AI Policy for Companies: What Employees Can and Cannot Do

Open ChatGPT, paste in a piece of text, hit enter—done. Using artificial intelligence at work has never been this easy.

And that's exactly where one of the biggest problems for companies now lies.

What information are employees actually allowed to enter into ChatGPT, Claude, Gemini, Copilot, or other AI systems? Can an employee have a customer contract summarized? Can accounting analyze invoice data with AI? Can a developer send internal source code to an AI tool? And who's liable if the AI gives a wrong answer?

Companies that don't answer these questions still end up with AI in the workplace. It just happens without any control.

An AI usage policy for employees isn't meant to prevent AI use. It's meant to define how artificial intelligence can be used safely, transparently, and productively.

Since February 2, 2025, the EU AI Act's AI literacy provisions have applied. Since August 2, 2026, further core parts of the AI Act have become applicable, and European and national oversight has gained significant weight. The AI Act was most recently amended by the Digital Omnibus, which entered into force in July 2026. The basic obligations for small businesses are covered in EU AI Act for Small Businesses: What SMEs Need to Know (Read article)

For companies, now is the right time to turn a vague "just use ChatGPT, but be careful" into a clear company rule.

Why every company needs an AI policy

In many companies, the actual way AI gets used looks something like this:

One employee uses their personal ChatGPT account. Another has Gemini open. The development team runs Claude. Microsoft Copilot may already be built into existing software. Someone pastes a customer email into an AI tool while a colleague has an AI analyze a spreadsheet.

Management often only knows part of this.

This phenomenon is often called Shadow AI: employees using AI tools outside a centrally governed and controlled company process. How widespread Shadow AI has become, and how companies can respond, is covered in Shadow AI at Work: When Employees Secretly Use ChatGPT (Read article)

The problem isn't the AI itself. The problem is missing governance.

A good AI policy answers at least five questions:

Which AI systems may be used? Which data may be processed? For which tasks is AI allowed? When must a human review the result? And who decides on new AI applications?

That clarity protects companies and employees alike.

The core principle: AI may assist—responsibility stays with people

A sensible company policy doesn't need an 80-page rulebook.

One central principle is enough as a starting point:

AI systems may support employees in their work. Decisions, approvals, and responsibility for work results remain with the accountable person.

That means, for example:

An employee may have AI draft a first version of a customer email. They may not send it unreviewed.

An administrator may have AI generate a PowerShell or Bash script. They should still understand it, review it, and—depending on impact—run it in a test environment first.

An employee may get AI help with a presentation. But invented sources, wrong figures, or hallucinations must not make it into the presentation unchecked.

This one principle already eliminates a large share of typical risks.

The traffic-light model for AI at work

For employees, an AI policy is easiest to follow when not every situation is described in legal language.

A traffic-light model works much better in practice.

Category	Meaning	Examples
🟢 GREEN	Use generally allowed	Brainstorming, general text, translating public content, phrasing help, summarizing publicly available information
🟡 YELLOW	Only with an approved AI system and defined rules	Customer communication, internal documents, source code, business analysis, draft contracts, personal information
🔴 RED	Prohibited without explicit approval	Passwords, API keys, credentials, highly sensitive information, uncontrolled HR data, confidential trade secrets, autonomous critical decisions

This model should be part of every company's AI policy. Employees need to be able to judge, within seconds, whether a given use is allowed.

What employees can normally do with AI

AI use is generally low-risk when no confidential or personal information is shared and the AI is simply used as a supporting tool.

Typical examples include brainstorming, structuring your own thoughts, drafting general phrasing suggestions, improving spelling and style, translating non-confidential text, building outlines, general research prep, or explaining publicly known topics.

Technical questions are generally fine too. An administrator might ask, for example:

"Explain the difference between systemctl restart and systemctl reload."

It becomes a problem once entire confidential configuration files, internal hostnames, passwords, certificates, or customer data become part of the prompt.

So the key question isn't "Am I allowed to use ChatGPT?" It's "What information am I sending to which AI system?"

When employees need to be especially careful

Many business AI use cases fall somewhere between clearly allowed and clearly forbidden.

A good example is a customer email. The request "Write me a friendly reply to a customer complaining about a late delivery" is fine on its own.

Pasting in the entire real customer email—including name, address, customer number, phone number, order details, and other personal information—is a completely different kind of processing.

Germany's data protection conference (DSK) explicitly recommends that companies consider privacy questions already at the stage of selecting, implementing, and using an AI system.

That's why an AI policy shouldn't classify personal information as simply "allowed" or "forbidden" across the board. The better rule is:

Personal or confidential data may only be processed with AI systems explicitly approved for that purpose, and in line with the defined privacy and access framework.

That still leaves room for privacy-compliant enterprise AI solutions.

Information that should never go into public AI systems unchecked

One of the most important rules in any AI policy concerns secrets.

Employees should not send information to unapproved AI services if its disclosure or loss could harm the company, its customers, or other people.

That especially includes credentials, passwords, API keys, private keys, authentication tokens, unpublished financial information, confidential customer data, internal security information, and protected trade secrets.

This applies not only to the chat itself but also to features like ChatGPT Memory: information an employee shares in a chat can, in some cases, be stored as a lasting memory that persists beyond that single conversation. For what ChatGPT Memory actually remembers, how it differs from chat history, and how to delete memories in a controlled way, see ChatGPT Memory Explained: What It Remembers – and What It Doesn't (Read article).

Source code is especially risky when it contains credentials or details about internal systems. A developer can absolutely use AI for coding. But:

DATABASE_PASSWORD=SuperSecret123
API_KEY=abc123xyz
INTERNAL_SERVER=10.10.50.12

does not belong in a prompt to any public AI service. This is exactly where the company policy should be unambiguous.

Personal ChatGPT account or company account?

This question absolutely belongs in a modern AI policy too.

For business tasks, employees should generally use the AI systems provided and vetted by the company.

A personal account largely escapes company oversight. IT may not be able to reliably manage central settings, access rights, contract terms, retention rules, or offboarding on that account.

So the following rule makes sense: business data may only be processed through AI services and accounts approved by the company for that purpose.

This doesn't prevent AI use. It simply moves it into a controllable framework.

Can internal documents be summarized by AI?

The answer depends on the document and the system used.

A public user manual is different from an internal revenue plan. A public whitepaper differs from an employment contract. And an approved enterprise AI system with proper privacy, contractual, and security controls differs from any random free web service.

Companies should classify documents. A proven scheme is: Public → Internal → Confidential → Strictly confidential.

From there, define which protection class may be processed with which AI system. That rule is much easier for employees to grasp than abstract privacy language.

AI and personal data

"No personal data in AI" sounds like a simple rule. For many companies, though, it would be nearly unworkable.

CRM systems, support tools, Microsoft 365 features, HR applications, or internal AI assistants sometimes work with exactly that kind of information.

The right question is: is this AI system approved for this specific processing activity?

Companies should assess purpose, legal basis, data minimization, access controls, data processing agreements, transfers, deletion, and technical safeguards.

Germany's data protection conference now provides extensive guidance on privacy-compliant AI use.

For employees, that should translate into one simple rule: only process personal data in AI systems the company has explicitly approved for that purpose.

AI must never receive passwords or secrets

This rule should apply without exception to ordinary AI chats.

Passwords, private keys, recovery codes, session cookies, access tokens, and API keys don't belong in an AI prompt.

For technical troubleshooting, anonymize instead. Turn:

ssh admin@internal-company-server.example
Password: MyRealPassword

into, for example:

ssh user@example-server
Password: [REDACTED]

The same principle applies to logs. Logs can contain email addresses, IP addresses, session IDs, tokens, user identifiers, or other confidential information.

Before handing anything to an AI system, such content should be cleaned up or routed through an approved internal process.

Can AI see source code?

A blanket ban would be unrealistic in most development teams today.

Coding assistants and AI agents can help developers considerably. What matters is the framework around them.

Companies need to define which repositories may be processed, which AI system is approved, whether secrets might be present, what write access an agent gets, and which actions require human approval.

AI agents in particular change the risk profile. A classic chatbot returns text. An agent might edit files, run shell commands, modify Git repositories, call APIs, or access databases.

That turns a privacy question into a classic access-control and IT security question as well.

Prompt injection is a relevant attack scenario here: manipulated content can try to alter an AI agent's behavior or trigger unwanted tool calls. This topic is covered in more depth in Prompt Injection Explained: How Attackers Hijack AI Agents (Read article)

Never blindly run AI-generated scripts

Administrators in particular benefit enormously from AI. ChatGPT and similar tools can produce PowerShell, Bash, SQL, Python, Dockerfiles, Kubernetes manifests, and configuration files.

But a convincing-looking command isn't automatically a correct one.

That's why the AI policy for technical staff should state: production changes still require the same review, test, approval, and rollback process as changes written by humans.

AI must not create a "shortcut to production." An AI-generated rm -rf /path/* is no less dangerous just because a language model suggested it.

For administrators and developers, the rule should be: AI → review → test → approval → production. Not: AI → production.

Employees must verify AI results

Generative AI can produce wrong information very convincingly. That applies to facts as well as sources, legal information, commands, configurations, calculations, or supposed product features.

The company policy should define when human review is mandatory.

That matters especially for external customer communication, contracts, quotes, technical changes, publicly published information, HR matters, and financial decisions.

The greater the potential damage from a mistake, the stricter the review should be. A four-eyes principle is a good fit for critical tasks.

Can AI make decisions about employees?

This is where things get especially sensitive.

The EU AI Act treats certain systems in employment, HR management, candidate selection, and access to self-employment as high-risk applications. That can include AI systems for sorting and scoring job applications. The corresponding high-risk rules for certain Annex III systems were pushed back to December 2027 by the most recent amendments.

Companies should clearly distinguish between "Help me write a neutral job posting" and "Automatically evaluate every applicant and decide who we hire." Those are completely different applications from a regulatory and organizational standpoint.

Emotion recognition at work is an especially sensitive area

The AI Act is even more explicit about AI-based emotion recognition.

AI systems for emotion recognition in the workplace generally fall under the AI Act's prohibited practices, aside from narrowly limited medical or safety-related exceptions. These prohibitions have applied since February 2025.

A company shouldn't simply deploy an AI system that claims, based on voice, face, or behavior, to detect whether employees are motivated, stressed, bored, or dissatisfied.

This isn't just a productivity feature anymore. It touches fundamental rights and an explicitly regulated area.

Don't forget the works council

If a company has a works council, its involvement in introducing AI systems can be legally relevant.

Under German law (§ 87(1) no. 6 of the Works Council Constitution Act), the works council has co-determination rights over introducing and using technical systems designed to monitor employee behavior or performance.

The AI Act also sets information duties toward employees and worker representatives for certain high-risk AI systems at work.

That's why an AI policy shouldn't be developed in isolation from HR, privacy, information security, and worker representation.

AI Act: employees need AI literacy

A frequently overlooked requirement is Article 4 of the EU AI Act.

The obligation to promote AI literacy has applied since February 2, 2025. The current rule requires providers and deployers of AI systems to take measures supporting the AI literacy of people working with these systems on their behalf, taking into account their knowledge, experience, education, and context of use.

That doesn't mean every employee needs the same certificate. The European Commission explicitly clarifies that no single uniform competence level is mandated.

An employee who occasionally has AI draft text needs different skills than an administrator running an AI agent with database, API, or server access.

That's exactly why an AI policy and AI training belong together. More detail is available in AI Literacy Is Required—but Where Is the Mandatory Certificate? (Read article)

Does every AI result need to be labeled?

No.

A blanket rule like "anything made with ChatGPT must be labeled as AI" oversimplifies the AI Act.

The AI Act contains specific transparency duties—for example for certain deepfakes and certain AI-generated or manipulated content. Those transparency rules have been a particularly current topic since August 2026.

Companies should therefore define in their policy when AI use must be documented internally or disclosed externally.

Independent of any legal labeling duty, transparency can still make sense for editorial, contractual, or internal company reasons.

Who is liable for a wrong AI answer?

For employees, one practical rule matters most: "the AI said so" is not an approval process.

If an employee is responsible for checking an invoice, approving a contract, configuring a system, or reviewing a publication, AI cannot simply replace the required professional review.

That's why an AI policy should explicitly assign responsibilities. Whoever uses an AI output professionally needs to know when they may review it themselves and when additional professional or organizational sign-off is required.

Copyright: AI output isn't automatically free to use

For images, text, music, and code, one common misconception should be avoided: just because AI generated something doesn't automatically mean the result can be used freely without further review.

For business publication, potential copyright, trademark, license, personality, and usage rights need to be considered. That especially applies to marketing material, product images, software code, logos, music, and content that recognizably involves real people or protected works.

A company policy should require professional review of AI-generated content before any relevant external or commercial use.

The 10 most important rules of a good AI policy

1. Use only AI systems approved by the company for business tasks.
2. Never enter passwords, API keys, tokens, or other credentials into AI systems.
3. Process personal data only with solutions explicitly approved for that purpose.
4. Process trade secrets and confidential information only according to their protection class.
5. Check AI results for errors and hallucinations before use.
6. Test and review AI-generated code and scripts before deploying to production.
7. Never base critical decisions solely on unverified AI output.
8. Have new AI tools approved by the responsible team before business use.
9. Report suspicious AI output, privacy issues, or security incidents.
10. Train employees regularly, matched to their actual AI use.

That produces a policy employees can actually understand and apply.

Template for a short company AI policy

The following template can serve as a starting point for your own company policy.

1. Purpose

Artificial intelligence may be used to support employees in their daily work, make processes more efficient, and improve the quality of work results.

Responsibility for decisions and for results that are published or used in production remains with the accountable employee or manager.

2. Approved AI systems

Only AI systems and accounts approved by the company may be used for business activities.

New AI services must be reviewed and approved by the responsible team before business use.

3. Confidential information

Passwords, credentials, API keys, authentication information, and other secrets must not be entered into AI systems.

Confidential information and trade secrets may only be processed with systems explicitly approved for that purpose.

4. Personal data

Personal data may only be processed if the AI system and the specific use case have been approved by the company for that purpose.

5. Reviewing results

AI-generated information must be checked for accuracy, completeness, and plausibility before business use.

Additional approval or four-eyes processes can be required for critical decisions.

6. Software and administration

AI-generated code, scripts, database commands, and configuration changes must go through the same test, review, and approval process as changes written manually before being used in production.

7. Automated decisions

AI systems must not make autonomous decisions about applicants, employees, customers, or other people without separate review and approval when those decisions can significantly affect the people involved.

8. Copyright and usage rights

Before external or commercial use of AI-generated content, required rights, licenses, and internal approvals must be considered.

9. Security incidents

Suspicious output, potential data leaks, prompt injection attempts, unintended disclosure of confidential information, and other security incidents must be reported immediately through the defined reporting channel.

10. AI literacy

Employees who use AI systems professionally receive briefing or training appropriate to their role and the systems they use.

Who should be responsible for the AI policy?

Not every company needs a new full-time role for this.

In a small company, responsibility can sit jointly with management, IT, and privacy. Larger organizations can also involve information security, compliance, HR, legal, and worker representation.

More important than the title of the responsible person is a clearly answered question: who is allowed to approve a new AI system for the company?

If no one can answer that, even the best policy won't work.

Combine an AI policy with an AI inventory

Beyond the policy itself, a simple overview of the AI systems actually in use is worthwhile.

It can document things like system, vendor, responsible department, purpose, user group, permitted data types, and approval status.

Such an internal inventory isn't automatically a universal legal registration requirement for ordinary AI use. Organizationally, though, it can help significantly with keeping track of AI use and shadow AI. KI-Buster's guide on the EU AI Act for small businesses also recommends systematically recording and assessing the AI systems in use.

An AI policy shouldn't over-regulate AI

There are two extremes. One is: "Just go for it. Anyone can use any AI." The other: "AI is banned."

Both work poorly long-term. With the first, the company loses control. With the second, usage often just moves to personal accounts, personal phones, or other uncontrolled services.

The better strategy: allow where the risk is low, control where sensitive information is processed, and prohibit where the risk isn't acceptable.

That's the core of modern AI governance.

Conclusion: employees need clear rules, not blanket AI bans

Artificial intelligence is now a fixture of everyday work. The question is no longer whether employees will use ChatGPT, Copilot, Claude, Gemini, or other AI systems. The real question is: under what conditions are they allowed to?

A good AI policy for companies creates a simple framework for that. Employees know which systems they can use. They know which information doesn't belong in a prompt. They understand when results need to be checked. And they know who to ask when a new use case comes up.

Companies, in turn, reduce shadow AI, privacy risk, uncontrolled data leakage, and bad decisions—without giving up the productivity gains of modern AI.

We explain in detail which technical and organizational measures usefully complement such an AI policy in Preventing Shadow AI: Technical and Organizational Measures for Companies (Read article). One particularly effective technical complement is a self-hosted, local AI platform: we show how companies can run Ollama and Open WebUI for this in a privacy-friendly, GDPR-compliant way in Local AI for Companies: Running Ollama and Open WebUI Safely (Read article).

Given the AI literacy framework in place since February 2025, and the AI Act's advancing application and enforcement since August 2026, documented, understandable AI governance is now considerably more useful than a loose "be careful with ChatGPT."

The best AI policy isn't the one with the most bans. It's the one where every employee, after reading it, knows: what am I allowed to do, what am I not, and who do I ask if I'm unsure?

FAQ: AI policy for companies

Does every company need an AI policy?

There is no blanket legal requirement for a standalone document with exactly that name. But once employees use AI professionally, a clear internal policy is organizationally very sensible. It helps define approved systems, privacy, accountability, security rules, and AI literacy in a traceable way.

Can a company ban ChatGPT?

Companies can regulate the use of specific applications within their corporate IT and for business activities. In many cases, a list of approved services and clearly defined use cases works better than a blanket AI ban.

Can I enter customer data into ChatGPT?

Not without restriction. Personal customer data should only be processed with a system the company has explicitly approved for that purpose. Privacy and contractual requirements must be checked beforehand.

Can I use AI to write business emails?

In general, AI can help draft business communication. Confidential and personal information may only be processed according to company rules, and the responsible employee should review the content before sending.

Can ChatGPT review contracts?

AI can help with structuring, summarizing, or a first pass at analysis. An AI output should not replace a qualified legal review or approval without further checking.

Do I have to train employees on AI?

Article 4 of the EU AI Act has required providers and deployers of AI systems, since February 2, 2025, to take measures supporting the AI literacy of people working with these systems on their behalf. The type and scope should match actual use and risk.

Do I need a certificate for every employee?

No. The European Commission does not require one specific uniform certificate or identical competence level for every employee. Measures should match each person's knowledge, experience, and context of use.

Can AI automatically evaluate job applicants?

AI systems for recruitment, candidate selection, and certain HR decisions can fall under the high-risk categories of the AI Act. Such applications therefore need considerably more careful legal and organizational review than a normal generative AI assistant.

Can AI monitor employees based on their emotions?

AI-based emotion recognition in the workplace is generally among the prohibited AI practices under the AI Act. Narrowly limited exceptions exist mainly for medical or safety-related purposes.

Who should create an AI policy?

Ideally, management, IT, privacy, and information security work together—depending on the company, HR, compliance, legal, and the works council should be involved as well.

Sources and currency

Fact-checked: September 1, 2026. Legal and regulatory information draws on the current EU AI Act including the Digital Omnibus amendments, information from the European Commission and the European AI Office, EUR-Lex, Germany's data protection conference (DSK), and the German Works Council Constitution Act.

10 ChatGPT Prompts for IT Support and Helpdesk (Read article)