Trade secrets and AI: customer drawings, recipes and NDAs in ChatGPT or on your own server

Can we upload customer drawings or recipes to ChatGPT?

Whether a customer drawing may go into ChatGPT or another cloud assistant depends first on the NDA or supply contract that came with it, and only then on the tool. For your own recipes, formulations and calculations the question is different: EU law protects them as trade secrets only if the company takes reasonable steps to keep them secret, and staff pasting them into tools the company has not approved weakens that case. An AI server under your control keeps such data inside the company and makes the steps easier to show.

We wrote this for owners, managing directors and heads of R&D, engineering, quality or sales in companies whose value sits in drawings, formulations, process parameters and price calculations: machine and tool makers, job shops, chemical, food, cosmetics and fragrance producers, and their suppliers. Personal data falls under the GDPR, which our article on local AI and GDPR covers; this article is about confidential business information. It is general information as of October 2026 and no legal advice. The directive leaves room for national rules, so your legal counsel confirms how it applies in your country and to your contracts.

What the Trade Secrets Directive asks of the holder

Directive (EU) 2016/943 of 8 June 2016 protects "undisclosed know-how and business information (trade secrets) against their unlawful acquisition, use and disclosure". Member States had to transpose it by 9 June 2018, and Article 1(1) allows them to give "more far-reaching protection", so national laws differ in detail.

Article 2(1) sets three conditions, and information has to meet all of them:

1. It is secret: not "generally known among or readily accessible to persons within the circles that normally deal with the kind of information in question".
2. It "has commercial value because it is secret".
3. It "has been subject to reasonable steps under the circumstances, by the person lawfully in control of the information, to keep it secret".

The third condition is the one a company controls. A recipe that every employee may paste into a public chatbot under any account is harder to defend as protected than one that is kept in a system with access rights and a record of who used it. Recital 14 also says what the protection does not cover: "trivial information" and "the experience and skills gained by employees in the normal course of their employment". The directive does not protect a secret against everything either. Article 3(1) makes independent discovery lawful, and the same applies to the "observation, study, disassembly or testing of a product or object that has been made available to the public", unless a valid contractual duty says otherwise.

Directive (EU) 2016/943, eur-lex.europa.eu, read 25 October 2026; the right-hand column is our reading, not legal advice.

Customer drawings under an NDA

A customer's drawing, specification or 3D model is the customer's trade secret, and your company holds it under the NDA or the confidentiality clause of the supply contract. Article 4(3) makes use or disclosure unlawful for a person "in breach of a confidentiality agreement or any other duty not to disclose the trade secret", and also for one "in breach of a contractual or any other duty to limit the use of the trade secret". Whether uploading a drawing to a cloud service is a disclosure under your NDA is a question of the contract's wording, which your legal counsel reads.

Before anyone in sales or engineering uses an AI tool on customer documents, read the NDA for these points:

• Recipients. Does it list who may receive the information, such as employees with a need to know, named subcontractors or affiliates? A cloud provider is a further company with its own staff and contractors.
• Purpose. Is use limited to quoting, producing or servicing a part? Analysing the drawing with a tool for another purpose may fall outside it.
• Location and processing. Does it require the data to stay on your own systems or in the EU, or forbid processing by third parties without written consent?
• Return and destruction. At the end of the project the drawings have to be returned or destroyed. With AI, copies also sit in the chat history, the search index, the logs and the backups.
• Notice. Does it require you to inform the customer of a breach? A drawing that went into an unapproved tool can trigger such a clause.

Where the contract is silent or unclear, ask the customer for written consent before the first upload. Our article on AI quoting from drawings for job shops describes the quoting task itself.

What business AI subscriptions state about your inputs

Provider terms answer part of the question, and they differ between personal accounts and business plans. We read OpenAI's pages on 25 October 2026 because ChatGPT is the tool most staff already use; other providers publish their own terms.

OpenAI Enterprise privacy page (updated 8 January 2026) and OpenAI Business Terms (no date shown, sections 3 and 4), both read 25 October 2026. Our summary; the provider's current pages apply.

The business terms add that OpenAI "will only use Customer Content as necessary to provide you with the Services" and "will not use Customer Content to develop or improve the Services", and they count customer content as confidential information. For personal accounts the picture is different: OpenAI's help centre describes how users can opt out of having ChatGPT conversations used to help teach its models, which means the setting sits with the individual employee. Our guide to shadow AI and an AI usage policy covers that case.

For your own secrets, a business plan with these terms may be a reasonable step. For customer drawings, the question remains whether your NDA allows a further company, with its staff and contractors, to receive the data at all.

We supply AI servers that keep drawings, recipes and calculations in your building; send us the kinds of documents and the teams involved, and our team proposes a configuration and a first task.

Reasonable steps to protect recipes and drawings when using AI

The directive does not list the steps; "under the circumstances" leaves them to the case. The measures below are the ones a company can show when it introduces AI. They are our examples, and your counsel decides which ones your situation calls for.

1. Classify the information. Name the classes: customer drawings under NDA, own formulations, price calculations and margins, process parameters, general documents.
2. Approve tools per class. State which tool may receive which class, and which classes go only to the internal assistant.
3. Give access by role. Development sees formulations, production sees the parameters it needs, sales sees prices; the assistant only searches documents the user may open.
4. Keep customers apart. One collection per customer or project in the search index, so an answer for customer A never draws on customer B's drawings.
5. Record use. A query log shows who asked what and which documents the answer used.
6. Delete at the end. When a project ends, remove its documents from the index and the chat histories within the period the NDA sets.
7. Train and remind. Staff learn the rules before they get access, and the rules sit in the employment and confidentiality agreements.

These are the same measures you would take for a file server with sensitive folders. The assistant adds the search index and the logs as new stores that hold copies of the documents.

What an AI server of your own changes

When the model runs on your own server, a drawing or a recipe in a prompt reaches no outside company. The model is a file on the server; answering does not change it or send the question elsewhere. You decide the retention of each store, the access rights follow your directory, and the documents can stay in a network segment without internet access, as our guide to an air-gapped LLM server describes.

Three things stay with you. People inside the company can still misuse what they can see, so access by role matters more than the hosting. Suppliers who install or support the server see the system: our Private AI/ML page states that engineers work in your environment under your access controls, take no copies of production data out of it, and sign an NDA before technical detail. And the server needs the same care as other systems holding secrets: encryption, updates and logged administrator access, as our guide to securing a GPU server sets out.

Reading drawings is a task for vision-language models, which take the image of a page as input; our guide to vision-language models for document AI explains the two-stage pipeline. For formulations and lab reports, our article on formulation R&D with AI on your own lab data goes into the R&D side.

Which server for confidential R&D, quoting and quality work

The size follows from the number of questions running at the same moment. With the example values of our company-size guide, a company of 100 people has about four in its busiest hour. Our users-per-card rule takes 90 per cent of the 95.6 GiB that the driver reports on a 96 GB RTX PRO 6000, less 3 GiB per card and less the model. With gpt-oss-120b, an open-weight model from OpenAI under Apache 2.0, that leaves room for 19 conversations of up to 32,000 tokens at the same moment on one card (a token is a word or part of one).

Conversations at 32K by our users-per-card rule for gpt-oss-120b with a 16-bit cache, as in our company-size guide; four cards as two copies across pairs; power by our estimate (cards at 600 W, processors, plus 15 per cent for the rest of the server).

The entry server on our AI servers page has two RTX PRO 6000 Blackwell Server Edition cards, two AMD EPYC processors, 384 GB of DDR5 ECC memory and two NVMe drives. One card runs the language model; the second is split by MIG, NVIDIA's way of dividing one card into isolated parts, for the search models and a document model. The server draws about 1.8 kW by our estimate, fits one 230 V 16 A circuit and belongs in a server room or a rack; the four-card server draws about 3.4 kW and needs 32 A or three-phase. Our company-size guide shows how head count turns into conversations at the same moment.

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 the 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, delivered by our engineering partner Vixen.UNO, sets up the model, the document search with access rights per user, separate collections and the query log. The legal reading of your NDAs stays with your counsel; we deliver the technical measures they can point to.

FAQ