AI and the Works Council: a legally sound rollout.
How to introduce innovation in a legally sound way, without losing the works council or compromising compliance.
30 minutes, online, free and non-binding.
AI projects at SMEs rarely fail because of the technology, they fail due to concerns about surveillance, data protection, and regulation. The solution is sovereignty through architecture: anonymization before the AI, operation on your own hardware, and rules agreed with the works council that are technically enforced. That rules out misuse instead of merely promising it won't happen.
Rolling out AI without losing the works council
It's the classic dilemma of German SMEs: you know you need to automate to stay competitive. But the moment the word "AI" comes up, the alarm bells start ringing across the company.
The works council fears surveillance, legal fears data protection violations, and employees fear for their jobs. We call this state the "valley of death." On one side stands the technological vision, on the other the reality of German regulation (Works Constitution Act, GDPR, EU AI Act). Many projects die here, before they've even begun.
Our experience from SME projects: the problem is almost never the technology. It's the fears, legitimate fears, that leadership often underestimates. An impressive demo in front of the board doesn't solve a single compliance problem.
Why so many AI projects at SMEs fail
More and more SMEs are launching AI initiatives, yet many never get past the pilot phase. The reasons keep repeating:
- Regulatory uncertainty: the EU AI Act is in force, but many companies don't know which risk class their planned application falls into.
- Works council pushback: under Section 87(1) No. 6 of the Works Constitution Act, the works council has a co-determination right over technical systems capable of monitoring employees.
- Data protection concerns: the fear that personal data ends up on US servers triggers a reflexive "no" from the legal department.
- Missing strategy: many companies start with the tool instead of the problem.
We don't solve legal problems with lawyers, we solve them with technology.
From the employee representatives' point of view, an AI system is a black box. That's why systems need to be built so that misuse is technically ruled out, not through promises or contracts, but through an architecture that leaves no other option.
Sovereignty through architecture: three security layers
1. The digital gatekeeper (privacy by proxy)
Before the AI reads a question, it's anonymized. Names, addresses, and phone numbers are removed. The gatekeeper works on rules, not AI, and can be inspected by the works council at any time.
2. The data vault (on-premise)
The AI runs on hardware physically located at your premises, with no outside connection. A "digital shredder" clears short-term memory after every prompt, so no conversation content remains on the server.
3. Ethical safeguards (compliance rules)
We define rules such as: "The AI may not write performance reviews of employees." The system blocks such requests technically. Every rule is agreed jointly with the works council and cannot be disabled in the system.
Example scenario: the blocked law firm
A commercial law firm wants to automate case file analysis. Clients send hundreds of pages of contracts that an associate works through over days. An AI could extract the key points. But the project is stalled because the works council has two concerns:
- "The AI measures which lawyer reads the slowest": the fear of covert performance monitoring.
- "Client data ends up on a server in the US": the fear of the CLOUD Act.
Both concerns are legitimate. Many cloud AI solutions log which user made which request. The solution isn't to talk the works council into it, it's to rule out the risks technically:
- The AI runs on a dedicated server in the firm's own server room, with no cloud and no internet.
- The digital gatekeeper removes client names and file numbers before analysis.
- Usage statistics are only captured in aggregate. No one can see who made which request, or when.
On this basis, a works agreement can be negotiated.
In a free initial call, we will look together at what matters most for you.
Example scenario: the skeptical manufacturing plant
An automotive supplier wants to introduce visual AI inspection that detects surface defects on metal parts. The plant manager halts the project with one question: "Can we be sure the AI won't one day decide whether an employee is working well enough?"
The question is legitimate. Vision systems could, in theory, also capture work speed, break behavior, and movement patterns. The answer is a limitation built into the hardware: the camera covers only the inspection zone at the end of the conveyor belt, and is mounted so it can't be redirected without breaking a seal. The system has no access to the company network, stores no images, and only passes pass/fail signals to the PLC.
What's convincing isn't the promise that the AI only inspects parts, it's that it's visibly incapable of anything else.
The EU AI Act: what SMEs need to know now
The prohibitions and the AI literacy obligation have applied since February 2, 2025, and the transparency obligations since August 2, 2026. With the Digital Omnibus (Regulation (EU) 2026/1744), the EU has pushed back the high-risk obligations: to December 2, 2027 for applications under Annex III, such as in HR, and to August 2, 2028 for AI in regulated products. For SMEs, that means:
- Risk classification: every AI system must be assigned a risk class. Many business applications fall into the "limited risk" category, but the boundaries are fluid.
- Transparency obligation: people must be able to recognize that they're interacting with an AI, for example a chatbot, and AI-generated content must be labeled (Art. 50 of the EU AI Act).
- High risk in HR: AI for candidate selection, promotion, or performance evaluation counts as high-risk AI. Employers then must, among other things, ensure human oversight and inform employee representatives and employees in advance (Art. 26 of the EU AI Act). With the right architecture, this is manageable from the start.
- AI literacy: companies must take measures to ensure their staff use AI competently (Art. 4 of the EU AI Act).
The most common mistake: reading the delay as an all-clear. Prohibitions, AI literacy, and transparency already apply today, and whoever builds the right architecture now won't have to scramble to catch up in 2027. As of October 29, 2026, not legal advice.
Example scenario: the healthcare provider
An outpatient care service with multiple locations plans its routes in Excel. When staff call in sick, mornings turn into improvisation. An AI-assisted route planning tool is meant to handle routes, fair distribution, and rescheduling when staff are out.
The problem: the AI needs access to sensitive data such as patient addresses, care levels, and medical specifics. These fall under the GDPR and the duty of confidentiality under Section 35 of the German Social Code, Book I.
The solution is a two-stage system: the AI works only with pseudonymized records. Instead of "Mrs. Schmidt, Hauptstraße 15, care level 3", it sees "Patient 4711, Zone North-7, care tier C". The re-identification happens only on the dispatcher's local system. That way, the data protection officer can verify data security directly.
What does this get you?
- Relief on recurring analysis tasks such as file review, quality control, and document processing.
- A structured path to a works agreement instead of months of principled debate.
- Possible funding: see our overview of Digitalization Funding Programs 2026.
- Data protection built in from the start: privacy by design instead of retrofitted safeguards.
The perhaps most important value can't be measured in euros: sovereignty. The certainty that your data belongs to you, and that your employees see AI as a tool, not a threat.
The example scenarios describe typical situations, not named customers. Legal statements are not a substitute for case-specific legal advice.
Before you reach out
Does the AI really have to run on-premise?
Not necessarily. We offer both on-premise and managed cloud. What matters is that data sovereignty stays with you. For sensitive industries like law firms, healthcare, and finance, we recommend on-premise.
How long does implementation take?
A typical pilot project takes 4 to 6 weeks. That covers requirements analysis, technical integration, coordination with the works council, and training your employees.
What does an AI solution for SMEs cost?
Costs depend on scope and solution, and we estimate the effort after the initial consultation. Our overview of the 2026 funding programs shows which options fit.
Is the AI GDPR-compliant?
The architecture follows the privacy-by-design principle: the digital gatekeeper anonymizes personal data before it reaches the AI model. The specific assessment is done case by case, together with your data protection team.
Ready for the next step?
Talk to us about how to roll out AI together with your works council and data protection team. We'll get back to you within 24 hours with a proposal for the initial consultation.

