AI guidelines for SMEs define who may use which AI tools for which tasks, with which data, and under what level of human oversight. A clear AI policy limits risks, protects company knowledge, and answers a concrete question for your team: "Am I allowed to use AI for this task – and if so, under what conditions?"
Especially in family businesses or small teams, usage often begins without central coordination: One employee tests a text assistant, a colleague creates a spreadsheet using a private account, and another person connects a free tool to their email inbox. This creates risks for data protection, trade secrets, and quality.
In my work with SMEs, I've seen two problematic reactions: complete rejection due to uncertainty and untested use due to time constraints. Neither is economically sound. A better approach is a manageable AI code of conduct for company practice that aligns with the actual risk, processes, and the size of your team. Such an AI guideline for employees establishes clear boundaries without categorically prohibiting sensible applications.
An AI policy should enable responsible use. A mere prohibition list, on the other hand, increases the risk that employees will use private accounts and undocumented alternative methods.
What AI guidelines for SMEs must specifically regulate
A company AI policy is a binding internal regulation for the selection, use, and control of AI systems. It should clearly answer at least five questions:
- Tools: Which applications are permitted, permitted with restrictions, or prohibited?
- Data: What information may be entered, uploaded, or processed via interfaces?
- Responsibilities: For which work steps may AI be used?
- Control: Who checks the results and who grants approval?
- Responsibility: Who decides on new tools, incidents, and exceptions?
An AI policy in a family business doesn't need to be a comprehensive legal manual. For a company with up to 20 employees, a clear and concise basic document with appendices for a tool list, data classes, and approval process can suffice. The crucial factor is that the regulations are understood, followed, and regularly updated in daily work.
The quick start guide for small teams
If you need clear boundaries in the short term, start with five decisions. These initial AI guidelines for SMEs don't replace a full review, but they do reduce acute risks:
- Use only centrally approved and documented accounts.
- Do not enter any personal data, trade secrets or confidential documents without prior review and approval.
- Have every relevant result checked by a professionally responsible person.
- Do not send or publish AI content to customers without checking it first.
- Report unsafe expenditures, data leaks, and suspicious system reactions immediately.
Next, identify which applications your team is already using. A pragmatic AI maturity check will help you assess not only the technology but also processes, data, and skills.
Distinguish between permitted, restricted, and prohibited uses.
The tool list is the central working basis of your AI policy. However, simply listing permitted AI tools is not enough. For each tool, also document the approved purpose, the permissible data, the responsible role, and the required quality control.
Erlaubte Nutzung
You can classify applications and activities as permitted if their purpose, account, terms of service, and data processing have been reviewed. Typical examples include:
- Collection of ideas using publicly available or explicitly released information,
- linguistic revision of self-written, non-confidential texts,
- Summary of public documents,
- Creation of internal drafts without personal, customer or partner data,
- Translation of approved content followed by human review.
Restricted permitted use
Activities permitted under certain conditions are restricted. These may include draft contracts, calculations, application documents, health information, customer data, program code, or automated publications. These applications require at least documented approval, appropriate permissions, and a technical review.
Prohibited use
Any prohibited use must be clearly stated. Typical prohibitions include:
- Entry of passwords, access data or secret keys,
- Uploading confidential customer, personnel or contract documents without authorization,
- Using private AI accounts for business information,
- Installation of unverified browser extensions or apps,
- Bypassing internal permissions and approvals,
- Fully automated publication without expert review,
- Decisions about people based solely on an AI output,
- Passing on unverified AI results as confirmed facts.
Even free tools need to be reviewed. Free plans may include limited contractual rights, allow the use of data for training purposes, or lack administrative functions. Therefore, carefully examine the specific product, plan, settings, terms and conditions, and intended use.
Clearly defining data protection and data classes
AI guidelines for SMEs should differentiate data according to its sensitivity level. Four classes are often sufficient for a small business:
- Public: Previously published information that may generally be entered into a released tool.
- Internal: Work documents without special confidentiality, the processing of which is only permitted in company-approved accounts.
- Confidential: Offers, calculations, strategies, partner data and non-public documents that require explicit approval.
- Special protection: Personal data, special categories of personal data, access data, and key trade secrets. Their processing using AI is generally prohibited or only permitted after a documented legal and technical review.
The principle is: No data should be entered without consent and a suitable legal basis. This applies not only to names or email addresses. Meeting notes, CVs, customer histories, photos, order data, or combinations of seemingly harmless information can also contain personal data.
Anyone processing personal data using AI must, in particular, comply with the principles outlined in Article 5, establish a legal basis according to Articles 6 and 9, comply with the data processing agreements stipulated in Article 28, and meet the security requirements of Article 32 of the GDPR. You can find business decision-making support in our article on GDPR and AI for SMEs.
When using external providers, also check whether a data processing agreement is required, where data is processed, which subcontractors are involved, whether input is used for training purposes, and when content is deleted. This check must be carried out before approval.
Protecting trade secrets and family knowledge
In family businesses, a significant part of the company's value lies in its experiential knowledge: pricing logics, supplier relationships, recipes, production details, negotiation patterns, and knowledge about long-standing customers. If this knowledge is copied uncontrollably into private AI accounts, the business loses track of storage locations, access, and potential further use.
An AI policy therefore not only protects against data leaks. Documentation makes decisions and working methods comprehensible for the next generation. It prevents critical processes from depending on a single person, a private account, or an undocumented prompt.
Copyright, sources and attribution
AI output may be factually incorrect, contain copyrighted material, or be subject to terms of use that are unsuitable for its intended purpose. Therefore, stipulate that employees must verify the origin of input from texts, images, videos, music, and program code, as well as any identifiable third-party rights.
EU law currently contains no specific, uniform rules across the Union regarding the copyright protection of AI-generated content. Purely machine-generated content without sufficient human creative contribution is generally not protected under the concept of a work based on human creativity. For AI-supported works, the specific human contribution and the applicable law are decisive.
For business practice, this means:
- Only use input for which you have sufficient rights or permissions.
- Examine relevant results for identifiable takeovers and third-party rights.
- For important content, document the tools used and the human editing steps.
- Clarify usage rights before using AI-generated images, texts, music, or program code commercially.
- Indicate AI support where required by laws, industry rules, platform specifications or the specific communication situation.
Internal documentation is not the same as public labeling. Both requirements must be reviewed and addressed separately.
Human-in-the-Loop as a mandatory quality process
In your company policy, "human-in-the-loop" means that a technically responsible person reviews an AI result before it is used for a relevant decision, publication, or customer communication. This person must have the necessary time, expertise, and authority for a thorough review.
Depending on the task, the testing step should include the following points:
- Are the facts, names, dates, and sources correct?
- Are the calculations and conclusions comprehensible?
- Does the tone of voice match the brand and the specific situation?
- Does the result contain biases, disadvantages, or inappropriate statements?
- Were data protection, confidentiality and copyright taken into account?
- Is it clear who gives the final approval?
Article 14 of the EU AI Act contains requirements for human oversight of high-risk AI systems. This does not create a blanket legal obligation for a specific human-in-the-loop procedure for every use of AI. Therefore , legal oversight of high-risk AI and your company's quality process should be conceptually separated.
Regardless of the legal classification, human quality control in customer communication, offers, calculations, and public content is operationally sound. AI can support expertise but does not assume responsibility.
Clearly assign responsibilities
In a small business, one person can take on several roles. Nevertheless, the responsibilities must remain clearly distinguishable:
- Managing directors: approves the AI policy, determines the acceptable risk, and provides time and resources.
- Person responsible for AI: manages the tool list, coordinates inspections, documents decisions, and organizes updates.
- Person responsible for data protection: assesses the processing of personal data and involves external experts if necessary.
- IT responsibility: Manages accounts, permissions, interfaces, logging, and technical security measures.
- Subject matter expert: Review the results for content and approve them for their area.
- Employees: They comply with the rules, document relevant usage, and report uncertainties or incidents.
Additionally, define a deputy. In a family business, the entire approval process should not depend on a single person who may be temporarily or permanently unavailable.
Minimum technical requirements for small businesses
An AI policy without technical implementation is insufficient. For most SMEs, the following measures represent a sensible minimum standard:
- company-owned rather than private user accounts,
- Multi-factor authentication for all relevant accounts,
- centralized granting and withdrawal of authorizations,
- Separate roles for usage and administration,
- regular monitoring of active accounts and extensions,
- Logging of important interfaces and automated actions,
- no uncontrolled connections to email, cloud storage or customer systems,
- Regular backup of important configurations and documentation.
Prompt manipulation and unreliable expenditures
Prompt manipulation , also known as prompt spoofing, refers to the attempt to manipulate an AI system into performing unwanted actions by injecting instructions. For example, a manipulated website, file, or email might contain instructions that a connected assistant mistakenly interprets as a work order.
The more access a system has to internal data, emails, or tools, the more restrictive its permissions and shares must be. Separate read, write, and execute permissions, limit access to the necessary scope, and require explicit confirmation for critical actions.
Browser extensions and untested interfaces can also transmit data or trigger actions. Therefore, they must be included in the formal approval process.
What to do in the event of a security incident
A security incident could be an accidentally uploaded customer document, a compromised account, an incorrect publication, or uncontrolled interface access. Your policy should include a short reporting procedure:
- Stop using the service and disconnect the affected connection.
- Save evidence, logs, and messages, and do not delete them.
- Inform the internal contact person immediately.
- Document the affected data, persons, systems and times.
- Block or change access and end active sessions.
- Have your data protection and notification obligations professionally reviewed.
- Fix the cause and then update the policy.
An open reporting culture is crucial for a swift response. If employees expect sanctions for every mistake, incidents are often reported late. Deliberate circumvention and repeated violations require clear consequences; errors reported in good faith should be used to improve processes.
AI competence and employee awareness
According to Article 4 of Regulation (EU) 2024/1689, providers and operators of AI systems must take measures to the best of their ability to ensure that their staff and other persons working with AI systems on their behalf have a sufficient level of AI competence. This provision has been in force since 2 February 2025. Factors to be considered include technical knowledge, experience, training, and the context of use.
For a small team, this doesn't need to result in an extensive training program. A short mandatory training session should at least cover the following topics:
- which tools and activities are permitted
- which data must never be entered without verification,
- how AI spending is controlled,
- how uncertainty, errors and incidents are reported,
- why plausible answers can be factually incorrect,
- which rules apply to sources, copyright and attribution.
Document the date, content, participants, and, if applicable, a brief practical test. Repeat employee awareness training after significant changes, new roles, and new applications. A distributed document alone is not reliable proof of AI competence.
Before and after: Example of a South Tyrolean family business
The following example is fictional: A family business with twelve employees uses three free text assistants in the office via private accounts. Proposals are uploaded for revision, meeting notes are summarized, and social media posts are published directly. The terms and conditions were not reviewed, and in case of an error, it is unclear who is responsible.
Following the implementation of the AI policy, the company operates with two approved tools and company accounts. Public information may be used for drafts. Customer data, quotes, and internal calculations are classified as confidential and require prior approval.
A designated person manages accounts and the tool list. Relevant subject matter experts review content, and a clear reporting process governs incidents. This ensures that experience and decisions remain traceable, even as responsibilities are passed on to the next generation.
A limited AI pilot project with clearly defined roles is usually more sensible than immediate approval for the entire operation.
Copyable framework for your AI policy
AI guidelines for SMEs must be adapted to the roles, data types, tools, and workflows of the respective company. The following template provides a concise basis for this. This template does not constitute individual legal advice.
1. Purpose: This policy enables the responsible use of AI and protects people, company knowledge, personal data and the quality of our work.
2. Scope: This policy applies to all employees, managers, external persons and systems that use AI on behalf of the company.
3. Tool list: Only the approved tools [Enter tools] via company accounts and for the documented purposes may be used.
4. Data Classifications: Data is classified as public, internal, confidential, or specially protected. Confidential and specially protected data may only be processed after documented review.
5. Prohibited Use: Access data, trade secrets, and personal data may not be entered into AI systems without authorization. Private accounts and unverified extensions are prohibited for company data.
6. Quality control: Relevant results are checked for facts, calculations, tone, rights, data protection and possible disadvantages before use by [insert role].
7. Documentation: For important results, the tool, purpose, data class, responsible person, test steps and release are recorded.
8. Incidents: Suspicious expenditures, data leaks, compromised accounts and rule violations must be reported immediately to [insert contact details].
9. Training: All affected persons will receive appropriate instruction before use. Training sessions and updates will be documented.
10. Review: The policy is reviewed at least every six months and also in the event of new tools, interfaces, legal changes or security incidents.
11. Violations: Deliberate or repeated violations will be dealt with according to internal rules. Reported errors in good faith serve to improve processes and do not automatically lead to sanctions.
Introduce AI guidelines for SMEs in 30 to 90 days
Days 1 to 10: Making usage visible
- Record all AI applications, accounts, extensions, and interfaces used.
- Document tasks, entered data, and people involved.
- Pause obviously risky use until the matter is resolved.
Days 11 to 30: Define risks and pilot rules
- Define data classes and prohibited uses.
- Select one or two tools and manageable use cases.
- Define responsibilities, the approval process, and incident reporting.
- Set up technical accounts, multi-factor authentication, and permissions.
Days 31 to 60: Testing in everyday life
- Test the rules with typical tasks.
- Limit testing steps and documentation to the necessary scope.
- Revise unclear wording together with the team.
Days 61 to 90: Mandatory implementation
- The management has the final version professionally reviewed and approves it.
- Employees receive a short mandatory training course with practical examples.
- Participation, tool approvals and responsibilities are documented.
- A date for the next inspection will be set.
If you want to collaboratively develop tools, processes, and responsibilities, Berger+Team, a freelance collective from Bolzano, can support you with consulting on AI and digitalization . The solution is tailored to the actual processes and risks of your business.
Questions and answers about the AI policy
Is an AI policy legally mandatory for every SME?
There is no general obligation for every SME to have a document entitled "AI Policy". However, obligations arising from data protection, information security, work organization, and the EU AI Act may make a documented policy necessary for business operations.
How long should an AI policy be for a family business?
For small teams, a concise, easy-to-understand guideline of just a few pages is often sufficient. Supplement the main document with an up-to-date tool list, data classes, and brief test instructions.
Are employees allowed to use free AI tools?
Free tools may only be used if the specific product, tariff, data processing, and intended purpose have been approved. A provider's reputation does not replace this verification.
Is it permissible to enter personal data into an AI system?
Only if there is a suitable legal basis, a clear purpose, necessary contracts, and appropriate security measures in place. Without documented review and approval, your policy should prohibit the input.
Does AI-generated content always need to be labelled?
There is no general labeling requirement for all AI support. Check legal requirements, industry rules, platform terms and conditions, and the specific communication situation.
Who is responsible for an incorrect AI output?
An AI system does not assume operational responsibility. Your policy must define who reviews the result and who grants final approval for publication, decision, or customer communication.
What needs to be checked during quality control?
At a minimum, check the facts, sources, calculations, tone, data protection, copyright, and potential discrimination. Document the verification process for results that are significant from a technical or economic perspective.
How can AI expertise be documented?
Document the training date, content, participants, roles, and, if necessary, a short practical test. Repeat the training if tools, risks, or tasks change significantly.
How often should AI guidelines for SMEs be updated?
Review the policy at least every six months. An additional review is required if new tools, tariffs, interfaces, data types, legal requirements, or security incidents are added.
What happens in case of a violation?
First, stop the risky use, secure relevant information, and assess potential consequences. Then, differentiate between a well-intentioned error report and a deliberate circumvention, ensuring that incidents continue to be reported promptly.
Who should legally review the AI policy?
For personal data, employee data, automated decisions, or confidential core processes, appropriate data protection or legal advice should be sought. A template can expedite internal coordination, but it does not replace a review of the specific application.
Conclusion: Clear rules instead of blanket bans
Effective AI guidelines for SMEs combine permitted AI tools, data classes, responsibilities, human-in-the-loop integration, technical security, and employee awareness training into a comprehensible operational system. This results in less uncertainty for SMEs and family businesses, more transparent decisions, and better protection of their experiential knowledge.
Start with the applications your team actually uses. Set clear boundaries, allow approved use cases, and determine who reviews and approves results. AI remains a tool; judgment and responsibility remain with humans.
