Today
Camera checks, learning controllers and assistants are moving into machines and devices, often as a bought-in module. Whether a function makes the product high-risk under the AI Act is rarely asked before delivery. When a learning function takes on a safety task, risk management, technical documentation, logging and an assessment by a notified body often come on top.
An example
Conveyra Fördertechnik reviews its range before the next model year. The agent finds five AI functions, among them a camera with self-learning person detection that stops the BK 10 t bridge crane as soon as someone is in its travel path. It classifies this one as high-risk because, as a self-learning safety component, it must be assessed by a notified body under the Machinery Regulation, and lists risk management, logging and the evidence for the assessment as open. For the service chat in the app it names the duty to disclose, and three functions such as the failure forecast on the hoist carry no specific duties. Product safety reviews the classification with the camera's supplier.
Your data, wherever it lives
Which function uses machine learning is scattered today: in specifications and BOMs in the PLM, in tickets and wiki pages of software development, in suppliers' data sheets as PDFs and in spreadsheets on the drive. demi reads, understands and links them, with your rights. The AI finds every learning function in the portfolio in days instead of months and justifies its classification, and your experts decide.
Where the data lives today
- PLM
- Tickets
- Wiki & intranet
- PDFs
- Office files
- DMS & archive
- Folders & drives
How demi solves it
Three layers in this order: connect the data first, then the rights, then the AI.
- 1 · ConnectData Layer
Data sources in this case: PLM, software development, DMS and the suppliers' documents. demi replaces nothing and links, per product and variant, the functions, software versions, specifications, risk assessments and the suppliers' information on their modules. demi writes back only after approval.
- 2 · PermitPermission Layer
Development documents and supplier information are visible only to development, product safety and legal. Every classification is confirmed by a responsible person you appoint, and each decision is recorded with its reasoning.
- 3 · ActIntelligence Layer
The agent recognises AI functions from descriptions, software versions and data sheets and justifies each classification with the article of the regulation and the place in the document. Unclear cases are marked as open instead of being classified; the legal assessment stays with your experts.
How it runs
- 1
The agent searches BOMs, software versions and specifications for every function that uses machine learning, including bought-in parts.
- 2
For each function it collects purpose, location, supplier and role for safety.
- 3
It classifies each one under the AI Act: prohibited, high-risk, transparency duty or no specific duties, always with reasons.
- 4
For every classification it lists duties, missing evidence and deadlines.
- 5
Your experts review the classification and product owners release it.
What changes
Every AI function in the portfolio known and classified
Missing evidence visible early
Fewer surprises in the conformity assessment
What does it bring you?
Check it with your own numbers
Every starting value is an assumption you can change. The calculation sits right below, step by step.
Your numbers
Result
+58 hfreed per year
Assumptions and calculation
demi finds every AI function in your products and justifies its classification; you decide.
4 h → 2.8 h per case
How we calculate
4 AI functions × 4 h × 30% × 1.0 (manual) × 12 months = 58 h
Value per year, before demi costs
Freed per year+58 h
Assumption: labour cost of €49.50 per hour in German manufacturing in 2025 (Destatis), plus workplace, IT and overheads.
= Value per year, before demi costs
+€3,770
gained per year
An estimate, not a promise. The starting values are assumptions for an industrial company with 150 to 500 people.
Combine with more cases and get the packageQuestions about this use case
What counts as an AI function under the AI Act?
A system that infers from its inputs how to respond, such as image analysis with a trained model. Fixed, programmed rules and classic control engineering usually do not count. The agent shows how it recognised a function, and your experts decide the borderline cases.
When is an AI function high-risk?
Mainly when it performs a safety function in a product that a notified body must assess anyway, for example under the Machinery or Medical Device Regulation, or when it works in one of the areas the regulation lists, such as critical infrastructure. The agent gives the reason with the article for every function.
What if the AI sits in a bought-in part?
Then the agent shows which information and evidence the supplier must provide and what is missing, and drafts the request to them. You remain responsible for the whole product as the manufacturer, so the module is part of the classification.
Does every software update mean a new classification?
Only when the purpose, role or learning behaviour of a function changes. The agent compares each new software version with the classified one and flags where a new assessment is needed.
