Searching Standards With AI: Why the Citation Decides Whether the Answer Is Usable
An AI answer about a standard without a file and page reference is not merely inconvenient in engineering — it is unusable. Why that is, and how to recognise a system that holds up.
Standards are versioned, get withdrawn and cross-reference each other. An AI answer to a question about a standard is therefore only usable if it cites the document, the edition and the page. Without that reference you can neither check whether the cited edition still applies, nor use the statement in a risk assessment or conformity assessment.
Why questions about standards are a special case
Most questions in day-to-day engineering tolerate an approximate answer. A question about a standard does not.
Anyone who needs to know which requirements apply to a guard does not need a plausible summary. They need the passage — document, edition, clause. Because what counts in the end is not the answer but the evidence that it comes from the edition in force.
That is not a formality. In conformity assessment the burden of proof sits with the manufacturer. A statement whose origin nobody can reconstruct is worthless in the technical file.
What makes standards technically hard to search
Four properties that together create the problem:
They are versioned. There is no such thing as DIN EN ISO 12100; there is an edition of it. Cite the wrong edition and you have cited wrongly — even if the sentence itself is correct.
They get withdrawn and replaced. Machinery Directive 2006/42/EC is superseded from January 2027 by the Machinery Regulation (EU) 2023/1230. Anyone looking something up today has to know which legal status applies to their product.
They cross-reference each other. A question about functional safety leads from the risk assessment under DIN EN ISO 12100 to the evaluation of control system parts under DIN EN ISO 13849. The answer rarely sits in a single document.
They are licensed. Standards are copyrighted. Where they are stored and who processes them is not merely a question of convenience.
The four properties and what they mean for a search system
| Property | Consequence for search |
|---|---|
| Versioned | The edition has to be part of the answer, not just the number |
| Get replaced | The system must show which document it is reading from — not answer from memory |
| Cross-reference | An answer spans several documents; every step needs its own citation |
| Licensed | The documents must not leave the controlled environment |
Where full-text search runs into its limit
Full-text search in a PDF reader finds occurrences. That is enough as long as you know the term something is filed under.
With standards you often do not. The question is "may the operator intervene during operation"; what has to be found is a clause about interlocking guards. Between question and passage sits technical vocabulary — and that is exactly what full-text search excludes.
Then there is the reference chain. Anyone who has to assemble an answer from three documents does that three times over with full-text search — and afterwards cannot be sure they did not miss a fourth.
What an AI assistant changes — and what it does not
An assistant working on your own body of standards shifts the effort: from "where does it say that" to "is that correct". That is the real gain. Not the answer itself, but the shortened path to the referenced passage.
What it does not change: the professional assessment. Whether a requirement applies in a given case, which exemption takes effect, how it is weighted in the risk assessment — that remains an engineering decision with responsibility behind it. A system that blurs this distinction is more dangerous than useful.
This is why the citation is not an add-on feature. It is what makes the assessment possible in the first place.
How to recognise a system that holds up
Four checks that can be settled in a ten-minute demo:
- Does every statement point to a document and a page? Not "according to DIN EN ISO 13849", but the specific file and the specific page, reachable in one click.
- Is the edition visible? A standard number without a year is half a citation.
- Does the system say when it finds nothing? An assistant that composes an answer to every question has no awareness of its limits. "I cannot find anything on that in the stored documents" is the more valuable answer.
- Where do the documents sit? With licensed standards that is a legal question, not a preference. EU hosting is the default here, not a nice-to-have.
Point 3 is underrated. A system that would rather sound plausible than admit the information is missing produces exactly the error you can least afford in standards research: a statement nobody questions any more, because it was phrased so confidently.
What this looks like in practice
KoAssist answers questions from the documents a company already holds — standards, directives, internal specifications, project records — and supplies the file and the page with every statement. The chain of evidence is not a separate view; it is part of the answer.
What that means day to day shows in adjacent tasks: the same mechanism carries looking things up in existing technical documentation and onboarding new design engineers, where the same questions come up again and again.
What KoAssist does not do: produce standards, assess conformity, or take on responsibility. It shortens the path to the passage. The judgement stays with the engineer — and that is the right division.
FAQ
Why is an AI answer about a standard without a citation not enough?
Because the answer cannot be verified. Standards exist in several editions, are withdrawn and replaced. Without naming the document, the edition and the page, there is no way to establish whether the statement comes from the edition in force. In conformity assessment an unreferenced statement is worthless, because the burden of proof sits with the manufacturer.
Can I simply upload standards to ChatGPT and ask questions?
Technically yes, practically problematic. First, standards are copyrighted and licensed; passing them to a public service is usually not covered. Second, the link to the edition is missing: the model answers without making clear which edition the statement came from, or whether it came from training data at all.
What is the difference between full-text search and AI-based standards search?
Full-text search finds occurrences of a term; AI search answers a question. The practical difference shows up in reference chains: anyone looking up safety requirements gets passed from one standard to the next. Full-text search returns hits per document; an assistant can summarise the chain — but then has to cite every step.
How do I recognise a system that holds up for questions about standards?
By four properties: every statement points to a document and a page. The edition is visible, not just the standard number. The system says when it cannot find something instead of composing an answer. And the documents stay where licensing allows them to stay.
Does AI replace standards research by a subject-matter expert?
No. It shortens finding the passage, not assessing it. Whether a requirement applies in a specific case, which exemptions take effect and how it is weighted in the risk assessment remains a professional decision carrying responsibility. The time saved is in searching, not in judging.
Roschiwal + Partner: Making a Large Document Archive Accessible, Drawings Included
Data that has grown over years is not a search problem, it is an access problem. How Roschiwal + Partner put its own archive to use, and why CAD and drawings are the harder part.
A chatbot for technical documentation: what it actually has to do
A chatbot on your own technical documentation sounds straightforward. In practice its usefulness is decided at four points, and generic chatbots fail at all of them.
What is design automation?
The term covers very different things, from a parametric template to a full configurator. A look at the stages, and at the precondition that is almost always underestimated.
Less searching.
More engineering.
In a 30-minute demo, we show KoAssist working with your own documents and discuss setup, integrations and pricing for your team size.