Standards in your company AI: what DIN, EN and ISO licences allow for RAG and assistants

Can standards go into ChatGPT or an internal AI?

In most cases, not without a further licence. DIN Media, which sells DIN standards, states on its support page that its terms currently exclude using standards in AI and offers a separate AI licence for internal, isolated AI environments; ISO's current terms for its Online Browsing Platform forbid embedding, prompting or querying ISO content with AI systems unless a separate licence allows it. Uploading a purchased PDF to a public chatbot is therefore outside both sets of terms, and an internal assistant over standards needs the AI licence and a set-up that matches its conditions.

This page is for the head of quality, engineering or compliance who wants the standards in the company assistant, and for the IT team that builds it. It quotes the terms as published by DIN Media and ISO on the dates given, explains what the EU copyright rules on text and data mining add, and then describes how such an assistant is built on your own server. It is general information, not legal advice: your legal counsel confirms how the terms apply to your licences.

DIN Media's terms and the AI licence for DIN standards

DIN Media's support article on importing standards into ChatGPT and AI, read on 25 October 2026, is short. It says that the general terms and conditions "currently exclude direct integration into AI systems", that DIN Media is developing a usage model, and that "Starting in the fourth quarter of 2025, it will be possible to obtain an appropriate AI license."

The licence page, headed "Standards in AI Applications? How to Use Them Legally", describes that offer. The AI licence extends the usage rights of purchased or subscribed DIN standards to "internal use, and additional processing in internal AI applications". Three conditions are listed:

1. The company is a DIN member with internal reproduction rights, or holds a DIN network licence together with DIN membership.
2. It has bought or subscribed to the digital DIN standards concerned and holds the network licence for them.
3. The use "takes place in an internal, isolated AI environment without training public or external AI models."

There are two models. Either the company licenses all DIN standards it holds that will be processed in its internal, encapsulated AI system, or it subscribes to a collection on DIN Media's Nautos platform and buys an AI licence for all standards in that collection. The number of user accounts "is determined by the required reproduction rights for DIN documents, not by the AI license", so the AI licence adds no users. Standards may be used in the internal AI application "for the duration of the license"; details are governed by the contracts.

The FAQ on the same page adds three points that matter for the project. A network licence or the legal provisions on text and data mining are not a sufficient basis for AI systems such as ChatGPT or Gemini. Using standards in AI without the licence, in DIN Media's words, means companies "infringe protected usage rights and risk injunctive relief, retroactive license fees, and claims for damages". And for wrong answers, the human-in-the-loop principle applies: "results should be checked against the original document", and the user is responsible for decisions based on AI output.

On the scope, the page says the offer currently covers "DIN, except VDE; ISO may follow", and that there is "currently no comparable AI license offering" for VDE and VDI standards. The page gives no technical specification of "isolated", and it does not say whether a company's own on-premise system qualifies outside Nautos, so ask DIN Media to confirm your design in writing.

What ISO's terms say about ISO standards in AI and RAG

ISO's Online Browsing Platform (OBP) terms of use, published on 18 September 2026 and read on 25 October 2026, contain a section headed "Artificial Intelligence and Machine Learning Uses". It excludes "training, fine-tuning, or evaluating language models" and also "embedding, prompting, querying, comparing, or otherwise enabling an AI or ML models or system to ingest" ISO content, unless a separate licence agreement issued by ISO, an ISO member body or a distributor appointed by ISO expressly allows the use. The same section states that "ISO expressly opts out of the text-and-data-mining exception under Article 4 of Directive (EU) 2019/790" and that the licensee may not rely on statutory text and data mining exceptions for AI use.

The copyright notice on ISO's website is broader. With the exception of its Open data page, "no ISO content may be used for any machine learning and/or artificial intelligence and/or similar technologies", including to "prompt or otherwise enable artificial intelligence or similar tools to generate responses". The OBP terms govern the OBP and its content; they do not address PDFs bought from the ISO Store or from a national member body, whose own terms apply.

EN standards, IEC and the other bodies

Which terms apply follows from where your copy came from. A DIN EN or DIN EN ISO document bought from DIN Media falls under DIN Media's terms; a document from the ISO Store or OBP falls under ISO's. IEC, CEN and CENELEC members in other countries, VDE, VDI and industry bodies publish their own terms, and we have not reviewed them for this page. Make a list of every source of standards in the company, with the licence each copy came with, before anyone builds an index. DIN Media's statement that VDE and VDI standards have no AI licence offer yet means those documents stay out of the index for now.

The text and data mining exception

Directive (EU) 2019/790 on copyright in the Digital Single Market defines text and data mining as "any automated analytical technique aimed at analysing text and data in digital form" to generate information such as patterns, trends and correlations. Article 3 gives research organisations and cultural heritage institutions an exception for works they have lawful access to. Article 4 gives everyone else an exception for "lawfully accessible works", but only if the use "has not been expressly reserved by their rightholders in an appropriate manner, such as machine-readable means" for content made publicly available online.

Two points limit what a company can draw from this. ISO has made that reservation in its terms, and DIN Media states that the text and data mining provisions do not suffice for AI systems. And Article 7(1) makes contract terms contrary to the exceptions in "Articles 3, 5 and 6" unenforceable; Article 4 is not in that list, so a licence can restrict commercial text and data mining. National laws transpose the directive and can differ in detail. Your legal counsel confirms whether any exception applies to your use; for a company assistant, plan on the licence.

Harmonised standards and the judgment of 5 March 2024

The Court of Justice ruled in case C-588/21 P, Public.Resource.Org and Right to Know against the Commission, on four harmonised standards adopted by CEN, three on toy safety and one on nickel release. It held that "the requested harmonised standards form part of EU law" and that there is an overriding public interest in their disclosure under the EU's rules on public access to documents, and it annulled the Commission's refusal to give access. The Court did not examine the ground on copyright. The judgment concerns access to documents held by the Commission; it does not say that standards bought from a standards body may be loaded into an AI system, and it changes nothing in the terms quoted above.

We build the server for an internal standards assistant and set up the model and retrieval through our Private AI/ML service; send us which standards collections you hold and how many people will use them, and our team replies with a configuration.

What an internal, isolated AI environment looks like

Neither DIN Media nor ISO publishes a technical definition, so the design below follows the words of the terms and leaves the confirmation to the licensor. A retrieval-augmented assistant (RAG: the system searches the documents first and the model answers from the passages it finds) needs five parts, and all five stay inside the company.

Our design summary; conditions quoted from DIN Media's AI licence page and ISO's OBP terms, read 25 October 2026.

The language model runs on the server, not through an external API, so no passage of a standard leaves the network and no outside provider can use it for training. An open model such as gpt-oss-120b, published under Apache 2.0, answers locally; the model itself is not trained on the standards. DIN Media excludes training "public or external" models and does not say whether fine-tuning an internal model on standards is covered, so leave fine-tuning out unless DIN Media confirms it. Retrieval is also the better technique here: a new edition replaces the old passages in the index without retraining anything.

Access is the second condition. DIN Media ties the number of users to your reproduction rights, so only the people covered by the licence may query the standards collection. Our guide to RAG on company data explains why that filter belongs in the search: a prompt that tells the model to hide documents is no access control. Connect the assistant to the company directory, so that leavers lose access the day their account is closed.

The answer should name the standard, the edition and the clause it used, and link the user to the original document. That supports the check DIN Media asks for, "results should be checked against the original document", and lets an auditor see which edition an answer came from. Keep the logs, but treat them as company data with a retention period, since they hold employees' questions.

Keeping licensed and unlicensed documents apart

Most companies hold standards from several sources under different terms, plus their own work instructions that quote them. Mixing them in one index makes it impossible to remove one licensor's documents later. Give each licence its own collection and store on every passage the issuing body, the document number, the edition, the licence reference and its end date.

That metadata does four jobs. The search filter uses it to limit each collection to its licensed users. A scheduled job deletes a collection, its vectors and cached passages when a licence ends, since DIN Media allows use only for the duration of the licence. A withdrawn edition is marked as such, so that the assistant does not answer from it by default. And VDE, VDI and ISO documents without an AI licence never enter the index, even if they sit on the same file share. Our guides to vector databases for RAG and to preparing company documents for an AI assistant cover filters and document hygiene.

Internal documents are a separate question. A work instruction that summarises a standard in your own words is your text; one that copies whole tables from it carries the standard's terms with it. Quality management decides which internal documents go into which collection.

Which server a standards assistant needs

A standards assistant is a modest workload. Questions are short, the passages are a few pages, and the users are an engineering, quality or purchasing team, not the whole company. Our guide to private ChatGPT server sizing by company size estimates that, with its example usage values, 500 employees produce about 20 requests in flight at the peak, and that one RTX PRO 6000 Server Edition holds about 19 conversations of 32K tokens with gpt-oss-120b. A standards assistant for a few dozen engineers uses a fraction of that.

Reference configurations on our AI servers page; conversations from our company-size guide (gpt-oss-120b, 32K context, 16-bit cache); power by our estimate (cards at 600 W, two processors, plus 15 per cent).

That is why the standards collection usually belongs on the company AI server that already serves other departments. The entry server with two NVIDIA RTX PRO 6000 Blackwell Server Edition cards, two AMD EPYC processors, 384 GB of DDR5 ECC memory and two NVMe drives runs the language model on one card and splits the other by MIG into up to four isolated instances for the embedding model, the reranker and test work. The same assistant then also answers from your own procedures, quality management documents or machine manuals, each collection with its own access rules. By our estimate the two-card server draws about 1.8 kW and fits one 230 V 16 A feed, which carries 2.94 kW at 80 per cent.

Without a server room, our Private AI/ML service runs the platform on dedicated hardware in a Tier-3 data centre in Lithuania, inside the EU. Whether a hosted platform counts as "internal, isolated" under your AI licence is a question for DIN Media before you choose it. Our article on ChatGPT Enterprise or an own AI server compares the two routes in general.

We supply the server assembled and tested before delivery, with the operating system, driver, CUDA and a container runtime on request; describe your standards collections and the teams that will use them, and we propose a set-up you can show to your licensor.

Before the first standard goes into the index

1. Inventory. List every source of standards in the company: DIN Media subscriptions and network licences, ISO purchases, other bodies, copies on file shares.
2. Licence. Ask each licensor whether an AI licence exists for your use and what it requires. Today that means DIN Media's AI licence for DIN standards; for ISO, a separate licence from ISO, a member body or a distributor appointed by ISO.
3. Design in writing. Describe the set-up (on-premise model, no external API, no training, access filter, deletion at licence end) and have the licensor confirm that it counts as internal and isolated.
4. Collections. One collection per licence, with metadata for body, edition, licence and end date.
5. Users. Map the licensed users to a role in the company directory and filter every search by it.
6. Answers. Cite standard, edition and clause in every answer, and train users to check the original.
7. End of licence. Schedule deletion of the collection and its vectors, and test it once before go-live.

What we supply

You can buy this system from us: our team sizes the configuration to your workload and budget and replies within one business day. We build AI servers to order with two, four or eight RTX PRO 6000 Server Edition cards, assembled and tested before delivery, delivered anywhere in the EU on one EU contract and invoice, with manufacturer warranty on every component handled through us; quantities and lead time are confirmed in writing with the quote. Our Private AI/ML service, with our engineering partner Vixen.UNO, sets up the model, the retrieval with access filters and the connection to your directory. The licences for the standards themselves come from the standards bodies; we do not sell or broker them. Our guide to buying a GPU server for LLMs lists every configuration, and our comparison of open LLM licences for company use covers the model's own licence.

FAQ