Blog · Published · Updated
AI in manufacturing and trade secrets: what a local model does not solve
Short answer
A model on the plant's server is not enough to keep a formula a trade secret. Under Article 11 of Poland's Act on Combating Unfair Competition, the steps the company takes to protect the information count too. So an AI assistant may only retrieve documents the person asking is allowed to see. With 500 operator questions a month, a €3,000 local app pays back in 1.7 months.
How we workDetails ↓
- 01Call30 minutes with an engineer, free
- 02PrototypeA working system on your process and data, free
- 03ImplementationTarget and fixed price in the contract; 50% at signing
- 04Acceptance50% once the agreed target is met
- model on your server or in the cloud in the EU
- quoted after the review
- same work in application with EU cloud model
- from €3,000
- task, data and equipment analysis
- €0
We belong to the OpenAI and Anthropic partner programmes and learn through their official courses. What this means ↓
Author: Syntalith · 3 min read
Statistics
Company data and AI: statistics 2026
1.7 mo.
payback of a €3,000 local app at 500 operator questions a month, 8 minutes of searching for instructions each, with a GPU server at the plant. Scenario.
48.8%
According to Eurostat (2025), 48.8% of EU companies that considered AI and did not use it named concerns about data protection and privacy.
Source: Eurostat, use of artificial intelligence in enterprises (2025) ↗
52.5%
According to Eurostat (2025), 52.5% of EU companies that considered AI and did not use it named unclear legal consequences as a reason.
Source: Eurostat, use of artificial intelligence in enterprises (2025) ↗
84%
According to the European Commission (2025), 84% of Europeans think AI at work needs careful management to protect privacy and ensure transparency.
Source: European Commission, Special Eurobarometer 554 (2025) ↗
17.3%
According to Eurostat (2025), 17.3% of EU manufacturing companies with at least 10 employees used AI.
Source: Eurostat, artificial intelligence by NACE activity (2025) ↗
Figures checked: 22 October 2026
What the legal definition of a trade secret requires
In Poland, Article 11(2) of the Act on Combating Unfair Competition covers technical, technological, organisational and other information with economic value that is not generally known or easily accessible to people in the industry. The third condition concerns the company itself: the entitled person must take due care to keep the information confidential. The provision also covers a particular combination of information: single elements may be known, and the value lies in putting them together.
A „confidential” label, a signed agreement or the location of the server does not meet all three conditions. Rolling out an AI assistant becomes part of those steps, so its design has legal weight. Assessing a specific incident is a lawyer's job, and this article describes the technical preparation.
A formula does not stay in its folder
At the plant, the process engineer sees the full formula, and the operator sees the instructions for their own stage. The assistant should help the operator find the current instructions. If it searches the whole collection, it can use content the operator has no access to in its answer. A model running locally still takes part in that badly designed flow.
| Place | What to settle with the data owner and the administrator |
|---|---|
| Reading PDFs and photos of documents | whether the file goes to a local tool or an external service |
| Search | whether document permissions limit the passages retrieved |
| Answer history | who sees the saved conversation and attachments |
| Diagnostic logs | whether an error records the full text of a document or question |
| Backups and service | who can restore the materials and what the service provider receives |
Hiding the button that opens the source is not enough if the content of a closed document has already reached the answer. Permissions must apply before the model, as the article on permission-aware RAG describes.
An access test before the rollout
- 01Preparationa fictional formula, two roles, the expected result written down before the test
- 02Direct questionan operator asks for the name of a formula they have no access to
- 03Access removedthe same question in an existing conversation after the role changes
- 04Document withdrawnwhat happens to the index, the history and the copies
- 05Recordthe run and the exceptions found stay as the acceptance document
A change in folder permissions does not prove that every copy changed: the search index, the conversation history and the logs have their own rules. So acceptance covers the failed attempts next to the correct answer for an authorised user.
What a local AI documentation assistant costs and when it pays back
- Today
- 500 questions × 8 min = 66.7 h (€2,327)
- With the system, a person checks the output in 15% of the time
- €349
- Model, 1 call per case, Claude Sonnet 5.5
- €8
- Cost of one case
- €4.65 today, €0.71 with the system
- Saving
- €1,970 a month
Payback of €3,000
1.5 months
The volume, the time and the checking share are the scenario's assumptions. €34.90 an hour is the EU average hourly labour cost for 2025 according to Eurostat. Model: 6,000 input tokens and 500 output tokens per call at Anthropic's rates, converted at the ECB rate of 1.1205 USD per euro on 22 October 2026.
In the local version, a GPU server replaces the API model, for example the Hetzner GEX45 at €214 a month, or hardware at the plant. With the checking at €349, the local version costs €563 a month and the payback grows to 1.7 months. That is the price of keeping formulas in your own infrastructure, which you compare with an assessment of an external service and its terms.
Who is responsible for what
The information owner and a lawyer
what needs protection and which use is allowed
a decision
The administrator
roles, copies, logs and service access
configuration
The contractor
data flow, permissions before the model, an access test
acceptance terms
We prepare the local model and its evaluation as part of private LLMs, and the register of tools and the rules for working with data in an AI Act and GDPR audit from €700 net. For the free process review, a description of the types of documents and user roles is enough, with no formulas and no process parameters.
Worked example · Scenario
How we count the payback: a documentation assistant in manufacturing
We run the same count for your documents in the free process review.
€4,653
a month for finding answers to 800 questions in documents
Payback: 0.8 mo.
Our assumptions are 10 minutes to find the right instruction for the product and version, and checking in 15% of the old time. €34.90 an hour is the EU average hourly labour cost for 2025 according to Eurostat, for companies with 10 or more employees. Model cost: Claude Sonnet 5.5, about 6,000 + 500 tokens per answer from documents, at the ECB rate of 1.1205 USD per euro on 22 October 2026.
See the full calculation →
