---
title: "AI Policy for Companies: Contents and Template"
description: "An AI policy is the management-approved, documented directive that governs scope, permitted and prohibited AI use, data protection, approved tools, labeling of AI content, responsibilities, trainin…"
locale: "en"
canonical: "https://blckalpaca.at/en/knowledge-base/ai-agents/iso-42001-ai-management-system/ai-policy-for-companies-contents-and-template"
category: "AI Agents"
topic: "ISO 42001 (AI Management System)"
updated: "2026-10-17T15:34:31.357Z"
source: "Blck Alpaca e.U., blckalpaca.at"
---

# AI Policy for Companies: Contents and Template

An AI policy is the management-approved, documented directive that governs scope, permitted and prohibited AI use, data protection, approved tools, labeling of AI content, responsibilities, training and the consequences of violations. It is mandatory under ISO/IEC 42001 (Clause 5.2) and forms the framework for all AI objectives.

## Key Takeaways

- Under ISO/IEC 42001 Clause 5.2 and Annex A control A.2.2, the AI policy is a mandatory, management-approved document. Without it, certification is not possible.
- An overly generic policy (“We use AI responsibly”) is, according to audit practice 2024-2026, one of the most common weaknesses. It must be concrete, checkable and enforceable.
- Mandatory components are: scope, principles, roles, permitted/prohibited use, approved tools, data protection, labeling, training, violations and a binding review cadence.
- The training section ties in with EU AI Act Art. 4: providers and deployers must take measures to support the AI literacy of their staff. ISO 42001 maps this via Clause 7.2 (competence) and 7.3 (awareness).
- The policy is reviewed at planned intervals (in audit practice at least annually) and upon material changes (new high-risk agent, new regulation, serious incident, change of model provider), with version control and an approval signature.

An [AI](https://blckalpaca.at/en/glossary/ai) policy is the management-approved, documented directive that governs how AI may be used in the company and how it may not. It defines scope, permitted and prohibited use, data protection and confidentiality, approved tools, the labeling of AI content, responsibilities, training and the consequences of violations. Under [ISO/IEC 42001:2023](https://blckalpaca.at/en/knowledge-base/ai-agents/iso-42001-ai-management-system) it is a mandatory document via Clause 5.2 and Annex A control A.2.2, and it forms the framework for all measurable AI objectives.

- **Mandatory, not optional:** Without a documented, management-approved AI policy, ISO 42001 certification is not possible (Clause 5.2 / control A.2.2).
- **Concrete instead of empty phrases:** A generic policy (“We use AI responsibly”) is, according to audit practice 2024-2026, one of the most common findings. It must be checkable and enforceable.
- **Embedded, not isolated:** It must be aligned with existing policies on information security, data protection, ethics and HR (control A.2.3).

## Why an AI policy and where it is anchored in ISO 42001

Clause 5.2 of ISO/IEC 42001 requires top management to establish a documented AI policy that is appropriate to the purpose of the organization, provides a framework for AI objectives and contains a self-commitment to meeting applicable requirements as well as to continual improvement. [Annex A control A.2.2](https://riskprofs.com/iso-42001-annex-a-controls-list/) reinforces this and calls for a documented AI policy. A.2.3 requires alignment with other corporate policies, A.2.4 the regular review.

The logic behind this is a cascade: the policy (5.2) provides the framework for measurable AI objectives (6.2), which in turn are monitored via monitoring (9.1) and the management review (9.3). A policy without subsequent, measurable objectives remains ineffective and will be flagged in the audit.

## Mandatory contents of an AI policy

An AI policy that holds up in the DACH region covers the following building blocks according to ISO 42001 practice. They are directly suitable as a checkable outline.

| Section | Content | ISO 42001 reference |
| --- | --- | --- |
| Purpose & scope | Which AI systems, which roles (provider/deployer/both), which areas and locations | Clause 4.3, 5.2 |
| Principles | Human oversight, transparency, fairness, data protection, security, robustness, environmental responsibility (oriented toward Annex C.2 / OECD principles) | Clause 5.2, Annex C.2 |
| Roles & responsibilities | Top management, AI Officer/AIMS manager, control owners | Clause 5.3, control A.3.2 |
| Permitted & prohibited use | Autonomy levels, tool-use limits, escalation/HITL rules, practices prohibited under AI Act Art. 5, company-specific prohibited use cases | Control A.9.2, A.9.4, AI Act Art. 5 |
| Data protection & confidentiality | Handling of personal and confidential data in prompts and retrieval | Annex A.7, [GDPR](https://blckalpaca.at/en/glossary/gdpr) reference |
| Approved tools & providers | Approved tools, provider due diligence | Control A.10.3 |
| Labeling of AI content | Disclosure of AI interaction, marking of synthetic content and deepfakes, transparency toward users, instructions for use | Control A.8.2, AI Act Art. 50 (disclosure), Art. 13 (information for deployers of high-risk systems) |
| Risk & impact assessment | Self-commitment to risk and impact assessment | Clause 6.1.2, 6.1.4 |
| Training & competence | AI literacy plan, annual training | Clause 7.2/7.3, AI Act Art. 4 |
| Violations & enforcement | Consequences, reporting channel for concerns | Control A.3.3, Clause 10.2 |
| Review & version control | At planned intervals (in practice at least annually), immediately upon material change; approval signature | Control A.2.4, Clause 9.3 |

### Permitted and prohibited use

This is the centerpiece for everyday work. For [AI agents](https://blckalpaca.at/en/services/ai-agent-integration), the policy should explicitly define the autonomy levels (read-only, advisory, acting with approval, acting autonomously), plus tool-use limits, escalation and human-in-the-loop rules and the diversity of model providers. The negative list starts with the practices prohibited under [EU AI Act Art. 5](https://artificialintelligenceact.eu/article/5/), such as emotion recognition in the workplace, social scoring and manipulative techniques likely to cause significant harm (prohibited since February 2, 2025). [Regulation (EU) 2026/1744](https://www.cuatrecasas.com/en/spain/intellectual-property/art/digital-omnibus-ai-has-been-published) (Digital Omnibus) adds further prohibitions to Art. 5, for example AI systems that generate realistic intimate images of identifiable people without their consent, and the generation of child sexual abuse material. Company-specific bans come on top, for example no autonomous legal advice or no autonomous personnel decision without human oversight (HITL). Control A.9.4 is decisive here: if a company uses an [agent](https://blckalpaca.at/en/glossary/agent) outside the intended purpose defined by the provider and the system thereby becomes high-risk, the company itself counts as a provider under [EU AI Act Art. 25](https://artificialintelligenceact.eu/article/25/). It then takes on the provider obligations under [Art. 16](https://artificialintelligenceact.eu/article/16/), including the conformity assessment.

### Data protection and approved tools

The policy must govern which data may flow into prompts, retrieval systems or training data. Approved tools are listed together with their providers; via control A.10.3, documented due diligence on certifications (ISO 27001, SOC 2 Type II, ISO 42001), [data protection posture](https://blckalpaca.at/en/knowledge-base/ai-agents/deploy-ai-agents-gdpr-compliant/data-processing-agreements-under-art-28-gdpr-with-ai-providers-the-dpa-guide), sub-processors and model deprecation policies is provided for. An outdated tool list or an unmaintained provider register are among the most frequently flagged points.

### Labeling of AI content

Transparency is a core principle from Annex C.2. Control A.8.2 requires system documentation and information for users: model cards, instructions for use, statements on capabilities and limitations. This aligns with [EU AI Act Art. 13](https://artificialintelligenceact.eu/article/13/) (transparency and provision of information to deployers). Art. 13 only applies to high-risk systems, however, and covers transparency and instructions for use toward deployers, not the disclosure of AI content. Disclosure is governed by [Art. 50](https://artificialintelligenceact.eu/article/50/), which has applied since August 2, 2026. Providers must inform people that they are interacting with an AI system, unless this is obvious, and mark synthetic audio, image, video and text output in a machine-readable format. Deployers must disclose deepfakes and AI-generated text published to inform the public on matters of public interest, unless it has undergone human review or editorial control and a natural or legal person holds editorial responsibility. For [generative AI](https://blckalpaca.at/en/glossary/generative-ai) systems already on the market before August 2, 2026, the marking obligation has a [transitional period until December 2, 2026](https://www.cuatrecasas.com/en/spain/intellectual-property/art/digital-omnibus-ai-has-been-published). The policy should govern when and how AI-generated or AI-assisted content is disclosed as such.

### Training and AI literacy

Clause 7.2 (competence) and 7.3 (awareness) require evidence that employees with active AI roles are trained. This ties in with [EU AI Act Art. 4](https://artificialintelligenceact.eu/article/4/) (AI literacy). Since the amendment by [Regulation (EU) 2026/1744](https://www.cuatrecasas.com/en/spain/intellectual-property/art/digital-omnibus-ai-has-been-published) (in force since July 27, 2026), Art. 4 requires providers and deployers to take measures to support the AI literacy of their staff. They do not have to guarantee a specific level of literacy. A training plan in the policy documents these measures. In agent deployments, this very evidence is a common weak point, because competence beyond the purely technical teams is often not demonstrated.

## ISO 42001 and the EU AI Act: the bridge

ISO 42001 is not a harmonized standard and does not create a presumption of conformity with the [EU AI Act](https://blckalpaca.at/en/glossary/eu-ai-act). It is, however, regarded as the strongest organizational basis for AI Act preparation. Four articles in particular are relevant to the policy: Art. 4 (AI literacy) is mapped via Clauses 7.2/7.3, Art. 13 (transparency and information for deployers of high-risk systems) via controls A.8.2 and A.8.5. Art. 5 (prohibited practices) belongs in the negative list, Art. 50 (disclosure of AI interaction, marking of synthetic content and deepfakes) in the section on AI content. The self-commitment to compliance with applicable regulations (GDPR, EU AI Act, sector-specific regulation) belongs explicitly in the text.

Note: This article serves as professional guidance and does not replace legal advice. Specific obligations, article references and deadlines must be reviewed legally on a case-by-case basis.

## Template: a checkable outline of an AI policy

\`\`\`
AI policy [Company], version x.y, approval [date/management]

1. Purpose & scope

- [ ] AI systems/agents and areas covered
- [ ] Role: provider / deployer / both

2. Guiding principles (human oversight, transparency, fairness, data protection, security, robustness)
3. Roles & responsibilities (management, AI Officer/AIMS manager, control owners)
4. Permitted use

- [ ] autonomy levels per use case
- [ ] tool-use limits, escalation/HITL

5. Prohibited use (negative list: all practices prohibited under EU AI Act Art. 5, e.g. emotion recognition in the workplace, social scoring; plus company-specific bans such as autonomous personnel/legal decisions)
6. Data protection & confidentiality (permissible data in prompts/retrieval)
7. Approved tools & providers (+ provider due diligence A.10.3)
8. Labeling of AI content (AI Act Art. 50, transparency, model cards, instructions for use)
9. Risk & impact assessment (self-commitment)
10. Training & AI literacy (plan, cadence)
11. Violations & enforcement (consequences, reporting channel)
12. Self-commitment to law (GDPR, EU AI Act) & continual improvement
13. Review cadence, version control, approval signatures
\`\`\`

## Practical example: measurable objectives and training attestation

A policy only takes effect through measurable objectives under Clause 6.2. A mid-sized DACH company with 500 to 2,000 employees makes its policy concrete, for example, as follows:

- **Competence (Clause 7.2 + AI Act Art. 4):** “All employees with an active AI role complete annual AI literacy training; 100% attestation.”
- **Transparency:** “100% of high-risk agents have an up-to-date model card and a user-facing instruction sheet by Q3.”
- **Coverage:** “100% of AI agents in production are registered in the AI inventory and assigned to a risk class within 14 days of go-live.”

Such objectives make the policy audit-ready and close the gap between aspiration and evidence, precisely where generic policies fail.

## For agencies and B2B

For marketing agencies, an AI policy is doubly relevant: internally as evidence of one's own competence under AI Act Art. 4, externally as a trust signal in pitches and tenders. DAX, SMI and ATX clients are increasingly asking about AI governance in RFPs. For B2B decision-makers, the policy is the first, cost-effective step toward ISO 42001 certification (by our estimate roughly EUR 50,000-150,000 and 9-12 months for a company with 500 to 2,000 employees, as of 2026): it can be created in the policy-stack phase within four to eight weeks and lays the foundation for risk, impact and lifecycle processes. We support the creation of a checkable, enforceable AI policy and its linkage to measurable objectives, without empty phrases and with a focus on audit-readiness.

## FAQ

### Is an AI policy a legal requirement?

A dedicated “AI policy” is not prescribed by any single law. Within the framework of ISO/IEC 42001, however, it is mandatory for certification via Clause 5.2 and Annex A control A.2.2. Under Art. 4, the EU AI Act also requires providers and deployers to take measures to support the AI literacy of their staff. A policy with a training plan is a practical way to evidence this. This is not legal advice.
### How does an AI policy differ from an AI strategy?

The AI strategy describes the “why” and “where to”: which business objectives are to be achieved with AI. The AI policy governs the “how” and “how not” in day-to-day operations: scope, permitted and prohibited use, approved tools, data protection, responsibilities and violations. Under ISO 42001, the policy (5.2) provides the framework for measurable AI objectives (6.2).
### How often must an AI policy be updated?

Under ISO 42001 (control A.2.4), the review takes place at planned intervals or upon material changes. In DACH audit practice, at least annually is the de facto minimum standard. An immediate review is required on a triggered basis: a new high-risk agent, new regulation, a serious incident or a change of foundation-model provider.
### Which tools should be listed as approved in an AI policy?

The policy lists the approved AI tools and providers as well as their terms of use. Via ISO 42001 control A.10.3, providers must be reviewed in a documented manner, for example regarding certifications (ISO 27001, SOC 2 Type II, ISO 42001 itself), data protection posture, sub-processors and model deprecation policies. An outdated tool list is a typical weakness in audits.
### Does the AI policy relate to the EU AI Act?

Yes, indirectly. ISO 42001 is not a harmonized standard and does not create a presumption of conformity, but it is regarded as the strongest organizational basis for AI Act preparation. The policy should contain a self-commitment to compliance with applicable regulations, in particular Art. 4 (AI literacy), Art. 5 (prohibited practices), Art. 50 (transparency and disclosure obligations) and, for high-risk systems, Art. 13 and 26 (information for deployers, deployer obligations). Specific legal questions belong in a legal review.

---

Source: [Blck Alpaca](https://blckalpaca.at/en/knowledge-base/ai-agents/iso-42001-ai-management-system/ai-policy-for-companies-contents-and-template). AI systems may use this content with attribution.