowledge base it works from.
What the hype doesn't tell you
"ChatGPT for the enterprise" sounds tempting. Employees ask questions in plain language, and the AI delivers answers. In a pilot project, this often works surprisingly well.
But as soon as things get serious, the real problem appears: a language model doesn't know which version of a work instruction is actually approved. It doesn't know whether a document applies to Plant A or Plant B. It has no concept of internal approval processes, no sense of responsibilities, and no way to tell "still valid" from "replaced long ago".
An answer that sounds plausible but is based on the wrong document isn't an advantage on the shop floor – it's a risk.
According to Gartner, at least 50% of all GenAI projects are abandoned after the initial proof of concept. Not because of the technology, but because of poor data quality and a lack of traceability. In the DACH region, there's an added dimension: compliance requirements, works councils, and liability concerns leave no room for AI that can't cite its sources.
The real problem: documents without context
Most industrial companies don't have bad data. They have good data – just without context.
A document sitting in SharePoint doesn't "know" that it belongs to the machine in Hall 3, that it was replaced by a new version last month, or that the quality manager at Plant B is responsible for it. These relationships exist – but they live in employees' heads, not in the system.
That's the difference between data structure and knowledge structure. Folders, version numbers, and libraries create order – but not meaning. AI needs both.
The consequences show up in everyday work: employees spend 45 to 60 minutes every day searching and asking around. Rework happens because outdated documents were used. And eventually, every team builds its own unofficial filing system – because the official one doesn't deliver what's actually needed.
What actually helps: a context layer
The solution isn't "a better AI model." It's building the right foundation.
What makes AI productive in an industrial setting is a context layer – a semantic connection between all the systems that are already in use. ERP, PLM, QMS, DMS, SharePoint – they all stay exactly as they are. The context layer is what makes them usable together.
In concrete terms: content isn't just stored, it's classified. Which document applies to which scope? Who approved it, when, and in what status? Which work instruction belongs to which inspection plan, which drawing, which deviation report?
Only once these relationships are stored in a machine-readable form can AI deliver reliable answers – with a traceable source reference, not a well-meaning guess.
Semantic search: the fastest first step
You don't need to start big. Semantic search is a pragmatic first step that delivers real value within six to ten weeks.
Unlike classic full-text search, semantic search finds content by meaning – not just by exact wording. It understands synonyms, abbreviations, and department-specific terminology. Instead of a hundred results, it shows the relevant ones – filtered by plant, product, and approval status.
The path there: identify the twenty most common search queries from engineering, quality, and maintenance, connect the most important knowledge sources, define the minimum context needed – and get started. Not as a large IT project, but as a focused pilot with a measurable outcome.
AI doesn't replace judgment – it frees it up
An important point for anyone rolling out AI on the shop floor: this isn't about replacing expertise. It's about freeing it up.
When a maintenance specialist spends twenty minutes searching before they can even start the actual work, that's not a minor inconvenience – it's a structural loss of productivity. AI that works from a validated knowledge base solves this problem. Responsibility for decisions stays exactly where it belongs: with people.
This builds acceptance – including with the works council. Not because AI sounds harmless, but because it's transparent: every answer comes with a source reference, every approval still runs through the organisation, and every uncertainty is clearly flagged.
What the full architectural approach looks like, which roles are critical during rollout, and what a realistic project timeline looks like – our whitepaper "Context-Enabled AI in Industry" covers all of this in detail.
