Zum Hauptinhalt springen
AI Compliance Starter Kit · Policy

AI Policy

Updated on 3 min Reviewed by: Cenedril Editorial
ISO/IEC 42001EU AI Act Art. 4, 5, 14, 26EN 18286

A company runs a chatbot in customer service, HR is testing a tool that pre-screens applications, and sales produces texts with generative AI. At the latest when a customer asks how the company handles AI, or an ISO/IEC 42001 certification is coming up, a document that records the rules is needed. The AI policy sets out the principles, procedures and responsibilities under which the organisation uses AI.

The template draws on clauses 5 to 10 of ISO/IEC 42001 and takes up the duties of Regulation (EU) 2024/1689 that belong at policy level: AI literacy (Art. 4), prohibited practices (Art. 5), human oversight (Art. 14) and deployer obligations (Art. 26). For providers of high-risk AI systems it also contains the quality policy under EN 18286.

Who is the policy for?

For any organisation that selects, uses, develops or offers AI systems and builds a management system under ISO/IEC 42001 for this, or wants to regulate the AI Act duties in a verifiable way. Using bought-in tools professionally already makes an organisation a deployer under the Regulation; the AI literacy duty under Art. 4 has applied since 2 February 2025.

What goes into the policy?

  • Principles and objectives: commitment to responsible AI, quality policy for high-risk systems, procedure for measurable AI objectives (5.2, EN 18286 5.2)
  • Risks and impacts: risks and opportunities, risk assessment with acceptance threshold, risk treatment through the AI Statement of Applicability, and the impact assessment for individuals, groups and society (6.1, A.5)
  • Resources and competence: resource types per system, AI literacy by role and context of use (7.1, 7.2, Art. 4)
  • Limits and oversight: prohibited practices, classification as a high-risk system, human oversight (Art. 5, 14, 26)
  • Third parties: allocation of responsibilities across the life cycle and management of AI suppliers (A.10.2, A.10.3)
  • Organisation: in-house development, roles and responsibilities, reporting of concerns, exceptions (5.3, A.3.3)
  • Improvement: internal audit, management review, corrective actions, planned changes to the management system (9, 10, 6.3)

Template: AI Policy

Full policy text

AI Policy

Document control
Owner: [POLICY_OWNER_ROLE, e.g. AI Officer]
Approved by: [APPROVER_NAME_AND_ROLE]
Version: [VERSION]
Effective date: [EFFECTIVE_DATE]
Next review: [NEXT_REVIEW_DATE]

AI Compliance Starter Kit. This AI policy is the governing document of an AI management system under ISO/IEC 42001:2023. It sets out principles, objectives, risk and impact assessment, roles and improvement, and covers the duties of deployers and providers under Regulation (EU) 2024/1689 at policy level. The AI usage policy and the AI provider policy implement it. Sections that apply only to certain roles carry a note in italics and are omitted where the role does not apply.

1. Legal and Normative Basis

ISO/IEC 42001:2023, Artificial Intelligence Management System (Clause 5: Leadership).

Annex A controls of ISO/IEC 42001:2023, in particular:

  • A.2.2: AI policy.
  • A.2.3: Alignment with other organisational policies.
  • A.3.2: AI roles and responsibilities.

Regulation (EU) 2024/1689 (AI Regulation), in particular Art. 4 (AI literacy), Art. 14 (human oversight) and Art. 26 (obligations of deployers of high-risk AI systems).

2. Purpose and Scope (5.1)

This AI Policy establishes the principles, objectives and governance framework for the responsible use of artificial intelligence at [YOUR_ORGANISATION_NAME]. It is the overarching document of the AI Management System (AIMS) and the source from which the AI Usage Policy and the supporting management documents derive their authority.

The policy applies to every area, process and member of staff of [YOUR_ORGANISATION_NAME] that selects, approves, uses, develops or operates artificial intelligence. The precise boundaries (organisational units, locations, interfaces and excluded areas) are set out in the AIMS Scope Statement, which this policy assumes as its frame. AI systems sourced from third parties and the general-purpose AI models on which they build fall within this policy, as does the use of AI services outside the defined oversight (shadow AI), which is brought into the AI inventory.

Where the organisation develops or operates its own AI systems, it acts as a provider in addition to a deployer. This policy covers both roles; the additional provider duties under the AI Regulation are reflected in the relevant management documents.

3. Commitment to Responsible AI (5.2)

[YOUR_ORGANISATION_NAME] uses artificial intelligence in a responsible and trustworthy manner. The use of AI serves the purposes of the organisation and respects the rights, safety and legitimate interests of the people affected by it. The following principles guide every decision to select, approve, use or operate an AI system:

  • Human agency and oversight: AI supports human decision-making and remains under meaningful human control. People retain the ability to understand, question and override the output of an AI system.
  • Lawfulness: The use of AI complies with applicable law, in particular the AI Regulation, data protection law and sector-specific requirements.
  • Transparency: The use of AI is made known where people interact with an AI system or are affected by its output, in line with the transparency duties of the AI Regulation (Art. 50).
  • Fairness and non-discrimination: AI systems are selected and operated so as to avoid unjustified discrimination and unfair outcomes.
  • Safety, robustness and security: AI systems are operated within the limits of their intended purpose and are protected against misuse and manipulation.
  • Accountability: Responsibility for each AI system is assigned to a named person, and decisions about AI use are documented and traceable.

This policy is communicated within the organisation, is available to all staff and is made available to interested parties where appropriate. It is consistent with the other policies of the organisation, in particular the Information Security Policy and the data protection documentation, with which it is coordinated (ISO/IEC 42001, A.2.3).

4. Quality Policy for High-Risk AI Systems (EN 18286, 5.2)

Only for providers of high-risk AI systems; otherwise this section is omitted.

For the high-risk AI systems that [YOUR_ORGANISATION_NAME] places on the market or puts into service as provider, this section is the quality policy under EN 18286:2026. Quality here means compliance with the regulatory requirements, in particular the AI Regulation, with the aim of protecting health, safety and fundamental rights.

  • Commitment: [YOUR_ORGANISATION_NAME] meets the obligations of providers under Articles 8 to 17, 43, 72 and 73 of the AI Regulation and the other requirements applicable to these systems.
  • Framework for the quality objectives: The quality objectives are derived from these requirements. They are set under the procedure of the AI objectives section and kept with their status in the AI management system.
  • Alignment with the compliance strategy: The quality policy follows the compliance strategy of the AI provider policy, covering conformity assessment procedures, post-market monitoring, reporting of serious incidents, data management and modifications.
  • Effectiveness: Top management maintains the effectiveness of the quality management system and reviews it in the management review.

The quality policy is part of the documentation of the quality management system; the AI provider policy sets out the thirteen aspects of Article 17. With this policy, every member of staff acknowledges the quality policy. Deviations from the quality management system can affect the performance of an AI system, regulatory compliance and the risks of the organisation; they are reported and handled under the procedure for non-conformities.

5. AI Objectives (5.2)

[YOUR_ORGANISATION_NAME] sets objectives for the responsible use of AI that are consistent with this policy and measurable where practicable. The objectives translate the principles above into operational targets.

The following method applies to setting them. Objectives are derived from the principles of this policy and decided by top management. For each objective the organisation records the information need it serves, the type of measurement together with its target value or assessment interval, the data source of the measured value, the responsible person and the target date. Their state is maintained continuously in the AI management system and evaluated at least annually in the management review, which determines whether an objective is continued, adjusted or replaced.

The individual objectives and their current state are not part of this policy. They are maintained separately because they change at shorter intervals than the review cycle of this policy.

Where [YOUR_ORGANISATION_NAME] develops or operates its own AI systems, the organisation additionally sets its own objectives for responsible development, for example on the maintainability, data quality, robustness and security of the systems. These development objectives are anchored across the phases of the development life cycle and are maintained and reviewed together with the other AI objectives (ISO/IEC 42001, Annex A.6.1.2).

6. Actions to Address Risks and Opportunities (6.1.1)

[YOUR_ORGANISATION_NAME] takes the issues determined in the context analysis (4.1) and the requirements of interested parties (4.2) into account when planning the AI management system. Actions are determined to address risks and opportunities that could affect the ability of the management system to achieve its intended results, to prevent or reduce undesired effects and to achieve continual improvement.

In the language of ISO 31000 a risk is the effect of uncertainty on objectives, and that effect can go in either direction. Under that umbrella an opportunity would be a risk with a positive sign. Because this is unwieldy in day-to-day work, [YOUR_ORGANISATION_NAME] distinguishes between risks as possible negative effects and opportunities as possible positive effects. Both mean the same thing: the outcome is uncertain, and that uncertainty bears on the objectives of the organisation.

Opportunities of using AI are therefore considered deliberately and weighed against the risks. They include the benefit for which an AI system is introduced in the first place, as well as the improvements that follow from working through the risks: cleaner documentation of processes, clearer responsibilities and well-founded trust from customers and partners.

The risk environment of using AI follows from the external issues of the context analysis. Where no external issues have been determined there yet, the risk environment is set with the next round of the context analysis.

7. Risk Assessment and Risk Treatment (6.1.2, 6.1.3)

[YOUR_ORGANISATION_NAME] assesses the risks of using AI following a uniform procedure. For each AI system, the possible hazards are determined from the model type and a hazard catalogue. Impact and likelihood are rated on defined scales; the risk score follows from both values. The scale in use is held in the AI risk register; the methodology itself is a chapter of this policy and is approved together with it.

An acceptance threshold is defined for the risk levels, above which a risk is treated. The levels and the threshold are set in the AI risk register.

Each risk is assigned to a risk owner who is responsible for its assessment and treatment. The risk assessment is repeated annually and renewed on occasion for material changes, such as a new or materially changed AI system, a new legal situation or a result from an internal audit or the management review. Where a risk is rated higher or lower than initially determined, the re-rating is recorded with a comprehensible justification.

Risk treatment follows a fixed sequence. For each system, a decision is taken for each control on whether it is applied; where a control is not applied, the decision is justified. After the effect of the applied controls, the remaining residual risk is reassessed and checked against the acceptance threshold. Where the residual risk lies within the threshold, the risk owner confirms its acceptance; otherwise further measures are defined. The selection of the applied controls and the justifications for the exclusions are kept in the AI statement of applicability; the assessments and acceptance decisions per cycle are retained as a record of the risk treatment.

8. AI System Impact Assessment (A.5.2, A.5.3)

[YOUR_ORGANISATION_NAME] assesses, for every AI system in use, the possible consequences for individuals, for groups of individuals and for societies. The assessment follows a uniform procedure, accompanies the system across its entire life cycle and complements the risk assessment with the perspective of the people affected.

An impact assessment is carried out before a new AI system is used for the first time, and it is renewed as soon as the purpose or deployment context, the technology in use or the degree of automation, or the data types and data sources processed change materially. Independently of this, the impact assessment is repeated yearly.

The procedure covers the identification of the possible consequences with their sources and triggering events, the analysis by severity and likelihood, the evaluation against the defined thresholds, the definition of mitigating measures as well as the documentation and the reporting to management. For each system, the processes in which it is used, the period and frequency of use, the categories of affected persons and the risks of harm are recorded, together with the foreseeable misuse of the system and the effects on society and the environment.

In similar cases an impact assessment may build on an assessment already carried out for a comparable AI system or on the impact assessment of the provider. The entries taken over are checked against the organisation's own deployment context, completed and kept as an assessment of its own. Where it emerges during use that an entry no longer holds, the impact assessment is updated; every substantial change reported in the AI register is passed to the responsible person for review.

The results of the impact assessment feed the risk assessment. Every foreseeable misuse identified is assigned to a risk scenario and assessed there by severity and likelihood, so that consequences for affected persons and the risks of the organisation are handled in the same procedure.

The assessment is carried out by the person named for the impact assessment; the result is signed off by the AI officer, and by top management where the result lies above the acceptance threshold. The results are the basis for the decision to release the system or to discontinue its use, for the definition of mitigating measures and for the design of the human oversight. Where an impact assessment identifies a consequence above the acceptance threshold, the system is only used once mitigating measures have been defined and put in place.

The assessment addresses the affected groups of people and the wider society. For each system the affected groups are determined from a catalogue that includes those the standard names as requiring particular protection (children and minors, people with disabilities, older people, employees), and for each group it is recorded which impact areas are affected: fairness, accountability, transparency and explainability, security and privacy, safety and health, financial consequences, accessibility and human rights. In addition, the societal impact fields affected are determined: environmental sustainability, economic effects, government and public administration, health and safety, and norms, culture and values.

The results of every impact assessment are kept as a record per system and retained for [RETENTION_PERIOD, e.g. 10 years]. A repeated assessment creates a new version, and the previous one remains available for reference. The records are available for internal audits, the management review and information provided to authorities.

Where a data protection impact assessment already exists for the processing an AI system feeds into, the impact assessment draws on it: it cross-references the relevant sections or incorporates relevant parts, and complements them with the fundamental rights perspective. What is recorded there is not collected again: for each system the linked processing activities are drawn on with the status of their data protection impact assessment, the consultations held, the next review and the risks recorded, and their harm categories are taken over. The data protection documentation itself stays untouched; the link is kept in the AI management system alone.

Where the classification triggers the fundamental rights impact assessment as a legal obligation, the result is notified to the competent market surveillance authority. The status of that notification is kept per system, with the date and the location of the proof of submission; the authority is the same one named for reporting serious incidents. Where a notification is not required in an individual case, the reason is recorded on the system: the exemption under Article 46(1) or reliance on an existing assessment under Article 27(2).

For AI systems where the classification triggers the fundamental rights impact assessment under Article 27 of the AI Regulation as a legal obligation, the same procedure covers that obligation, and the systems concerned are marked accordingly. For the remaining systems, the impact assessment remains the instrument under ISO/IEC 42001, which applies irrespective of the risk class.

9. Resources (7.1)

[YOUR_ORGANISATION_NAME] determines and provides the resources needed for the establishment, implementation, maintenance and continual improvement of the AI management system. These include qualified staff with clearly assigned roles, the time needed to carry out those duties, and a budget for tools, training and external support.

The use of AI adds further resource types that are determined per system: the components of the AI system itself, the data resources used together with their origin and quality, the tooling resources including algorithms and models, the system and computing resources, and the human resources, meaning people and their expertise. The need is reviewed and adjusted on significant changes and at least annually in the management review.

For each AI system, the computing and system resources are recorded in the AI register, per life cycle stage: the hosting location and provider, the region and the type of compute (for example CPU, GPU or a managed cloud service), together with the target environment for operation, the memory and storage needed and the network requirement. A system that is trained locally and operated in the cloud therefore carries its own entries for both stages. The responsible party per stage follows from the parties of the same system and is not recorded separately. This record supports the assessment of availability, dependencies and the environmental footprint of the resources.

10. AI Literacy (7.2)

[YOUR_ORGANISATION_NAME] ensures that staff and other persons who operate or use AI systems on its behalf have a sufficient level of AI literacy. The level is appropriate to the role, to the technical knowledge, experience and training of the persons concerned and to the context in which the AI systems are used. This commitment rests on ISO/IEC 42001, Clause 7.2, under which the competence needed is determined, the required competence is ensured and evidence of it is retained. Art. 4 of the AI Regulation requires providers and deployers to take measures that support the development of AI literacy; the measures described here cover that requirement.

AI literacy covers an understanding of how the AI systems in use work, their capabilities and limits, the risks they carry and the rules of this policy and the AI Usage Policy. The need for AI literacy is determined per role; the resulting training is assigned, recorded and refreshed. Completion is tracked and reported to management.

For the provider roles it is evaluated how the following factors influence competence needs (EN 18286, 7.2.3): the intended purpose of each system and its reasonably foreseeable misuse; the nature of the AI technologies used and the data processed; the relationship between intended purpose, misuse and the risks identified in risk management, including significant effects on affected persons; and the effect of usability and accessibility given the diversity of intended users, including persons with disabilities.

From these factors, the knowledge certain roles need is derived per system, for example on the rights of those affected in an area of use, on detecting bias or on validation after retraining. The derivation draws on the AI register, the classification, the AI risk register, the datasets, where available the fundamental rights impact assessment, and the conformity documentation. In the competence overview it is recorded per topic which roles need the knowledge and whether it is present or being built; knowledge to be built is acquired through training, induction or external support. The general competences are assigned in the competence matrix, which for providers also covers development and validation, data governance, technical documentation and conformity, and post-market monitoring. Persons in provider roles know their duties in the quality management system and have access to the competence they need to understand the regulatory requirements and the intended purpose (7.2.2); this is evidenced by completion of the assigned training and by acknowledging the provider policy, which they confirm in addition to this policy.

11. Prohibited Practices and High-Risk Systems

11.1 Prohibited Practices (Art. 5)

[YOUR_ORGANISATION_NAME] does not use AI systems for the practices prohibited under Art. 5 of the AI Regulation, such as manipulative or exploitative techniques, untargeted scraping of facial images, social scoring, generating or manipulating intimate depictions of identifiable persons without their explicit consent and generating child sexual abuse material (both from 2 December 2026), risk assessment of individuals to predict criminal offences based solely on profiling or personality traits, real-time remote biometric identification in publicly accessible spaces for the purposes of law enforcement, or biometric categorisation and emotion recognition where the regulation forbids it. Before an AI system is approved, it is checked against the prohibited practices; a system that falls under a prohibition is not introduced or is withdrawn from use.

11.2 High-Risk Systems (Art. 26)

AI systems are classified by their use purpose to determine whether they are high-risk under the AI Regulation.

For each high-risk system the organisation operates, the deployer duties under Art. 26 apply. Their operational detail, namely use in accordance with the instructions, human oversight, input data, monitoring and reporting, logs and notifications, is governed by the AI Usage Policy in the chapter “Deployer Obligations for High-Risk Systems”.

Only for public authorities and bodies acting on their behalf; otherwise this paragraph is omitted.

As a public authority, or as a body acting on behalf of one, the organisation also meets the additional duties for the operation of high-risk AI systems, in particular registration under Art. 49 and entry into the EU database under Art. 71 of the AI Regulation.

Where the organisation acts as a provider of a high-risk system, the additional provider duties are met through the AI Provider Policy and the AIMS management documents.

The classification, the assigned role (provider or deployer) and the resulting obligations are recorded for each system in the AI inventory and the obligations catalogue.

12. Human Oversight (Art. 14 and 26)

For every high-risk AI system, [YOUR_ORGANISATION_NAME] assigns human oversight to a named person who has the competence, authority and resources to perform it (AI Regulation, Art. 26 in conjunction with Art. 14). The person responsible for oversight is able to understand the capabilities and limits of the system, to monitor its operation, to recognise signs of malfunction or unexpected output, to disregard or override an output and to stop the system where necessary.

Human oversight is set up so that an output of the system is not acted upon automatically without the opportunity for a person to review it where the use purpose so requires. The arrangements for oversight per system are documented alongside the classification.

13. Allocation of Responsibilities with Third Parties (A.10.2)

As a rule, several parties contribute to an AI system. For every system, [YOUR_ORGANISATION_NAME] records which party carries which phase of the life cycle and which part of the duties it takes on. A distinction is drawn between the parties providing data, those supplying a model or an algorithm, and those developing, integrating, operating and monitoring the system. The allocation is held in the AI register per system and is agreed with the parties involved.

Responsibility that is not expressly assigned to a party remains with [YOUR_ORGANISATION_NAME]. An agreement moves tasks; accountability towards authorities and affected persons is unaffected by it. Where a party takes on a duty, it is settled which evidence it supplies and at what interval.

Where an AI system processes personal data, the role in which [YOUR_ORGANISATION_NAME] acts is determined in addition: as controller, as joint controller, as processor, or in several of these roles across the life cycle. The allocation is agreed with data protection and held in the data protection impact assessment.

Where [YOUR_ORGANISATION_NAME] makes an AI system available to a third party, it is developed under the rules of this policy and handed over with the documentation the third party needs for its own duties. The transfer of the provider role and the agreements with suppliers of integrated components are governed by the AI provider policy.

14. Steering of AI Suppliers (A.10.3)

Where [YOUR_ORGANISATION_NAME] obtains services, products or materials that feed into an AI system, the procurement follows a defined path. The full range is covered: the single learning algorithm, the pre-trained model, training data, computing capacity, components and software libraries, as well as the complete AI system used standalone or embedded.

Before use it is settled which documents the supplier provides. These include the description of the system or model, the details on training data and the limits of its performance, the instructions for use and the information on security. This requirement is agreed contractually, as is cooperation on requests for information, technical access, support in operation, and notification of incidents and changes.

Selection and monitoring follow the type of supplier, the extent of its service and the risk arising from the system concerned. A supplier whose model works inside a high-risk system is steered more closely than a tool with no bearing on a decision about persons. Contribution, risk rating, state of documents, cycle and last assessment per supplier are held in the AI supplier register. The risk rating follows the highest risk class of the systems supplied, and the assessment cycle follows the rating.

For every system and component obtained, it is documented where it acts within our own environment, through which interfaces it is connected and which data flow in the process.

Where a system obtained does not work as intended, or affects persons, groups or society in a way that runs counter to these rules, [YOUR_ORGANISATION_NAME] requires corrective action from the supplier. The matter runs as an AI observation or AI incident and, where a cause is to be removed, as a corrective action. Where the remedy fails to arrive, use of the system is restricted or ended.

15. Development of own AI systems (A.2.2)

Only where the organisation develops AI systems itself or takes on the provider role; otherwise this section is omitted.

Where [YOUR_ORGANISATION_NAME] develops an AI system itself, or modifies a third-party system such that the provider role arises, the development follows a defined life cycle. The purpose of the system, the persons affected and the legal requirements are settled before anything is built; the classification under the AI Regulation happens at the outset and determines which duties apply to the undertaking.

A prohibited practice under Article 5 is not developed. Where the classification results in a high-risk system, the requirements of Articles 8 to 15 are part of the development requirements rather than a later check. Before first use, the impact assessment, the design of human oversight and the technical documentation are in place.

For data that feed into training, validation and testing, origin, suitability and quality are recorded. Every system is tested against the defined criteria before release and the results are retained. After release the behaviour in operation is monitored; deviations run as an AI observation or an AI incident.

How these rules are carried out is set out in the AI provider policy with the quality management system under Article 17; the evidence per system sits in the technical documentation. Responsibility for development is named in the AI roles.

16. Roles and Responsibilities (5.3)

Top management assigns and communicates the roles, responsibilities and authorities for the responsible use of AI (ISO/IEC 42001, Clause 5.3 and A.3.2). The following key roles are established:

  • Top Management: Bears overall accountability for the AIMS. Approves this policy, allocates the necessary resources and reviews the responsible use of AI.
  • Person responsible for AI: Coordinates the AIMS, maintains the AI inventory and the classification, oversees the obligations per system and is the point of contact for the responsible use of AI. This person is named in the AI roles and responsibilities.
  • AIMS operation: Runs the AI inventory, the classification and the obligations per system day to day and prepares the reporting on the responsible use of AI.
  • Persons responsible for human oversight: Perform the human oversight of the high-risk systems assigned to them and act on signs of malfunction or unexpected output.
  • Competence and training: Determines the AI literacy needs per role and ensures that the resulting training is assigned, recorded and refreshed.
  • Incident reporting: Operates the channel for reporting incidents, malfunctions and concerns regarding AI systems (A.3.3), handles incoming reports including anonymous ones, and forwards reportable serious incidents.
  • Supplier management: Manages the providers of the AI systems and models in use and the contractual security and AI requirements placed on them.
  • Fundamental rights impact assessment: Coordinates the fundamental rights impact assessment where it is required and is the interface to the data protection documentation.
  • AI system security: Owns the security of the AI systems in use across their lifecycle, including access protection, robustness against attacks and the protection of models and data.
  • Data protection liaison: Checks whether personal data in AI systems requires a data protection impact assessment and coordinates the use of AI with data protection management.
  • Risk and risk acceptance: Identifies, evaluates and treats AI-related risks and decides on the acceptance of residual risk.
  • Performance monitoring: Defines metrics for, measures and reviews the performance of the AI systems in operation and initiates action on degradation.
  • Legal compliance: Determines the legal requirements that apply to the AI systems and maintains the evidence of their ongoing fulfilment.
  • Reporting to leadership: Reports the performance and conformity of the AI management system to top management.
  • Data quality: Owns the suitability, representativeness and lawful provenance of the data used to train, validate and operate AI systems.
  • AI development: Owns the responsible development of the organisation's own AI systems across the lifecycle, including requirements, testing and release.
  • Business units: Report new or changed AI applications in their areas, comply with this policy and the AI Usage Policy and support the oversight of the systems they use.
  • All staff: Use AI in accordance with this policy and the AI Usage Policy and report incidents and weaknesses through the established channels.

The authorities of the individual roles are set out in the AI role catalogue.

The complete assignment of these responsibilities to specific persons is documented in the AI roles and responsibilities (RACI).

17. Reporting of concerns (A.3.3)

Anyone with a concern about the role of [YOUR_ORGANISATION_NAME] regarding an AI system reports it through the organisation's AI reporting channel ([AI_REPORTING_CHANNEL, e.g. reporting form or mailbox]). The route is open to employees and to contracted persons alike, that is freelancers, temporary staff and supplier personnel, and it applies across the whole life cycle of a system, from procurement to decommissioning.

The person reporting chooses between a confidential and an anonymous report. Anyone reporting anonymously leaves no address; the report is handled all the same, although a follow-up question is then not possible. Every report enters the process as an AI observation, is acknowledged where a return channel exists, and is answered within the time targets set for AI reports.

Persons who report and persons who investigate are protected from retaliation. Any disadvantage suffered because of a report made in good faith is not accepted at [YOUR_ORGANISATION_NAME] and is itself handled as an incident. Access to case data is limited to the persons charged with handling it.

Reports are handled by persons with the necessary competence and the authority to investigate the matter and initiate remedies. Where a report concerns a risk to health, safety or fundamental rights or remains unresolved, it is escalated to top management in good time.

18. Deviations and exceptions (A.2.2)

This policy applies to every use of AI at [YOUR_ORGANISATION_NAME]. Where a rule cannot be met in an individual case, a decision is taken and recorded; a silent workaround counts as a breach, not as an exception.

An exception is decided by the AI officer together with the person named for risk acceptance. Where the deviation concerns a high-risk system or a duty under the AI Regulation, top management decides. There is no exception from the prohibited practices of Article 5.

Every exception carries the system concerned, the rule departed from, the reason, the compensating measures, the deciding person and a time limit. It ends with that limit; an extension is a new decision. Exceptions are held where the other decisions on the system are held: as an accepted risk in the risk treatment or as the decision on the AI observation, in each case with justification and compensating measure.

The standing exceptions are an input to the management review. Where deviations from the same rule accumulate, the rule is reviewed rather than the exception extended.

19. Performance evaluation and improvement (9.1 to 9.3, 10.1, 10.2)

The effectiveness of the AI management system is measured, audited and reviewed. Measurement follows the AI objectives and the metrics held per system in the AI register; the measured value and its date are recorded there. Evaluation takes place at the interval defined for each metric.

19.1 Internal audit (9.2)

The AI management system is audited internally at planned intervals. The audit programme defines for each audit what is examined, against which criteria and with which objective; the objective is recorded in the audit plan and carried through to the report. The audit runs against this policy, against the requirements of ISO/IEC 42001 and against the duties of the AI Regulation.

Auditors do not audit their own work. When an auditor is assigned, the assignment is checked against the AI responsibilities; where that person carries responsibility for the area under audit, another person is appointed or an external audit is commissioned. Results are recorded with finding, evidence and classification, and reported to management.

19.2 Management review (9.3)

Top management reviews the AI management system at least annually. Inputs are the status of the AI objectives, the metrics, the results of internal audits, incidents and observations, open corrective actions, standing exceptions, feedback from interested parties and changes in the legal environment. Decisions are recorded with responsibility and deadline.

For providers of high-risk AI systems the review also measures the effectiveness of the quality management system by how it protects health, safety and fundamental rights. Criteria are the number and severity of serious incidents, the deviations in the post-market monitoring evaluations, the residual risks above the acceptance threshold and the timeliness of reports to authorities. Further inputs are: the effectiveness of the risk controls, the actions against the risks of the quality management system, the results of closed investigations, cases and exceptions in the procedures of the provider QMS, improvement recommendations and new or changed harmonised standards per system.

Frequency and depth of the review follow the risks of the systems. In addition to the annual review, a serious incident, a substantial modification of a high-risk AI system or a residual risk above the acceptance threshold leads to an extraordinary review. At closing, the quality and AI objectives with their target values and latest result are taken into the record of the review.

19.3 Nonconformity and corrective action (10.1, 10.2)

Where a nonconformity occurs, action is taken first to limit the consequence, and the cause is determined afterwards. For every corrective action it is checked whether a similar nonconformity already exists elsewhere or could occur; the result of that check is recorded in the case. The action itself is run with a responsible person and a deadline, and its effectiveness is reviewed after implementation.

Where an information security management system exists alongside, the corrective actions of both systems are managed together. Cases arise directly from incidents, observations, audit findings and management review results; their status is visible at any time. The continual improvement of the AI management system follows from these cases and from the decisions of the management review.

Corrective actions are appropriate to the effects of the nonconformities encountered. Evidence of the nature of the nonconformities, the actions taken and the results of the corrective actions is retained as documented information in the corrective action case.

20. Confidentiality

In applying the AI Regulation, the organisation comes into possession of third-party intellectual property, trade secrets and source code. This concerns three cases (Article 78): documents and information from providers, importers and notified bodies, test results and source code from conformity assessments, and information from contact with authorities.

Such information is shared only with named persons who need it for their task. It is held in the AI records at the appropriate classification level and is not stored on systems outside the approved environment. Once the purpose has lapsed, it is returned or deleted.

21. Reference Documents

This policy is put into practice through the following documents, which set out the operational rules and records:

  • AIMS Scope Statement: Defines the boundaries and applicability of the AI Management System.
  • AI Roles and Responsibilities (RACI): Assigns each AI responsibility to a named person.
  • AI Usage Policy: Sets out the day-to-day rules for staff on how AI may and may not be used, including the use of approved AI services and the handling of inputs and outputs.
  • AI inventory and obligations catalogue: Record the AI systems in use, their classification, their assigned role and the resulting obligations per system.
  • AI incident reporting: Provides the channel for reporting incidents and malfunctions of AI systems.

22. Changes to the AI Management System (6.3)

[YOUR_ORGANISATION_NAME] carries out changes to the AI management system in a planned manner. A change is triggered in particular by a new AI system in the register, a change of role towards a system, a changed scope, a new legal situation, and results from an internal audit or the management review. Before a change is implemented, its purpose, its possible consequences, the resources required and the responsible person are recorded. Top management ensures that changes preserve the integrity of the management system.

This is distinct from the declaration of substantial modifications to an individual high-risk system under Article 43(4) of the AI Regulation, which is kept per system in the AI register.

23. Relief Based on Company Size

Only where a relief based on company size applies (Art. 62 for micro-enterprises and SMEs, simplified quality management under Art. 63, small mid-caps); otherwise this section is omitted.

The organisation meets the elements of the quality management system in a simplified manner under Article 63(1); the relief presupposes that no partner enterprises and no linked enterprises within the meaning of Recommendation 2003/361/EC exist. It concerns only elements of the quality management system; under Article 63(2) the remaining obligations remain unchanged, including the requirements of Articles 9 to 15 and the reporting and monitoring duties of Articles 72 and 73. The organisation continues to meet the substantive duties of this policy in full.

24. Review and Maintenance

[YOUR_ORGANISATION_NAME] reviews this policy at least annually and whenever material changes occur, such as the introduction of new AI systems, changed use purposes, new legal requirements or changed application stages of the AI Regulation. Each review and each change is versioned and documented with a justification. The person responsible for AI prepares the review and proposes changes; top management approves them.

Sources

Frequently asked questions

Do I need an AI policy if we only use AI tools we have bought in?

The AI Act does not prescribe a document with this name. As soon as AI tools are used professionally, however, the organisation is a deployer and has to ensure a sufficient level of AI literacy under Art. 4. The policy is the simplest way to regulate this in a verifiable manner. For ISO/IEC 42001 certification it is mandatory (clause 5.2); the sections on in-house development and the quality policy are then omitted.

How do the AI policy, the AI usage policy and the AI provider policy fit together?

The AI policy is the governing document and addresses management and the responsible persons. The AI usage policy turns it into rules for all employees. You only need the AI provider policy if your organisation places AI systems on the market or puts them into service as a provider; it is the quality management system under Art. 17.

Can I merge the AI policy with the information security policy?

The policy is aligned with the other policies (ISO/IEC 42001, A.2.3) and refers to information security and data protection. As a document of its own it can be approved and reviewed separately, which makes the evidence easier in a certification audit. Shared procedures such as corrective actions and the management review can serve both management systems.

No Cookies!

This wiki collects nothing, bakes nothing and leaves nothing behind. There's nothing to consent to. Privacy doesn't get better than this.