An AI use policy is a short document, approved by management, that tells your staff which AI tools they may use, with which data, under whose responsibility, and what happens if the rules are broken. This guide gives you the clauses to include, the Swiss and European legal framework to know, a template to copy and adapt, and the indicators to keep the policy alive.
Why put an AI usage policy in place in your company?
Artificial intelligence (AI) is already inside the company, often without management having decided it: one employee pastes a contract into a public assistant, another switches on an AI feature in a piece of software. A written policy replaces these isolated habits ("shadow AI") with rules everyone knows.
What an AI use policy is for
- Regulatory compliance: the revised Swiss Federal Act on Data Protection (FADP, nFADP) and, depending on your activity, the EU GDPR set strict rules on personal data.
- Risk reduction: a clear policy prevents data leaks, unsupervised automated decisions and algorithmic bias.
- Trust: transparent governance reassures employees, clients, partners and investors.
- Consistent tooling: a well-defined policy aligns AI tools with company objectives and stops you buying the same thing several times.
The legal framework at a glance
There is currently no general Swiss law dedicated to AI. The existing rules therefore already apply to AI tools. These are the texts to know before you draft, and they do not replace legal advice for your situation.
| Text or authority | What it changes for your policy |
|---|---|
| FADP (revised, nFADP) (Swiss Federal Act on Data Protection, in force since 1 September 2023) | Applies to any processing of personal data, including by an AI tool. It provides in particular for a duty to inform, a record of processing activities (with exceptions for some small businesses), a data protection impact assessment (DPIA) where processing may present a high risk, rules on automated individual decisions and notification to the FDPIC of data security breaches likely to result in a high risk. |
| FDPIC (Federal Data Protection and Information Commissioner, EDÖB in German) | Has stated that the FADP applies directly to AI-based processing. Its publications are the reference for what Swiss authorities expect. |
| EU AI Act | Applies in stages, depending on the risk level of the system. According to the European Commission, the prohibited practices and the AI literacy obligation (staff training) have applied since February 2025, the rules on general-purpose AI models since August 2025 and most of the regulation since 2 August 2026. The rules on high-risk systems in certain areas (including employment and HR) have been postponed to 2 December 2027, and those on systems embedded in products to 2 August 2028. It can concern a Swiss company whose AI systems or their outputs are used in the EU. |
| GDPR | May apply to a Swiss company that has an establishment in the EU, offers goods or services to people in the EU or monitors their behaviour. |
| Professional secrecy and contractual obligations | Professional secrecy (lawyers, fiduciaries, doctors and others) and the confidentiality clauses in your contracts still apply when you use AI. |
| FINMA | For financial institutions, supervisory communication 08/2024 expects governance and risk management adapted to AI. |
On public policy, the Federal Council decided in February 2025 to ratify the Council of Europe Convention on AI and to prepare a bill for consultation by the end of 2026, supplemented by non-binding measures (see the Federal Chancellery page in the references). Follow this work: it could change your obligations. To structure your risks, the NIST AI RMF 1.0 framework (govern, map, measure, manage) is a recognised and easy-to-adopt base.
Which clauses should an AI policy include?
An AI use policy has to be complete but readable: aim for a few pages that every employee can actually read. These are the essential clauses.
Roles and responsibilities
- Stakeholders: identify who is responsible for managing, overseeing and using AI tools.
- User responsibility: set out employees' obligations, in particular human review of outputs before any external use.
- Oversight: state who monitors compliance and handles AI-related incidents.
Objectives and scope
- Strategic objectives: define how AI supports company goals (automating tasks, improving customer service).
- Scope: list the AI tools, software and technologies covered, including AI features built into existing software and free online tools.
Acceptable use of AI systems
- Approved tools: the list of tools validated by the company, with their intended use. Any other tool requires prior approval.
- Prohibited data: specify which data must never be entered into a non-approved tool (client data, sensitive personal data, trade secrets, information covered by professional secrecy).
- Restrictions: list prohibited uses (discriminating, manipulating, deciding alone on a matter with significant effects for a person).
- Human review: every AI output is checked by a person before it is sent to a client or used to make a decision.
Inventory of AI uses
Keep a simple register of each AI use: tool, provider, purpose, data involved, internal owner, approval date and risk level. This register is the basis for your risk analysis and your record of processing activities, and for responding to an inspection or a client request.
Data protection and regulatory compliance
- Legal compliance: name the applicable texts (FADP, GDPR where relevant, sector rules).
- Data handling: specify how data is collected, stored and protected, and where it is hosted. For sensitive data, prefer controlled hosting, for example in Switzerland.
- Information and consent: inform the people concerned when AI is involved in processing that affects them, and obtain their consent where the law requires it.
- Minimisation and deletion: process only the data you need and plan for its deletion.
- Impact assessment (DPIA): when processing may result in a high risk to people's personality or fundamental rights, document an impact assessment and the measures chosen. Profiling, automated decisions and large-scale processing of sensitive data are indicators of such a risk.
Clauses to require from AI providers
A policy does not protect you if your contracts do not back it up. For each AI tool provider, check at least:
- whether your data and prompts are used to train its models (the answer you want is no);
- where data is hosted and processed, the sub-processors and any transfers abroad;
- security measures, retention period and conditions for deleting data;
- the obligation to notify you promptly of a security incident;
- audit rights or the certification reports available;
- ownership of the generated outputs and the warranties in case of an intellectual property dispute;
- reversibility: getting your data back and leaving the service without lock-in.
Ethics, bias, transparency and explainability
- Transparency: tell the people concerned when they are interacting with an AI or when an AI contributed to a decision affecting them.
- Explainability: for important decisions, the company must be able to explain the criteria used. Document the model, its limits and its input data, and keep a human able to review the decision.
- Fairness: describe how bias is looked for and corrected (representative test sets, regular reviews).
- Social responsibility: build your ethical principles into the choice and use of AI systems.
Intellectual property
- Inputs: do not submit to an external tool any content you have no rights to or that is confidential.
- Generated content: review outputs before publication. The legal status of AI-generated content remains uncertain in several countries: do not promise exclusivity without legal advice.
- Third-party code and documents: check licences before reusing an output in a product delivered to a client.
Sanctions and handling breaches
Set out what happens when the policy is not followed: a reminder and training, a disciplinary measure under employment law, suspension of access to the tool. Also remind staff of the external framework: the FADP provides for criminal fines of up to CHF 250,000 for certain violations specified in the law (generally intentional ones), which in principle target the responsible individuals, and the AI Act provides for high administrative fines (up to EUR 35 million or 7% of worldwide turnover for prohibited practices, with lower ceilings for other breaches, and the lower of the two amounts for SMEs). A breach of your internal policy alone is not an offence in itself, but it can engage the company's liability. The most common practical consequences are contractual first: losing a client, a tender or a certification.
An AI use policy template to copy
Here is a ready-to-adapt outline. Replace the items in square brackets, delete what does not apply to you, and have the final text approved by your management and, if possible, a lawyer.
Artificial intelligence use policy – [company name] Version [number] · Effective date: [date] · Policy owner: [role] · Next review: [date]
1. Purpose and scope. This policy governs the use of any artificial intelligence tool or feature (generative assistants, AI features built into software, translation, transcription or analysis tools) by [employees, agents and service providers] of [company name].
2. Approved tools. Only the tools listed in the AI use register, kept by [role], may be used for work. Any request for a new tool goes through [procedure or address].
3. Data. It is forbidden to enter the following into a non-approved tool: [client data, sensitive personal data, trade secrets, information covered by professional secrecy, passwords]. Approved tools may process: [data categories].
4. User responsibility. Each employee checks the results produced by AI before using or passing them on, and remains responsible for how they are used. No decision with significant effects on a person (hiring, dismissal, credit, sanction) is made solely by an AI system.
5. Transparency. We inform clients and the people concerned when [an AI interacts with them or contributes to a decision that concerns them].
6. Data protection. AI processing complies with the FADP [and the GDPR where applicable]. An impact assessment is carried out before any high-risk processing. Data is hosted in [location].
7. Providers. Every AI provider is assessed before use against [assessment grid]: no reuse of our data for training, processing location, sub-processors, security, reversibility.
8. Intellectual property. Confidential content, or content we do not hold the rights to, is not entered into external tools. Generated content is reviewed before any publication.
9. Incidents. Any data leak, serious error or unexpected tool behaviour is reported without delay to [role or address].
10. Training. Every employee completes AI use training on joining, then [frequency].
11. Breaches. Failure to comply with this policy may lead to [measures provided for], without prejudice to the sanctions provided by law.
12. Review. The policy is reviewed at least once a year and after any significant regulatory or technical change. Approved by: [name, role, date].
Governance and oversight: setting up clear mechanisms
Ethics and AI governance committee
In an SME, a small group is enough: one member of management, one person in charge of IT or security, a legal or compliance contact and a representative of the users.
- Responsibilities:
- approve new AI uses and tools;
- monitor risks and incidents;
- ensure ethical and regulatory compliance;
- keep the inventory of uses up to date.
Review and audit process
- Regular audits: schedule internal checks, and external ones if the risk calls for it.
- Model updates: make sure tools are updated and retested after a version change.
| Step | Description |
|---|---|
| Risk identification | Analysis of the ethical, legal and technical risks linked to AI. |
| Setting up an audit | Regular assessment of AI systems by internal or external experts. |
| Report and recommendations | Presenting results and improvement proposals to the committee. |
Communicating the policy and training employees
A policy nobody knows about does not protect anyone. Plan short training sessions on joining and every year, concrete examples from your own line of work (what can be pasted into the tool, what cannot) and a contact point for questions. Leaders set the example by applying the policy themselves.
| Action | Goal |
|---|---|
| AI training | Make employees aware of what is at stake with AI. |
| Internal communication | Share good practice and updates. |
| Integration into processes | Align AI with company values. |
Reviewing and updating your policy: frequency and method
- Frequency: review the policy at least once a year, and after any major regulatory or technical change (a new tool rolled out, a change in the law, an incident).
- Method:
- Internal audit: assess how the policy is actually applied.
- Stakeholder consultation: involve employees, clients and experts.
- Update: adapt the text, then circulate the new version with a summary of the changes.
Building AI into business processes
A policy makes the most sense when it supports specific use cases. Not every process suits AI: start with those that are repetitive, well documented and cheap to get wrong.
Identify the processes suited to AI
- Map existing processes: list repetitive, time-consuming tasks and those that rely on large volumes of data.
- Assess the automation potential: keep the tasks that can be automated without compromising quality, ethics or compliance.
- Analyse the available data: check that it exists, that its quality is good enough and that its use is permitted.
Example: automating HR processes
| HR process | AI solution | Points to watch |
|---|---|---|
| Recruitment | Help with screening applications | Risk of bias and automated decision: a person decides, and candidates are informed. |
| Performance management | Analysis of performance data | Employee monitoring: legal basis, proportionality and prior information. |
| Workforce planning | Forecasting staffing needs | Data quality and human review of the assumptions. |
Measuring the impact of AI in your company
To know whether the policy works, track a few indicators (KPIs) chosen before rollout. For return on investment calculations, see our guide measuring AI ROI and choosing KPIs.
Main KPIs to track
- Policy coverage: share of AI uses inventoried and approved, share of employees trained.
- Compliance and security: number of AI-related incidents or data leaks, audit results.
- Operational efficiency: average time saved, reduction in errors in the processes concerned.
- Perceived quality: client and team satisfaction, share of AI outputs corrected during review.
- Return on investment: gains measured against implementation costs, from a documented baseline.
Checklist to assess the impact of AI
- Have you defined clear KPIs and a baseline value for each AI project?
- Is the data used of good quality and compliant with regulations?
- Have you measured the time and costs actually saved?
- Do employees and clients see an improvement?
- Have you reviewed the bias and errors in your tools?
- Are the results aligned with your strategic objectives?
Anticipating technological and regulatory change
AI moves fast, and so do the rules around it. A light but regular watch is better than one big update every three years.
- Regulatory watch: follow FDPIC publications, the Confederation's work on AI regulation, GDPR developments and the AI Act timetable if it concerns you.
- Technology watch: follow new models and AI features built into your software, because they widen the scope of the policy without anyone having decided it.
- External expertise: have a digital-law lawyer review the policy when major changes occur.
- Flexibility: plan a fast update process, without waiting for the annual review.
Steps for a successful implementation
Step 1: define the objectives
Identify the company's specific needs and the expected results.
Step 2: select tools and providers
Assess market solutions using the provider clause checklist above. Choose tools compatible with your systems and your security level.
Step 3: train the teams
Run training sessions and provide practical resources.
Step 4: roll out gradually
Start with a pilot project, adjust processes based on the results, then extend.
Step 5: measure and optimise
Track the KPIs you defined, correct course, and update the policy.
| Step | Goal |
|---|---|
| Define the objectives | Identify needs and expected results. |
| Select tools | Choose suitable technologies and providers. |
| Train the teams | Ensure effective adoption by employees. |
| Roll out gradually | Limit risks and test the solutions. |
| Measure and optimise | Keep improving performance. |
Developing an AI risk management strategy
A risk management strategy helps prevent incidents and strengthens stakeholder trust.
Identify and assess risks
- Algorithmic bias: biased training data can lead to unfair decisions.
- Regulatory non-compliance: AI must comply with the law, notably on data protection.
- Data security: AI systems must be protected against cyberattacks and leaks.
- Impact on jobs: automation changes roles and calls for support.
- Errors and hallucinations: a model can produce answers that are wrong but convincing.
Put preventive measures in place
- Regular audits to spot bias and vulnerabilities.
- Ongoing training of teams on risks and good practice.
- Rigorous testing in a controlled environment before deployment.
- An incident management plan to react quickly when something goes wrong.
Case study: a services SME sets rules for its AI use
This scenario is illustrative: it does not describe a real client and contains no measured figures.
Context: a consulting SME in French-speaking Switzerland finds that several employees are using a public AI assistant to summarise client documents.
Approach:
- Management lists the tools actually in use (inventory).
- It chooses an assistant hosted under a controlled contractual framework and blocks the others for client data.
- It drafts a two-page policy using the template above and appoints an owner.
- It trains the teams with examples taken from their own files.
- It measures at three and six months: unlisted tools detected, employees trained, incidents reported.
What matters: the value comes from clear rules and from following them up, not from the length of the document.
Common mistakes and how to avoid them
- No written policy: write a clear, accessible one.
- Lack of training: run regular training with concrete examples.
- Banning without offering an alternative: if there is no approved tool, shadow AI grows. Offer an approved one.
- Regulatory non-compliance: consult an expert to validate the text and the provider contracts.
- Unidentified bias and errors: test the tools and keep human review.
- A frozen policy: schedule an annual review and another after each significant change.
Checklist for an effective AI policy
- Have you identified the risks of using AI in your company?
- Do you have an inventory of AI tools and uses?
- Does the policy contain clauses on prohibited data, compliance and providers?
- Is a person or a committee responsible for AI governance?
- Have employees been trained in responsible use of AI?
- Is there an incident management plan?
- Is the policy reviewed at least every year?
Approve, track and roll out each new version
A policy that changes without a framework quickly becomes unreadable. A few simple rules keep it reliable:
- one owner of the text (the governance committee or its delegate), who approves every change before it is published;
- a change log at the top of the document: version number, date, what changed, and who or which body approved it;
- event-driven reviews, not only calendar-driven ones: a new tool, a change in the law, an incident, feedback from users;
- a summary of changes sent with each new version, and a short training reminder when everyday rules change.
| Step | Question to ask |
|---|---|
| Assessment | Where does the policy no longer cover real usage (incidents, exception requests)? |
| Consultation | Have IT, legal, HR and end users been heard? |
| Update | Are the affected clauses reworded without weighing down the rest of the text? |
| Approval | Has the committee approved the version that will be published? |
| Communication | Has every employee received the new version and the summary of changes? |
| Follow-up | Do the indicators (compliance, reported incidents, feedback) show an improvement? |
Check before publishing
- The version has been approved by the governance committee.
- The ethical principles are aligned with the company's values.
- Roles and responsibilities are assigned, including IT (security and implementation), legal (compliance) and HR (awareness).
- Employees know where to find the text and whom to ask.
- Monitoring and audit mechanisms are in place.
FAQ: frequently asked questions about AI use policies
-
Why is an AI policy necessary? To ensure responsible, ethical and compliant use, and to put a framework around tools your teams are probably already using.
-
Is there an AI use policy template I can copy? Yes, the outline in this article is meant to be adapted. It does not replace review by a lawyer, especially if you handle sensitive data or data covered by professional secrecy.
-
Which AI tools are covered? All those used in the company: generative assistants, AI features built into Microsoft 365 or other software, transcription and translation tools.
-
Does the FADP apply to AI? Yes, as soon as personal data is processed. The FDPIC has stated that the law applies directly to AI-based processing.
-
Does the AI Act concern a Swiss company? Possibly, if you place AI systems on the EU market or if their outputs are used in the EU. Serving European clients is not enough on its own: check your situation with a lawyer.
-
Do I need an impact assessment (DPIA)? It is required when processing may result in a high risk for individuals; profiling or automated decisions are indicators. It can be documented simply: purpose, risks, measures.
-
What if my employees already use a non-approved AI tool? Take stock without punishing anyone straight away, offer an approved tool and set a date from which only authorised tools are allowed.
-
How much does an AI policy cost? It depends on the size of the company: internal drafting time, possible legal review, training and, where relevant, approved tools. We do not give a typical amount, because it varies too much with the context.
-
How do I ensure algorithm transparency? Document the use, its limits and its input data, test results regularly and inform the people concerned.
-
Which tools can detect bias? Representative test sets and regular human reviews, complemented by the open-source audit libraries described in the NIST frameworks.
-
How do I assess the quality of the data used? Look for inconsistencies, gaps and bias, and validate the data before giving it to an AI system.
-
Who should oversee the policy? A designated person or a small committee (management, IT, legal, users).
-
How often should the policy be updated? At least once a year and after any significant regulatory or technical change, or incident.
-
Can AI replace employees? It mainly automates repetitive tasks. What matters is supporting changes in roles and keeping decisions on important matters in human hands.
-
How do I handle internal resistance? Explain the concrete benefits, involve the teams in drafting and offer training.
-
How do I involve stakeholders? Through short workshops with employees, management and, where relevant, clients and partners, to gather real use cases.
Conclusion
A clear, short and enforced AI use policy protects your data, your clients and your reputation, while letting your teams use AI with confidence. Start from an inventory of what is already in use, adapt the template above, train your teams, then measure and reassess every year. If you would like support, contact houle.ai.