Much of the AI that organisations use today does not fall under the heavy rules for high-risk systems: a summarisation tool for civil servants, a chatbot on the website, an assistant that searches policy documents. For those systems the AI Act does not require a risk management system, technical documentation or human oversight. Organisations that want to arrange those things anyway can draw up a code of conduct for it. That is what Article 95 of the AI Act makes possible, and what the AI Office and the Member States must encourage.
Suppose six municipalities in a region want to agree joint rules for the AI tools their staff and residents use. Do they have to? What does a code give them, and what does it not? After that we look at a bank, an insurer, education and HR. The examples are fictional.
In short
Article 95 sits in Chapter X of the AI Act (codes of conduct and guidelines). It places the obligation on the AI Office and the Member States: they encourage and facilitate the drawing up of codes of conduct. Organisations themselves do not have to draw up or sign a code.
A code can do two things. Paragraph 1: the voluntary application of some or all of the requirements of Chapter III, Section 2 (Articles 8 to 15, such as risk management, data governance, record-keeping, transparency, human oversight and accuracy) to AI systems that are not high-risk. Paragraph 2: voluntary commitments for all AI systems, including by deployers, based on clear objectives and key performance indicators, for example on ethics, environmental sustainability, AI literacy, inclusive design and vulnerable groups.
Article 95 attaches no presumption of conformity, no approval and no fine of its own to a code, and a code does not replace any legal obligation. The Digital Omnibus on AI (Regulation (EU) 2026/1744) amended only paragraph 4: the interests of small mid-cap enterprises now count as well. Article 95 has applied since 2 August 2026 (Article 113).
This article is about codes of conduct under Article 95. They are different instruments from the code of practice for general-purpose AI models under Article 56 (see the code of practice for general-purpose AI and which tech companies signed it) and the code of practice on transparency of AI-generated content under Article 50 (see the EU code of practice for AI content). The difference is set out below.
What Article 95 does, and what it does not
Article 95 is addressed to the AI Office and the Member States. Paragraph 1 says they "shall encourage and facilitate" the drawing up of codes of conduct, paragraph 2 that they "shall facilitate" it. There is no obligation for providers or deployers to draw up or follow a code. Recital 165 gives the aim: developing AI systems that are not high-risk in accordance with the requirements of the Regulation may lead to a larger uptake of ethical and trustworthy AI in the Union, and providers "should be encouraged" to create such codes.
What Article 95 does not do:
- No obligation. The title speaks of "voluntary application". Having no code does not infringe Article 95; the duty to encourage rests with the AI Office and the Member States.
- No presumption of conformity. A presumption of conformity with the requirements for high-risk systems arises, among other things, from harmonised standards (Article 40), common specifications (Article 41) and the cases of Article 42. Article 95 does not attach such an effect to a code of conduct. If a system does fall under the high-risk rules, it must meet them, code or no code.
- No approval or register. Unlike codes of conduct under the GDPR, Article 95 provides for no approval by a supervisory authority, no register and no body that monitors compliance with the code. Paragraph 1 does mention "related governance mechanisms": the code itself can lay down who monitors compliance.
- No fine of its own. Article 95 imposes no obligation on providers and deployers and does not appear in the fine provisions of Article 99(3) to (5). In our reading, failing to follow a voluntary code is therefore not in itself an infringement of the AI Act. A code may of course contain commitments that are also legal obligations, such as the chatbot disclosure in Article 50(1); that obligation then applies regardless of the code. And what an organisation promises in a code may bind it in other ways, for example in a contract or, towards consumers, through the rules on unfair commercial practices (Article 6(2)(b) of Directive 2005/29/EC): a trader that indicates it is bound by a code and then fails to honour a firm, verifiable commitment in it may be acting misleadingly. The AI Act does not regulate that.
Whether a code that was followed counts when a fine is set for another infringement, the AI Act does not say. The list of circumstances in Article 99(7) does not mention codes of conduct; point (g) does cover "the degree of responsibility of the operator taking into account the technical and organisational measures implemented by it". In our reading, a code that was actually implemented can serve there as evidence of measures, but that is not settled.
Paragraph 1: applying the high-risk requirements voluntarily
Paragraph 1 concerns AI systems that are not high-risk. A code can provide that "some or all" of the requirements of Chapter III, Section 2, are applied anyway, "taking into account the available technical solutions and industry best practices". Section 2 consists of Articles 8 to 15; Article 8 governs compliance with the requirements as such, Articles 9 to 15 contain the requirements:
| Article | Requirement | What a code could take from it |
|---|---|---|
| 9 | Risk management system | A short risk analysis per application, repeated yearly |
| 10 | Data and data governance | Rules on which sources a system may use |
| 11 | Technical documentation | A model or system card per application |
| 12 | Record-keeping | Keeping inputs and outputs for later review |
| 13 | Transparency and provision of information to deployers | Instructions for use stating limits and known errors |
| 14 | Human oversight | Who checks the output, and what is never automated |
| 15 | Accuracy, robustness and cybersecurity | Testing before go-live and after updates |
The right-hand column gives our examples, not requirements of the law. Recital 165 says the requirements can be "adapted in light of the intended purpose of the systems and the lower risk involved", and mentions "model and data cards" as an example of best practice. A code therefore does not have to copy the high-risk requirements one to one.
Paragraph 2: own objectives for all AI systems
Paragraph 2 is wider. It concerns the voluntary application "of specific requirements to all AI systems", so high-risk systems as well, and expressly "including by deployers". Such a code is based on "clear objectives and key performance indicators to measure the achievement of those objectives". The Regulation lists five elements, "but not limited to":
- Ethics (a): applicable elements of the Union ethical guidelines for trustworthy AI.
- Environment (b): assessing and minimising the impact of AI systems on environmental sustainability, including as regards energy-efficient programming and efficient design, training and use.
- AI literacy (c): promoting AI literacy, in particular of persons dealing with the development, operation and use of AI.
- Inclusive design (d): an inclusive and diverse design, including through inclusive and diverse development teams and stakeholder participation.
- Vulnerable groups (e): assessing and preventing negative impacts on vulnerable persons or groups, including as regards accessibility for persons with a disability, as well as on gender equality.
On point (c): Article 4 already requires providers and deployers to take measures that support the development of AI literacy. A code can describe how you do that and how you measure whether it works. The code does not replace the Article 4 obligation; in our reading, a code you actually implement does help to show which measures you have taken. See also demonstrating AI literacy.
Paragraphs 3 and 4: who draws up a code
A code can be drawn up by individual providers or deployers, by organisations representing them, or by both (paragraph 3). Interested stakeholders and their representative organisations, including civil society organisations and academia, can take part. A code can cover one or more AI systems, "taking into account the similarity of the intended purpose of the relevant systems". A single company, a sector association or a partnership of municipalities can therefore make a code.
Paragraph 4 asks the AI Office and the Member States to take into account the specific interests and needs of SMEs, including start-ups, and since the Digital Omnibus also of small mid-cap enterprises (SMCs). The Omnibus replaced paragraph 4 for that purpose (Article 1(35) of Regulation (EU) 2026/1744). What an SMC is follows from Article 3, point (14b): the definition in point (2) of the Annex to Commission Recommendation (EU) 2025/1099. The rest of Article 95 was not amended.
Do not confuse: codes of conduct and codes of practice
The AI Act distinguishes two terms: a code of conduct under Article 95 is something other than a code of practice under Article 56 and Article 50. In the Dutch text the difference is "gedragscode" against "praktijkcode".
| Code of conduct, Article 95 | Code of practice for AI models, Article 56 | Code of practice for AI content, Article 50(7) | |
|---|---|---|---|
| Subject | Voluntary application of requirements to AI systems | Obligations of providers of general-purpose AI models (Articles 53 and 55) | Detection, marking and labelling of artificially generated or manipulated content |
| Who encourages | AI Office and Member States | AI Office | Commission |
| Assessment of the code | None; the Commission evaluates the instrument (Article 112(7)) | Commission assesses and publishes whether the code is adequate (Article 56(6)) | Commission assesses whether adherence to the code is adequate to ensure compliance with the obligations in paragraphs 2 and 4 (Article 50(7)) |
| Legal effect | None in the AI Act | Providers may rely on it to demonstrate compliance until a harmonised standard is published (Article 53(4)) | No presumption of conformity (recital 41 of Regulation (EU) 2026/1744) |
The codes of practice do not grant a presumption of conformity either. Recital 41 of Regulation (EU) 2026/1744, the Digital Omnibus, states that the codes of practice referred to in Article 50(7) and Article 56(6) "have limited legal effect, and in particular do not grant a presumption of conformity". For a code of conduct under Article 95 the Regulation provides even less: there is no assessment of an individual code by the Commission.
Evaluation by the Commission
Article 112(7) provides that by 2 August 2028 and every three years thereafter, the Commission evaluates the impact and effectiveness of voluntary codes of conduct to foster the application of the requirements of Chapter III, Section 2, and possibly other additional requirements for AI systems that are not high-risk, "including as regards environmental sustainability". That concerns the instrument as a whole, not the approval of a code. How the Commission will go about it and which codes it will include is not yet settled.
Codes of conduct under the GDPR: a different instrument
Anyone working in privacy knows codes of conduct from Article 40 GDPR. That is a different instrument:
- Approval. Associations and other bodies representing categories of controllers or processors may prepare a code and submit the draft to the competent supervisory authority, which approves it if it provides sufficient appropriate safeguards (Article 40(2) and (5) GDPR). Approved codes are registered and published (Article 40(6) and (11) GDPR).
- Monitoring. A code contains mechanisms for mandatory monitoring of compliance by an accredited body (Article 40(4) GDPR and Article 41 GDPR). That monitoring by an accredited body does not apply to processing carried out by public authorities and bodies (Article 41(6) GDPR).
- Legal effect. Adherence to an approved code may be used as an element to demonstrate compliance (Article 24(3) GDPR and Article 32(3) GDPR), compliance is taken into account in a DPIA (Article 35(8) GDPR) and when a fine is set (Article 83(2)(j) GDPR).
Article 95 has none of these three. A code under Article 95 is therefore not an approved code within the meaning of Article 40 GDPR and, in our reading, does not count under Article 35(8) GDPR. If you want a code to work under the GDPR as well, it must follow the Article 40 GDPR route; that is a separate procedure with the data protection authority.
Main example: six municipalities draw up a code
Six municipalities in a region work together in a fictional partnership. Their staff use a summarisation tool for long council documents and consultation responses, and the municipalities jointly run a chatbot on their websites that points residents to forms and opening hours. Neither system decides on a benefit or a permit.
Classification first. An AI system intended to be used by or on behalf of public authorities to evaluate whether someone is eligible for essential public assistance benefits and services is in principle high-risk (Annex III, point 5(a)). The summarisation tool and the signposting chatbot do not do that. If the municipality starts using the summarisation tool to assess applications for special assistance, the picture changes, and a code does not help. In our reading the municipality may then itself become the provider of a high-risk system, because it modifies the intended purpose (Article 25(1)(c)), unless an exception of Article 6(3) applies, for example for a purely preparatory task. The high-risk obligations for Annex III apply from 2 December 2027. A good code therefore states expressly what the systems may not be used for.
What is already mandatory. Two things apply without any code. If the municipalities have the chatbot developed and put it into service under their own name, they are the provider, and since 2 August 2026 residents must be informed that they are interacting with an AI system, unless that is obvious (Article 50(1)). And each municipality must take measures for the AI literacy of its staff (Article 4). If the systems process personal data, the GDPR applies, with a DPIA where required (Article 35 GDPR).
What the code adds voluntarily (paragraph 1). In our reading the partnership can draw up the code as an organisation representing the municipalities (paragraph 3). It picks from Section 2:
- a short risk analysis per application, repeated yearly (after Article 9);
- a system card per application: purpose, sources, known errors and who owns it (after Articles 11 and 13);
- keeping chat conversations for a fixed period for quality review, within the limits of the GDPR (after Article 12);
- a fixed rule that a summary never ends up in a decision without the original being read (after Article 14);
- a test set of residents' questions that is rerun after every update (after Article 15);
- a joint committee that checks every year whether the commitments are being met (a governance mechanism within the meaning of paragraph 1).
Objectives and indicators (paragraph 2). The code sets one measurable indicator per objective:
- AI literacy (c): every staff member who uses the summarisation tool completes a short role-based instruction; indicator: the share of users who have completed it.
- Vulnerable groups and accessibility (e): the chatbot answers in plain language and always offers a route to a staff member; indicator: the share of conversations in which a resident asks for a person and is helped within one working day.
- Environment (b): the municipalities include energy use in procurement; indicator: the share of contracts in which the supplier reports that use.
Who takes part. Paragraph 3 allows stakeholders to be involved. The partnership asks a client council and a regional university of applied sciences to comment on the draft.
What the code does not deliver. No presumption of conformity, no approval by a supervisory authority and no exemption from Article 4, Article 50 or the GDPR. It does deliver a shared standard for six organisations, a record of measures taken and a framework that is ready if an application later turns out to be high-risk after all.
Example 2: the bank and the insurer
A fictional association of medium-sized banks draws up a code for AI its members use outside the high-risk categories: an internal knowledge assistant for staff and systems that detect financial fraud. Fraud detection is expressly excluded from the high-risk point on creditworthiness (Annex III, point 5(b)). A system that evaluates the creditworthiness of natural persons or establishes their credit score is high-risk, and there a code does not replace the obligations. For the fraud systems, the code can provide that every block is reviewed by a person within a fixed period (after Article 14) and that error rates are measured per customer group (paragraph 2(e)). Two things remain plain obligations here. A block may be a decision based solely on automated processing that significantly affects the customer; the rules of Article 22 GDPR then apply, including the right to obtain human intervention (paragraph 3), and that is not a voluntary commitment. And measuring per customer group is only possible with data the bank may process for that purpose; special categories of personal data are subject to the prohibition with exceptions of Article 9 GDPR.
An insurer uses AI to sort claims under home insurance policies. Risk assessment and pricing in the case of life and health insurance are high-risk (Annex III, point 5(c)); sorting home insurance claims is not on that list. The insurer can draw up its own code as an individual provider or deployer (paragraph 3) or join a code of a sector association.
Briefly on education and HR
Education. AI for admission, for evaluating learning outcomes or for monitoring tests can be high-risk (Annex III, point 3). A tool teachers use to prepare course material does not, in our reading, fall under it. A code for such a tool can provide that generated material is always checked by the teacher.
HR. AI for recruitment and selection, for decisions on terms of work, promotion and termination, for task allocation or for monitoring and evaluating performance and behaviour is in principle high-risk (Annex III, point 4; Article 6(3) has exceptions). An internal HR chatbot that answers questions about leave rules is not, in our reading. A code can provide that the chatbot does not handle questions about individual assessments and always refers to an HR adviser.
From when?
- Article 95 sits in Chapter X. That chapter is not among the exceptions in Article 113, third paragraph, and has therefore applied since 2 August 2026 (Article 113, second paragraph).
- The new paragraph 4 (small mid-cap enterprises) entered into force on 27 July 2026: Regulation (EU) 2026/1744 was published on 24 July 2026 and entered into force on the third day after that (Article 4 of Regulation (EU) 2026/1744). Because Article 95 itself has only applied since 2 August 2026, it has applied in its amended form from the first day.
- Articles 4 and 50 already apply: Article 4 since 2 February 2025, Article 50 since 2 August 2026 (for marking AI-generated content by systems that were already on the market before that date, the period runs until 2 December 2026, Article 111(4)).
- The high-risk obligations apply from 2 December 2027 for Annex III and from 2 August 2028 for Annex I (Article 113, third paragraph, point (c)). Applying the Section 2 requirements voluntarily now to systems that are not high-risk therefore means working with rules that apply to high-risk systems only later.
- The evaluation of voluntary codes of conduct by the Commission follows by 2 August 2028 and every three years thereafter (Article 112(7)).
What you do now
- Classify first. Put your AI applications in a register and decide per application whether it is high-risk. A code is for what remains, not for what is already mandatory.
- Separate mandatory from voluntary. Write down separately in the code what Article 4, Article 50 and the GDPR already require. That keeps it visible that those obligations apply even if the code changes.
- Pick from Section 2 what fits. Do not copy everything. A system card, a test set and a rule for human review are a good start for most applications.
- Make objectives measurable. Paragraph 2 asks for objectives and key performance indicators. Choose per objective one indicator you can really measure, and record who tracks it.
- Cooperate where you can. An umbrella body, a partnership or a sector association may draw up a code (paragraph 3). That saves work for smaller organisations.
- Do not promise more than you do. A code that exists on paper but is not implemented is no evidence of measures, and what you promise residents or customers may bind you in other ways.
- Want a GDPR code? Then follow the Article 40 GDPR procedure with approval by the supervisory authority. Such a code can only be drawn up by an association or other body representing a category of controllers or processors (Article 40(2) GDPR), not by a single organisation. A code under Article 95 is not enough for that.
Want to know which AI Act and GDPR obligations come into view for an application and whether it is high-risk? With the case check you describe the application in a few questions.
Frequently asked questions about codes of conduct under Article 95
Sources
Newsletter
Every Tuesday, the AI Act week ahead in 5 minutes
A practical briefing on deadlines, new guidance and enforcement, so you know what matters this week. No spam and you can unsubscribe in one click.
Practical and short · No spam · One-click unsubscribe