Skip to content

Trade Secrets and AI: Protecting Know-how When Using AI Tools

For many manufacturers, the real asset is not the factory but what people know: CAD models, tolerance chains, recipes, costing rules, test limits. AI tools in the browser make it easy for that knowledge to end up in someone else's system, one pasted quote or uploaded bill of materials at a time. This overview explains the legal logic behind trade secret protection and how it should drive your choice of AI operating model. It is not legal advice.

What counts as a trade secret, and why do AI tools matter?

A trade secret is information that is not generally known, has commercial value because it is secret, and has been subject to reasonable steps to keep it secret. AI tools matter mostly for the third condition: if your know-how flows into external tools without any rules, it becomes harder to show you protected it.

The EU definition sits in Article 2(1) of Directive (EU) 2016/943. The information must be secret, meaning not generally known among or readily accessible to people in the relevant circles; it must have commercial value because it is secret; and it must have been "subject to reasonable steps under the circumstances" by the person lawfully in control of it.

Germany transposed the directive with the Trade Secrets Act (GeschGehG), in force since 26 April 2019. Its definition in section 2 adds a requirement of a legitimate interest in secrecy. In the United States, the Defend Trade Secrets Act takes a similar line: under 18 U.S.C. § 1839(3), information only qualifies if the owner "has taken reasonable measures to keep such information secret".

Typical trade secrets in mid-sized industrial companies:

  • Design and production data: CAD models, process parameters, machine settings
  • Recipes and mixing ratios in chemicals, food or plastics processing
  • Costing: overhead rates, purchasing terms, margins per customer
  • Quality know-how: test limits, inspection routines, failure patterns from complaints
  • Sales knowledge: customer lists with terms, bidding strategies

The point for AI decisions is simple. The law does not protect confidential information by default. It protects information whose owner visibly protects it.

How does trade secret protection work in practice?

Protection works through prohibited acts and remedies. Anyone who acquires, uses or discloses a trade secret unlawfully can face injunctions, damages and, in serious cases, criminal penalties. Every one of these remedies depends on the information qualifying as a trade secret in the first place, which includes the reasonable-steps test.

The German act, which transposes the directive, shows the typical building blocks:

  • Lawful acquisition. Independent discovery and reverse engineering of a product that is publicly available or lawfully held are generally allowed, unless a contractual duty says otherwise.
  • Unlawful acquisition, use and disclosure. Copying documents or files without consent, breaching a confidentiality agreement, or using a secret received from someone who obtained it unlawfully.
  • Civil remedies. Injunctions, destruction or recall of infringing goods, information claims and damages.
  • Criminal law. In Germany, section 23 GeschGehG makes certain violations, for example to benefit a competitor or to harm the owner, punishable by up to three years in prison or a fine.

For AI use this creates a direct chain. If you ever need to enforce your rights against a competitor, a former employee or a supplier, you will need to show reasonable measures. If the same know-how is routinely pasted into consumer AI accounts with training enabled and no policy in place, the other side gets a ready-made argument that your measures were not reasonable.

Neither the directive nor the German act lists specific measures. They ask what is reasonable "under the circumstances". Useful reference points are the value of the information, the size of the company, the security standards in your industry and the effort a measure requires. That is exactly where the choice of AI operating model comes in: the more valuable the knowledge, the more control over its processing you can justify.

Which AI operating models exist, and what do they mean for secrecy?

There are four main operating models: consumer tools, vendor business plans, API access with extra data controls, and self-hosted models. They differ in who stores the data, whether it may be used for training, and how well you can document your measures later.

Consumer tools (free and individual paid plans). Each user controls their own privacy settings. Anthropic, for example, states that chats from consumer accounts can be used for training if the user allows it in their settings, and that conversations a user rates with feedback are kept for up to five years. The company has no central control. For trade secrets, this model is hard to defend as reasonable without further measures.

Business and enterprise plans. Vendors contractually exclude training on customer data. Anthropic says inputs and outputs from its commercial products are not used for model training unless you explicitly opt in, for instance by sending feedback. You also get central administration, single sign-on and audit logs. This is a solid building block, but the data still sits with the vendor.

API access with data controls. Companies building AI into their own applications use the API. According to OpenAI's documentation, API data has not been used for training by default since 1 March 2023, and abuse-monitoring logs are kept for up to 30 days by default. Zero data retention and processing in the European region require prior approval from OpenAI. The key advantage is that your application decides exactly which fields leave your systems.

Self-hosted models. Open-weight models run on your own hardware, in your data centre or on a server-class workstation. Data, logs and permissions stay entirely under your control. For your most valuable data class this is the strongest argument under the reasonable-steps test, at the cost of operational effort and usually some model quality. When this trade-off pays off is covered on the page about sovereign AI in-house; the guide to hardware sizing for local LLMs covers which server and GPU class you need.

Most companies end up combining models: a business plan for general writing, an API application for well-defined processes, and a self-hosted model for design and costing knowledge.

What do you need in place before AI tools touch your know-how?

You need three things: a list of which information actually counts as a trade secret, a mapping of that information to data classes, and named owners. Without these, you can neither choose the right tool nor prove later that your measures were reasonable.

Data. Build an inventory of sensitive information. Where do CAD files, recipes and costing sheets live: PDM, ERP, file servers, SharePoint? Who can access them? Many companies discover that access rights have grown over years and are already wider than necessary, AI or not.

Data classes. Three or four levels usually work:

ClassExamplesPermitted AI operating model
PublicProduct catalogue, press releasesany approved tool
InternalMeeting notes, generic process descriptionsbusiness plan with no training
ConfidentialCustomer terms, quote costingbusiness plan or API with data controls, named users only
Strictly confidentialRecipes, design data, process parametersself-hosted model only, or no AI

Systems. Find out which tools are already in use, approved or not. Browser extensions, AI features in office suites and assistants built into CAD or ERP software all count.

Organisation. Decide who approves tools, who assigns data classes and who handles breaches. If you have a works council (common in Germany and other EU countries) or a similar employee body, involve it early, since AI tools that log usage can touch co-determination rights.

How do you set this up step by step?

Work through five steps: identify secrets, define data classes, assign an operating model to each class, put rules, technology and contracts in place, and document everything. Documentation is not a formality. If there is ever a dispute, you will have to show that your measures were reasonable.

  1. Identify secrets. Walk through processes rather than folders: quoting, engineering, production, inspection, service. For each, ask which information would help a competitor most.
  2. Define data classes. Assign every identified item to a class. Keep the number of classes small so people can apply them without a lookup table.
  3. Assign an operating model per class. The higher the value and sensitivity, the more control over storage and access you need. Write down the reasoning: why a business plan is enough for "internal", why "strictly confidential" stays on your own hardware.
  4. Implement rules, technology and contracts.
    • An AI policy with an allow-list of tools and concrete examples per data class
    • Confidentiality clauses in employment and supplier contracts that explicitly cover AI tools
    • Network blocks for unapproved AI services where proportionate
    • Permissions in AI applications that mirror those in the source systems
    • A review of vendor terms: training exclusion, retention, storage location, sub-processors
  5. Train and document. Short, recurring sessions with real examples from your own processes work better than a one-off memo. Record who was trained, when, and which version of the policy applied. In the EU this training can be combined with the AI literacy duty under Article 4 of the AI Act; that duty covers competent AI use in general, while the focus here is protecting your know-how.

The same five steps form the backbone of any AI programme that goes beyond individual tools. The AI governance framework guide covers roles, approvals and policies in general; this article adds the legal reason why trade secret protection should shape your operating model. How to turn them into an ongoing structure with clear ownership is the subject of the page on a fractional head of AI.

What are the limits and risks?

The main limit: no operating model protects know-how on its own. A self-hosted model does not help if everyone can read everything, or if outputs leave the company by email without control. Equally, a ban without a usable alternative tends to push people towards personal accounts.

Other risks to keep in mind:

  • Shadow AI. If the approved tool feels clumsy, people switch to private accounts. That use is invisible to the company and undermines the measures you documented.
  • "Reasonable" moves over time. What counts as reasonable today can change with new tools, new industry practice or a rise in the value of the information. Policies need regular review.
  • Vendor terms change. Privacy pages and terms of service are updated frequently; the pages cited here reflect October 2026. Re-check before relying on any promise.
  • Outputs are data too. An AI-generated quote or a summary of an inspection report often carries the same secrets as its source and needs the same classification.
  • Export control and customer contracts. Beyond trade secret law, export rules and non-disclosure agreements with customers can restrict cloud processing further, especially for technical data from customer projects.
  • Self-hosting is not automatic protection. Your own models need patching, access control, logging and backups. A neglected server in a back room is not a reasonable measure.
  • No guarantee in court. Whether measures were reasonable is decided case by case. Good documentation strengthens your position but does not replace legal review.

How do you get started?

Start with a two-week stocktake: which AI tools are already in use, which three types of information are your most valuable secrets, and where do they live today? From that, derive a first policy with data classes and an allow-list, and offer a usable tool for each class.

A pragmatic first month:

  • Week 1: A short survey across departments on AI tools in use and typical tasks, plus a list of blockers such as "engineering is not allowed to upload anything".
  • Week 2: Agree data classes and operating models, adopt a first version of the policy, roll out a business plan for general tasks.
  • Next: For the "strictly confidential" class, check whether a self-hosted model justifies the effort. Often one well-defined use case is enough to start, such as searching inspection reports or design guidelines.
  • Ongoing: Training, an annual review of the policy and vendor terms, and a log of every change.

If you want leadership and department heads to make these decisions together, a half-day workshop is a good format: data classes, operating models and first use cases fit on one shared whiteboard. For the question of when your own in-house models pay off, the page on sovereign AI is the next step.

Note: this article is for orientation only and is not legal advice. For your specific situation, especially contract clauses and employment questions, consult a qualified lawyer.

Frequently asked questions

Does pasting a trade secret into ChatGPT destroy its protection?

Not automatically. What matters is whether the information is still secret afterwards and whether you had reasonable measures in place beforehand; with no rules or controls at all, proving that in court becomes much harder.

Is a business plan from an AI vendor enough as a secrecy measure?

It is an important building block, because vendors contractually exclude training on your data there. It only becomes reasonable in combination with data classes, access rights, confidentiality obligations and staff training.

When does a self-hosted model make sense from a trade secret perspective?

When your most valuable information, such as design data, formulas or costing models, needs to be processed regularly and you want full control over storage, logs and access.

Do the same rules apply outside Germany?

All EU member states implement the same directive with its reasonable-steps test, and US federal law requires reasonable measures as well. Case law and remedies differ, so check each jurisdiction where your secrets are created or used.

Is this legal advice?

No. It explains the logic of the law as orientation; involve a lawyer specialising in IP or employment law for contracts, policies and disputes.

Sources

  1. Directive (EU) 2016/943, Article 2 – Definitions legislation.gov.uk
  2. 18 U.S. Code § 1839 – Definitions (Defend Trade Secrets Act) law.cornell.edu
  3. German Trade Secrets Act (GeschGehG) – full text gesetze-im-internet.de
  4. Section 2 GeschGehG – Definitions gesetze-im-internet.de
  5. Section 23 GeschGehG – Criminal offence gesetze-im-internet.de
  6. OpenAI API Docs – Data controls in the OpenAI platform developers.openai.com
  7. Anthropic Privacy Center – Is my data used for model training? (commercial products) privacy.claude.com
  8. Anthropic Privacy Center – Is my data used for model training? (consumer products) privacy.claude.com