AIEI-STD-RAIP-01
Responsible AI Professional Standard
Requirements for people who develop, deploy, manage or use AI systems
Version 1.0 · Published · 10 November 2026
This standard sets out twenty-one duties for responsible judgment and conduct in AI-related work. For each duty it states what a person is expected to do, how much the duty asks of them given their role and the risk, and what evidence can show that it has been met.
Status of this edition
Version 1.0, published on 10 November 2026, is the first published edition of this standard. It replaces working drafts 0.1 to 0.9. Later editions follow their own review cycle, and each change is listed under Document control.
Document control
| Version | Change |
|---|---|
| 0.1–0.3 | Initial requirements, scope reconciliation and editorial structure. |
| 0.4 | Added the three explanatory notes to every requirement. |
| 0.5 | Clarified duties and limits, added risk-based application rules, completed assessment coverage and revised references. |
| 0.6 | Clarity revision. Requirements and notes rewritten in plainer language without changing what they require; the one substantive change is the fallback added to 8.1. Normative statements that had been left in informative notes moved into the requirements they belong to. Legal limits consolidated in 1.3; the legal-restriction rule moved from section 1 to 2.5. A short panel of specially used words added at 3.3. Definitions of “professional”, “role”, “escalation”, “scheme”, “consequential” and “” added to section 11. Section 12 restructured and extended: job analysis, assessment integrity including the use of AI in assessment, impartiality and scheme governance, lawful alternative evidence, and group-level critical failures. Annex B (informative reader’s guide by role) added. Section numbering unchanged. |
| 0.7 | Practical application revision. Clarified proportionate checks, shared decision records, and action where evidence is missing. Added short examples of role-specific action and acceptable assessment decisions. Refined reporting safeguards and stated AIEI’s intention to offer consultation, with support arrangements still to be settled. Aligned explanatory notes with requirements and softened overly categorical wording. Structure and duty numbering unchanged. |
| 0.8 | Practical assurance revision. Limits of the standard made binding in new 2.6; new 2.7 on working within the organisation’s arrangements; four new duties: 5.3 organisational AI rules and approved tools, 6.5 security of AI systems, 6.6 control of actions taken by AI agents, 8.3 AI incidents; 9.2 and 9.3 reduced to an internal escalation sequence; personal records removed from 8.1 and 8.2; Evidence lines under the duties removed in favour of 12.10; scheme ownership, recertification, AI use in assessment and protection of assessment added to section 12. |
| 0.9 | Consolidated round 1 review. Clarified proportionate security testing, agent autonomy, individual approval and controls where a platform cannot record or stop actions. Distinguished serious near misses from expected errors caught by adequate checks. Aligned escalation thresholds and senior decision-maker responsibilities; added secrecy safeguards, assessment protections and independent assessment of authors and reviewers. Resolved the ISO/IEC 17024 reference exception; aligned explanatory notes, updated references and removed revision colours. The job and task analysis remains open under 12.2. |
| 1.0 | First published edition, after review round 2. Scope in 2.1 extended to everyday use of AI systems at work; 1.1 states that the duties complement organisational arrangements; 2.5 settles what is recorded where law prevents an action and no lawful alternative exists; 4.2 limits referral to legal questions beyond the ’s competence and bars unjustified approval or reliance meanwhile; new 6.4 Requirement 4 on resourcing oversight; 6.6 states when one approval can cover a bounded operation; 7.2 sets what a person is told about a decision and how they can challenge it; the definition of “AI system” aligned with the OECD and AI Act wording. Duty numbering unchanged. |
How to read this edition
The standard is reproduced word for word. Everything added here is navigation, and it follows the division section 3.1 draws between text that binds and text that explains.
The standard’s own text is set at the same size throughout, whether the section is binding or informative – the chip above each heading says which.
Emphasised only inside genuinely normative text: sections 2, 3, 10, 11 and 12, and the paragraphs labelled Requirement in sections 4 to 9. It never appears in section 1 or the annexes, which the standard itself marks as informative.
The three notes under every duty are informative. They are set smaller, on a tint, behind their own tabs. A note never overrides the requirement above it.
A word the standard defines in section 11. Select it anywhere in the text to read the definition without leaving the sentence.
Every section header states what force its text carries, as section 3.1 sets it out.
The twenty-one duties
All twenty-one apply to everyone within the scope in 2.1 whenever their conditions arise. There is no reduced set for any role. What changes with the role is how much each one asks, under 2.2 – see Annex B for how that plays out for particular kinds of work.
4 · Care for affected people
5 · Competence
6 · Responsible practice
7 · Honesty and openness
8 · Accountability
9 · Integrity under pressure
1Purpose
1.1What this standard is for
This standard describes the judgment and conduct expected of an individual working with AI. These duties complement an organisation’s policies, controls and management systems, and apply whether or not such arrangements exist (2.7). This standard covers what the person is expected to do – including the duty to question a result, to acknowledge a limitation, and to act on a credible concern.
1.2What a certificate against this standard would mean
A certificate based on this standard would attest that a person demonstrated assessed competence and judgment within a stated role scope, at the time of assessment. It would not guarantee that person’s future conduct, and it would not guarantee the safety of any system they use.
1.3What this standard does not do
This is a private standard. It is not law, and it does not change law. Its limits, including that it confers no legal protection on a person who raises or escalates a concern, are set out as binding text in 2.6.
What the standard does require is enough technical and domain understanding for a person to recognise the limits of their own role and to act safely within them.
2Scope and proportionate application
2.1Who this standard applies to
This standard applies to a person who, in the course of their work:
- develops, trains, tests or maintains an ;
- selects, procures or integrates one;
- decides that one will be used;
- uses one, including by providing it with information or by generating, reviewing, relying on or acting on its outputs;
- configures, operates or delegates tasks to an (section 11); or
- oversees, audits, advises on or governs any of those activities.
It covers relevant work throughout the system’s life, including changes and withdrawal from use. Annex B is an informative guide to where each kind of role usually meets these duties first. The duties govern a person’s conduct when their conditions arise in that person’s work. What a certificate confirms is stated in 12.3 and 12.9.
2.2How much each duty asks of you
Requirement All twenty-one duties apply whenever their stated conditions arise. A shall apply each one according to the harm reasonably foreseeable from the work, their role in it, and the access and authority they actually have. In judging foreseeable harm they shall consider severity, likelihood, scale, reversibility and the vulnerability of the people affected. Low seniority does not by itself mean low risk.
2.3What has to be written down
Requirement For work that could affect safety, rights, access to essential services or livelihoods, a shall record the risk basis, the checks and safeguards used, and any unresolved limitation. For less work, an explanation and existing work records may be enough, unless a particular duty expressly calls for a record. A professional may rely on their organisation’s existing evidence where it is relevant, current and understood by them; a separate parallel process of their own is not required. A single record may serve more than one duty where it contains the evidence each requires.
2.4When access or authority is missing
Requirement Being unable to check something shall not be treated as evidence that it is acceptable. A who cannot complete a check the work needs shall seek the information or competent support and, until they have it, shall limit their own reliance on the result as 6.2 requires. Where the gap could lead to , 9.2 applies.
2.5When law prevents the action
Requirement Where a legal restriction prevents an action this standard would otherwise require, a shall seek a lawful alternative proportionate to the urgency. They shall record the restriction, and the absence of a lawful alternative where there is none, so far as the law permits. The restriction does not remove the duty; it changes what discharging the duty looks like, and 6.2 and 9.2 continue to apply. Nothing in this clause requires an unlawful disclosure or an unlawful record. Where the restriction arises from legal or medical professional privilege or a statutory duty of professional secrecy, the professional is not required to seek an alternative that would disclose the protected information. Nothing in this standard requires, or shall be read as permitting, the disclosure of information so protected.
2.6What this standard does not do
Requirement This is a private standard. It is not law and does not change law. It does not:
- certify a system, product or organisation;
- establish a general qualification in AI engineering or data science;
- create legal permission to do anything, or remove any legal duty;
- determine civil liability, or set a standard of care for any court;
- discharge, satisfy or evidence an obligation that law places on a provider, deployer, employer, controller or any other legal actor;
- confer any legal protection on a person who raises or escalates a concern under 9.2 or 9.3, or who reports a matter outside the organisation.
Applicable law prevails at all times. A lawful activity may still fail this standard, and conformity with this standard is not an answer to a legal claim. Where a legal restriction prevents an action this standard would otherwise require, 2.5 applies.
2.7Working within the organisation’s arrangements
Requirement Where the organisation for which a works has an AI policy, procedures or controls that apply to the work, for example as part of an AI management system or an information security management system, the professional shall meet the duties in this standard through those arrangements wherever they provide what the duty requires. Records, assessments and reporting channels shall be those of the organisation; a parallel personal process is not required, and shall not be created where it would place confidential or personal information outside the organisation’s systems. Where an arrangement is missing, insufficient for the risk, or would require an action that conflicts with a duty in this standard, the professional shall raise this with the person responsible, and 9.2 applies where its conditions are met.
An organisational instruction does not by itself discharge a duty in this standard, and a duty in this standard does not authorise a to bypass an organisational control.
3How to read this standard
3.1What is binding and what is not
“Shall” states a requirement. “Shall not” states a prohibition. “May” states something that is permitted, not required. Binding text is written in the third person (“a shall”); the notes address the reader directly.
Binding: sections 2 and 3, which set the scope and application rules; the paragraphs labelled “Requirement” in sections 4 to 9; section 10, which governs how the references are read; section 11, which defines the terms used here; and section 12, which states what a claiming to assess this standard must do.
Not binding: section 1, which explains the standard’s purpose and limits and contains no requirements; the three notes under each duty; the “Sources” line under each duty, which records provenance only; and Annexes A and B.
3.2The three notes under each duty
In practice gives examples, drawn from more than one kind of role. They illustrate the duty; they are not a checklist, and an assessor may accept equivalent evidence.
Met when describes evidence that could show the duty has been met. It helps you prepare for assessment, but it is not a separate pass criterion. The publishes its pass criteria under 12.6; those criteria must assess the binding requirements, and the examples here do not exclude equivalent evidence.
Does not require states a limit that is already present in the requirement or in section 2. It restates the limit for clarity; it does not create an exemption.
If a note conflicts with binding text, the binding text governs.
A shall not treat as a failure anything that a “Does not require” note excludes, unless the binding text requires it. Examples of suitable assessment evidence are given in 12.10 only; they are not repeated under the duties.
3.3Words used here in a particular way
Several ordinary words carry a defined meaning in this standard, among them professional, , , credible, serious harm, unsafe use, due care and . Section 11 is the only place where terms are defined, and it is binding. A reader new to the standard should read section 11 before sections 4 to 9.
3.4Finding the duties that apply to you
All twenty-one duties apply when their conditions arise. There is no reduced set for any role, and no duty is reserved to specialists. What changes with the role is how much each duty asks, under 2.2. Annex B is an informative guide to where each kind of role usually meets these duties first, and to the ones people in that role most often overlook.
4Care for affected people
Recognise who is affected
A shall identify the people and groups reasonably foreseeable as affected by their AI-related work, including those who are not clients or users. They shall give to those people’s safety, rights and legitimate interests. Where their own knowledge of the impacts is insufficient for the risk involved, they shall seek relevant perspectives or evidence.
A recruitment tool affects every applicant it screens, and none of them is its user. The manager who decides to buy it is within this duty as squarely as the engineer who builds it, and so is the adviser whose opinion clears its use, because the advice shapes who is exposed. Evidence about impacts can come from complaints data, domain expertise, published research or direct consultation. One ’s intuition about other people’s circumstances is not a substitute for any of these.
The can name the affected groups, including any who are not users, and explain how their interests changed a decision or an assessment. Where their own understanding was incomplete, they can show what they did about that.
Contacting every , satisfying every preference, or assuming a legal duty beyond applicable law.
Sources [OECD] 1.1–1.2; [NSPE] I.1.
Resolve competing duties openly
A shall give priority to ’s safety and rights over commercial, schedule, reputational or convenience pressures. Where legitimate interests genuinely conflict, they shall compare the feasible options, including not using AI at all, and shall consider who bears the harm that remains. They shall record the reasons for their recommendation. Acceptance of residual risk shall be referred to a person authorised to decide; a professional who holds that authority shall record the decision and its basis. An unresolved unsafe use engages 9.2. Neither approval nor commercial benefit excuses an unsafe use as defined in section 11, or conduct that the professional knows, or has been advised by a competent person, to be unlawful. Where a material question about whether the work is lawful exceeds the professional’s competence, they shall refer it to a competent person, such as legal, compliance or data protection, rather than decide it alone. Until the question is resolved, they shall not give an approval, or rely on a result, that they cannot justify.
A lower false-positive rate usually means more missed cases. The choice needs an explicit account of who is worse off under each option and of the constraints that apply, not a claim that one metric proves fairness. A procurement officer facing a fixed go-live date is in the same position as the model developer: the deadline is a pressure to be weighed, not a reason that settles the question.
The record identifies the conflict, the options compared, the risk that remains, and the person who took the decision.
Zero risk, ignoring costs, or settling a disputed policy question beyond the ’s authority.
Sources [NSPE] I.1; [NIST] 1.2.2–1.2.3, MANAGE 1.4. The decision procedure is an AIEI requirement.
5Competence
Work inside your competence
A shall understand the purpose, limitations and relevant failure conditions of the they work with, well enough to perform their own role safely. Where a task exceeds their knowledge or skill, they shall decline it, escalate it, or obtain competent support before undertaking it.
A clinician using a risk prediction needs to know which population it was validated on and what the score actually means; they do not need to be able to retrain it. An adviser reviewing an AI-assisted decision needs to understand enough about how the output was produced to tell whether the question they have been asked is the right one. Being unable to explain a system’s internal mechanism is not incompetence. The relevant question is narrower: can you judge whether this system is suitable for the task in front of you?
The explains the system’s role and limits in their own work, recognises where their competence runs out, and chooses an appropriate response to that.
General AI expertise, or a complete causal explanation of every output.
Sources [NSPE] I.2; [AIA] Art. 4, as contextual support. Article 4 places an AI-literacy obligation on providers and deployers; it is not an individual qualification requirement, and conformity with this standard does not discharge it.
Keep your competence current
A shall maintain the AI-related knowledge and skills their role needs, as systems, practice and applicable obligations change. They shall keep a dated record of relevant learning and of how it informs their work. Records of learning kept by the organisation may serve as this record where they show the same.
Learning counts whether it is a course, technical reading, a changed reporting rule in your sector, or a review of an incident relevant to your work. The record can be a few dated lines in a file you keep yourself: what you read or attended, and what you would now do differently. It does not have to be held by an employer, it does not have to follow any particular format, and it should not contain confidential .
A dated record links specific learning to current responsibilities, and identifies what the changed in their practice as a result, or why nothing needed to change.
A fixed number of hours, formal courses, or employer-funded training.
Sources [NSPE] III.2.e; [NIST] GOVERN 2.2. The individual learning record is an AIEI requirement.
Use only approved AI tools
Within the arrangements described in 2.7, a shall use and services only where the organisation has approved them for the purpose and for the category of information involved, and shall not enter confidential or personal information into a service that is not approved for it. A professional working independently or for a client shall apply the client’s rules; where none exist, they shall keep a short record of the AI services they use, the information each may receive and the provider’s terms on retention and reuse of that information, and shall follow it. Where no rule covers a use that could affect people or information, they shall ask the person responsible, or the client, before proceeding.
Most organisations now state which AI tools may be used and with what information. A consultant who pastes a client contract into a personal chatbot account breaches this duty even if the output is accurate; the same task in an approved service with the right settings does not. For a freelancer or a sole practitioner the record can be a few lines: the service, what may go into it, and whether the provider keeps or trains on the content. A rule that conflicts with another duty in this standard is raised under 2.7, not ignored.
The can show that the AI services they use are approved for the information involved or, working independently, can show their own record and that they followed it, and can explain what they did where no rule existed.
Writing the organisation’s policy, assessing the security of an approved service personally, or knowing rules outside the ’s role.
Sources AIEI requirement. Organisational AI policy context: [ISO 42001].
6Responsible practice
Assess the work and protect what it uses
Requirement 1 Before starting AI-related work, or changing its purpose, data, model or operating context, a shall assess the reasonably foreseeable effects . The assessment shall cover relevant misuse, privacy, security, accessibility, and wider social or environmental effects, at a depth proportionate to the risk. Where the work cannot be done responsibly, they shall raise that.
Requirement 2 , a shall confirm that the use of data, content, models and tools is permitted for the intended purpose, including under licence terms and intellectual property rights, either through an organisational rule that covers the use or through the person responsible, such as legal, compliance or data protection. They shall use only what is necessary and shall apply the protection and security controls that fall within their responsibility. Where permission remains unclear, they shall hold the affected use until it is resolved.
Requirement 3 A who selects or prepares data for training, fine-tuning, evaluation or retrieval (information the system looks up when producing an answer) shall record its source and provenance, and shall check that its quality and representativeness are adequate for the intended use and for the people affected. limitations shall be recorded and passed to those who rely on the system.
Uploading a client file to an AI service is a use of that information, even when the task feels routine; what the service may store, access or reuse is part of the question, and so is who processes it and where. For a deployment the same assessment asks whether a simpler approach would achieve the purpose with less harm. Some kinds of personal data, such as health or biometric data, need a specific legal ground in many jurisdictions; that is a question for a competent adviser, not one to settle alone. Licence terms work the same way: a model or dataset published openly may still restrict commercial use. Where the service in question is not approved for the information, 5.3 applies first.
The explains the foreseeable effects, who confirmed that the use of the information was permitted, and the protections that were theirs to apply. gaps were raised, and the affected use was held back where the requirement calls for resolution first.
Personally performing an organisation-wide impact assessment, guaranteeing a third party’s compliance, or obtaining consent where another valid basis permits the use.
Sources [OECD] 1.1, 1.2, 1.5(c); [GDPR] Arts. 5–6, 9, 28, 32 and Chapter V; [NIST] MEASURE 2.7, 2.10, 2.12. Legal scope and individual duties differ.
Check outputs before relying on them
Requirement 1 Before relying on an , a shall check that it is fit for the intended use, by a method proportionate to the consequences of error. The check shall use evidence or a test independent of the output’s own assertion that it is correct. A system’s confirmation of its own output is not verification. Agreement from another , on its own, is not verification either. A disclaimer attached to an output does not justify relying on it without adequate support.
Requirement 2 Where direct verification is not available, a shall assess the validation evidence that does exist, state the resulting uncertainty to those relying on the result, and limit use to what that evidence and the safeguards justify. Where they are insufficient for the consequences, the professional shall withhold reliance and shall seek support or escalate.
Requirement 3 For repeated use, a shall review the checks after changes or after evidence of failure. Sampling or system-level validation may take the place of individual review only where the risk justifies it, and the basis for that choice shall be recorded.
Check a factual claim against a . Test code against its requirements and its failure cases. Assess a prediction using the validation evidence and the case in front of you. A forecast cannot be checked against a future fact, but its uncertainty and its suitability can be. A second model can help find errors, but its agreement alone does not establish that an answer is correct. For someone who is not technical, the practical form of this duty is usually two questions: before I act on this, what could I check it against – and what happens if it is wrong? For generated text, check that cited sources exist and say what is claimed; fabricated references remain a common failure. A document produced with AI and presented as evidence of practice, such as a policy or a record given to an auditor, shows what was written, not what is done, and needs corroboration from records, observation or interview. An output with limited validation might reasonably be withheld from a decision or used only to flag a case for an independent check; either may meet this duty if the can explain why the evidence and safeguards support it.
The explains the method used, the evidence it rested on, its limitations, and the decision they then took about relying on the output.
Identical checks for every output, certainty about the future, or an external reviewer for every routine task.
Sources [NIST] MEASURE 1.1, 1.3, 2.3–2.5; [BUCINCA]. The reliance rule is AIEI’s formulation.
Examine unfair differences in effect
Where their work can affect people’s opportunities, treatment or rights, a professional shall examine the relevant data, model behaviour and outcomes for between groups, and their lawful access. Where feasible and , they shall consider relevant proxies and intersecting disadvantages. They shall explain the criteria used to judge whether a difference is justified, and shall record material limitations. An unexplained difference, or an inability to carry out a necessary examination, shall be raised with someone able to investigate it. Credible evidence of harmful discrimination engages 9.2.
Overall accuracy can conceal poor performance for a smaller group. A difference is not justified merely because a model found it, or because it predicts an organisational target well; the purpose, the consequences and the applicable rights are what decide. Where a has no lawful access to the attributes needed – the usual position for a procurement officer or a manager using a bought-in tool – the practical step is to put a specific request to the supplier or the responsible owner and record the answer. If the necessary evidence remains unavailable, 2.4 and 6.2 govern how far they may rely on the system. Lack of access limits what can be concluded. It never demonstrates the absence of bias.
The can show the examination, or a specific and justified request for it, explain which groups and outcomes were considered, and account for differences and evidence gaps. Where a necessary examination remained unavailable, they can explain how they limited reliance under 6.2.
Every to run statistical tests, unlawful collection of protected data, one universal fairness metric, or identical outcomes for all groups.
Sources [OECD] 1.2; [NIST] MEASURE 2.11; [GENDER]. [AIA] Art. 4a is relevant to lawful access in the EU.
Make oversight effective throughout use
Requirement 1 A assigned to review, approve or override shall exercise independent judgment, with enough attention and understanding to do so, and shall not approve outputs as a formality. They shall check that they have the information, the time and the practical ability to disagree or to seek intervention. Where any of these is absent, they shall withhold approval they cannot support and shall raise the deficiency.
Requirement 2 A responsible for operation, maintenance or oversight shall, within that responsibility, monitor relevant failures and complaints, and shall reassess after changes or new evidence of harm. They shall carry out or initiate correction, restriction or safe withdrawal where it is needed, using the organisation’s change and incident processes where they exist.
Requirement 3 Where their work reveals a error affecting a person, a shall correct it within their authority or refer it promptly to someone who can, and shall support any necessary notification and remedy . 9.2 continues to apply.
Requirement 4 A with authority to approve a use, or to allocate the resources it needs, shall within that authority provide the competence, time, information and authority that effective oversight requires. They shall not approve a use where essential safeguards cannot operate effectively.
A reviewer needs a usable route to reject an output, not only a button marked “approve”. Where rejecting takes an hour and approving takes a second, the oversight is nominal and the deficiency is the thing to raise. A model update, a new user population or a pattern of complaints can each invalidate an earlier assessment. A who does not own monitoring can still pass on a problem they happen to find. A manager who approves a tool sets the conditions every reviewer works in: a review queue with no time budgeted for it is a decision about oversight, not a neutral default.
The reviewer explains their own decision and can show they were able to act on disagreement. A with an ongoing responsibility shows relevant monitoring, reassessment, and either corrective action or a referral. A professional with approval or resource authority can show what they provided for oversight, or why they withheld approval.
A target override rate, re-deciding every case from scratch, or personally repairing a system outside one’s responsibility.
Sources [AIA] Art. 14(4); [NIST] MANAGE 4.1, 4.3; [BUCINCA].
Protect AI systems against attack and misuse
, a who builds, integrates or operates an AI system shall identify the reasonably foreseeable security threats specific to AI systems that apply to it, shall apply or request safeguards proportionate to those threats, and, according to their role, shall test them, arrange for them to be tested, or obtain evidence that they have been tested, before consequential use and after change. A professional who uses an AI system shall not deliberately circumvent its safeguards outside authorised testing, and shall report behaviour suggesting that the system has been manipulated or compromised under 8.3.
Threats specific to include manipulation of inputs or instructions, poisoning of training or retrieval data, disclosure of confidential information through outputs, extraction of models or data, and compromise of third-party models, datasets or components. A system that reads email, web pages or uploaded documents can be given instructions hidden in that content; an assistant connected to customer records can be led to disclose them. For a developer this duty usually means threat modelling, adversarial testing (red teaming) proportionate to the risk, and checking where models and datasets come from. For a buyer it means asking the supplier what testing was done and against which threats. For a user it means recognising unexpected behaviour and reporting it.
A who builds, integrates or operates a system can name the threats relevant to it, the safeguards applied or requested and the testing done, with its limits. A user can recognise signs that a system has been manipulated and knows where to report them.
Security expertise beyond the role, testing of systems the does not control, or disclosure of security details to people not entitled to them.
Sources AIEI requirement, informed by [NIST AML] and [OWASP LLM].
Keep control of actions taken by AI agents
Requirement 1 A who configures, operates or delegates tasks to an AI agent shall limit its permissions and access to information to what the task requires. They shall set and enforce the limits within which the agent may act without individual approval. Individual approval by an authorised person shall be required before an action takes effect if it falls outside those limits, could affect a person’s safety, rights, treatment or opportunities, or could have financially significant consequences. An approval may cover a single specified action, or a bounded operation whose actions, or assets, and limits have been specified and reviewed before the operation takes effect. Approval of a purpose or of a permission limit alone is not approval of actions that meet the approval conditions above but have not yet been specified. A new approval shall be obtained before the affected action takes effect where a material change invalidates the basis of an approval, or where an action that meets the approval conditions falls outside the approved operation.
Requirement 2 The shall ensure that the agent’s actions are recorded and that further actions can be stopped. Where the platform cannot provide either control, they shall raise the deficiency under 2.7 and restrict the agent’s autonomy or provide adequate alternative controls. Those controls shall provide reliable records and a way to prevent further action when needed. Until adequate controls are in place, they shall not permit the agent to take actions with consequences. Delegating a task does not transfer the professional’s accountability .
An agent that can send email, move money, change records or run code acts with whatever rights it is given. A routine message within agreed limits need not require individual approval merely because it cannot be recalled. A decision that affects someone’s rights or opportunities does require approval, even if it can later be reversed. Approval comes before the action, not as a review of a log afterwards. Approving a batch of 40 refunds, with the recipients, amounts and limits listed and checked in advance, covers the steps the agent takes to carry it out; a refund outside that list needs its own approval. Where a platform lacks logging or a stop function, external records and controls may suffice if they provide the required protection. Reporting the deficiency alone does not permit continued use with material consequences.
The can show the permissions given, how the limits of the agent’s autonomy are enforced, which actions need individual approval, where actions are recorded and how further actions are stopped, or the alternative controls and restrictions applied.
Individual approval of routine actions that meet none of the approval conditions above, personally modifying a platform outside the ’s control, or avoiding agents altogether.
Sources AIEI requirement, informed by [OWASP LLM] (excessive agency).
7Honesty and openness
Describe systems honestly
A shall not misrepresent a system’s capabilities, accuracy, validation or reliability. They shall give those relying on their description the limitations and uncertainty alongside the claim, and shall be able to trace a material performance claim to its evidence. Confidentiality does not justify an unsupported claim; where the supporting evidence cannot be shared, the claim shall be qualified accordingly. They shall correct a material misstatement they made or endorsed once they discover it.
A result from one population or one test setting does not establish performance everywhere. “95% accurate” can mislead unless the reader knows what was measured and on whom. A concise qualification the reader actually sees is worth more than a long technical appendix they never open. This duty binds the person who repeats a vendor’s claim in a procurement paper or a board slide as much as the person who generated it. A supplier’s certificate or independent assurance report supports a claim only within its stated scope, period and boundaries; presenting it as proof of a property it does not cover is a misrepresentation under this duty.
The claim traces to evidence, its limits travel with it, and any misstatement the discovered has been corrected.
Publishing confidential test data, or listing immaterial limitations.
Sources [NSPE] II.3.a; [AIA] Art. 13; [GENDER].
Make AI involvement and challenge understandable
Where AI shapes an interaction with a person, or a decision about them, a shall, , provide or enable clear and accessible information about that involvement. For a decision about the person, the information shall include the principal reasons and the material facts actually relied on, the role and limitations of the AI output, and how to challenge the outcome or correct inaccurate information. The account shall be supported by evidence; the professional shall state any material limit on what can be explained and shall not invent a rationale. The information shall reach the person in time for them to understand it and, where relevant, to challenge the outcome. The professional shall support a route to a human with authority to review the outcome and to arrange correction or remedy where it is warranted. Where information needed for meaningful review is unavailable, they shall seek it or competent support and shall limit reliance as 2.4 and 6.2 require. Where that route is missing, they shall raise the deficiency; an unsafe use engages 9.2. Where law restricts disclosure, 2.5 applies.
A designer can build the explanation and the review route; an operator can direct a person to it; a manager can insist that it exists before the tool goes live. Naming a complaints address is not enough if nobody behind it can actually change the outcome. Where the organisation already has a human review process that works, using it discharges this duty – nothing new has to be invented. An applicant whose application is rejected should learn which factors decided it and which facts about them were used, and should be able to correct an income figure that was wrong and have the decision looked at again.
The information and the route are understandable to the intended audience, and for a decision about a person they include the reasons and facts relied on, or the can show what they did, , about a route that was missing.
Disclosing model architecture, obtaining universal consent, or personally providing a reviewer the has no authority to appoint.
Sources [OECD] 1.3; [AIA] Arts. 50(1), 86; [GDPR] Arts. 13–15, 22(3). These provisions support the approach; they do not create a general duty for every AI-assisted decision, and AIEI’s professional scope is broader than them.
Do not misuse a person’s identity or likeness
A shall not use AI to imitate a real, identifiable person’s voice, likeness or identity with the intent to mislead others about who is communicating, or about what that person said or did, and shall not knowingly assist such a use. Where a non-deceptive use could reasonably be mistaken for an authentic representation, they shall make its synthetic nature clear to the intended audience. Any required permission or legal basis shall be checked under 6.1. Separate legal marking and disclosure obligations apply regardless of intent, and this duty does not displace them. A professional shall not use AI to create sexual or intimate imagery of a real, identifiable person without that person’s consent, or sexual imagery depicting a child, whether real or synthetic, and shall not knowingly assist such a use.
Consent to use someone’s voice does not license putting invented statements into their mouth. Satire and fiction are judged in context, but the test is what the intended audience would take the to be, not what the maker thought was obvious. Note that EU law requires certain synthetic content to be marked whether or not anyone intended to deceive; meeting this duty does not answer that legal question, and 2.6 explains why it cannot.
The purpose and presentation are non-deceptive, any necessary disclosure is clear to the intended audience, and the permission question was addressed under 6.1. No sexual or intimate imagery prohibited by the requirement was created or knowingly assisted.
Consent for every lawful depiction, or additional labelling where the representation could not reasonably be mistaken for authentic.
Sources [NSPE] I.5; [AIA] Art. 50(2) and 50(4). The prohibition on intentional impersonation is AIEI’s own rule.
Preserve legitimate scrutiny
A professional shall not design, approve or operate AI, alter records, or withhold an account, with the intent to defeat legitimate scrutiny. , they shall enable lawful examination by regulators, courts, auditors and entitled to it, and shall raise a inability to provide it.
Confidential information can usually be examined under controlled access rather than withheld outright. An honest limit on what a system can explain is different in kind from arranging matters so that it cannot be examined at all. For most people this duty lies dormant until the day a request arrives; what it asks then is that you do not quietly make the evidence harder to find. Legitimate scrutiny includes a request by an , or their legal representative, to examine an output relied on in proceedings concerning them, subject to applicable law and procedure.
Relevant evidence has been preserved and made available by an appropriate route, or a inability to do so has been raised.
Public release of source code, weights or trade secrets, or an explanation the system cannot support.
Sources [OECD] 1.3–1.5; [NIST] GOVERN 4.2.
8Accountability
Make your judgments attributable
Where a ’s judgment shapes an AI system’s risk, approval or use, or a decision made with it, they shall ensure, , that a record identifies their contribution and its date. Where the recording system is outside their control, they shall ask the person responsible for the record to include the contribution; that request, made through the organisation’s systems, may serve as the record. For a collective decision, they shall distinguish their own assessment, any material reservation they held, and the person or body that took the decision, so far as they know it.
An approval log or a dated review comment can show who assessed the evidence and who authorised the use. A team label alone does not explain who made a contribution. If you work inside a system you do not control, such as a bought-in platform or a client’s environment, a dated email to the record owner asking for your contribution to be recorded is both the request and the record. Copies in a personal account or on a personal device are not required and should not be made.
Another person can identify the ’s own judgment and tell it apart from the final collective decision.
Sole ownership of a team decision, formal sign-off powers, or accepting blame for actions that were not the ’s.
Sources [OECD] 1.5(a)–(b); [NIST] GOVERN 2.1. Named individual attribution is an AIEI requirement.
Keep a useful and protected record
A shall make records sufficient to reconstruct the basis of their AI-related decisions and assessments. The records shall identify the relevant system or version, the evidence relied on, the uncertainty, and any material reservations. They shall keep the records retrievable and protected, using the lawful retention arrangements that apply. Where no arrangement exists, they shall ask the person responsible to set one and shall record the request. They shall not keep personal copies of confidential or personal information outside the organisation’s systems in order to meet this duty. They shall not falsify a record, and shall not destroy one in order to obstruct a review, complaint or investigation.
A link to a controlled validation report can be enough, provided its version stays identifiable. Copying sensitive client records into a personal portfolio solves an evidence problem by creating a privacy problem; a short redacted account, or an inspection arranged for the reviewer, does the same job safely. A short decision note might read: “8 September: reviewed model v3 against validation report R12. Evidence does not cover applicants under 25. Recommended excluding that group from automated scoring pending further validation; referred to the service owner for decision.” The same note can support several duties. Its detail depends on the decision; a routine action does not need this level of recording. Records made under this standard may be disclosable in legal proceedings. This standard does not affect any privilege or protection against self-incrimination that may apply to them.
An authorised reviewer can reconstruct what was decided and why, and the can explain where the records are held, who can reach them, and for how long they are kept.
Verbatim minutes, logging every routine action, keeping unnecessary personal data, or retention contrary to law.
Sources [OECD] 1.5(b); [GDPR] Art. 5(1)(c), (e)–(f). The record contents and the fallback retention rule are AIEI requirements.
Report AI incidents
A who becomes aware of an AI incident affecting the work shall report it without avoidable delay through the organisation’s incident process or, where none exists, to the person responsible for the system. They shall preserve the information relevant to the incident and shall not alter or delete it except as the response requires. Where the incident also gives for a concern under 9.2, they shall raise that as well.
Incidents include harm caused by an output, a leak of information through an AI service, and an agent acting outside its intended purpose in a way that affects people or systems. A failed safeguard that created a credible risk of remains reportable even when another check prevented the harm. Many organisations have legal deadlines for reporting serious incidents to authorities, counted from when the organisation becomes aware; a prompt internal report helps the organisation meet those deadlines. An incident is something that has happened; a concern under 9.2 is about how the work is being done. The same event can be both.
The report was made promptly through the right route, with enough information to act on, and the relevant information was preserved.
Deciding whether the incident must be reported to an authority, investigating it personally, or reporting outside the organisation.
Sources AIEI requirement. [AIA] Art. 73 sets serious incident reporting duties for providers; this duty supports such obligations but does not discharge them.
9Integrity under pressure
Keep your judgment independent
A shall form and state their assessment of AI risks, limitations and outputs on the merits, and shall not change it in order to satisfy commercial, political or managerial preference. They shall disclose a conflict of interest to the person relying on their judgment, and shall seek independent review or withdraw from the assessment where the conflict cannot be adequately managed. A material change of assessment shall have a recorded reason. Continuing to participate in work about which the professional holds a credible concern remains subject to 9.2.
New evidence can properly change a judgment; pressure to meet a launch date is not new evidence. A personal financial interest in a supplier is the ordinary case calling for disclosure and a different reviewer. Restating an unwelcome assessment in less combative language, with the substance intact, is not a breach of this duty. Replacing the substance is. A manager may choose a different option that the evidence also supports. Disagreement alone does not engage 9.2; to believe the work meets one of its stated conditions does.
The stated assessment matches the evidence, any change of position has a recorded reason, and conflicts were disclosed and managed.
Winning the argument, or refusing to continue whenever a decision goes against the ’s advice.
Sources [NSPE] II.3, II.4.a, III.5.
Raise concerns and stop unsafe contributions
Requirement 1 Where a has to believe that AI-related work is unlawful, unsafe, or producing systematically wrong or unfairly discriminatory outcomes, they shall record the concern and raise it without avoidable delay through the organisation’s channel for such concerns or, where there is none, with their line manager. Where the use is an as defined in section 11, or where the professional knows, or has been advised by a competent person, that the use is unlawful, they shall also cease or restrict their own contribution to the use in question, so far as it is within their control, and shall act immediately where serious harm is imminent.
Requirement 2 Where immediate cessation would foreseeably cause greater , the shall instead take the safest lawful interim action within their competence, urgently seek an authorised decision, and record the reason, the safeguards and the time set for review. This exception permits only the minimum continued use needed for a safe transition. It does not permit continuation for commercial convenience.
Requirement 3 A shall reassess an unresolved concern at a time justified by its urgency. Where the response is inadequate, they shall follow up or use another available channel, and shall apply 9.3 when its conditions arise. Resumption of a restricted contribution shall rest on evidence that the concern has been resolved or adequately controlled.
Restricting your own contribution is usually smaller than it sounds. A caseworker can stop treating a score as decisive without switching off the service; a reviewer can stop signing off; a developer can stop shipping the affected component; an adviser can withdraw the opinion that was relied on. Where an abrupt shutdown would threaten continuity of essential care, a controlled transition is the safer course, and Requirement 2 above is what covers it. An acknowledgement of your complaint is not evidence that the risk has been controlled. For example, a disputed score can be held out of a decision while the case proceeds through a supported manual route. Resumption needs evidence relevant to the concern, such as corrected data and a fresh check of the affected results. Below the threshold for ceasing, raising and recording the concern is what this duty asks.
The concern, the response and the review timing are recorded. The restricted what was within their control, or can justify a safe transition. Any resumption is supported by evidence.
Proof of wrongdoing, power to shut down an organisation, public disclosure, or repeating an unchanged report after an adequate response. Legal restrictions are addressed in 2.5.
Sources [NSPE] II.1.a; [NIST] 1.2.3, MANAGE 2.4; [OECD] 1.4(b). Timing and safe-transition conditions are AIEI requirements.
Escalate unresolved serious matters
Requirement 1 Where a concern raised under 9.2 involves of serious AI-related harm, or of a serious AI-related breach of law, and has not been within a time justified by the urgency, a shall escalate it within the organisation: to the next level of management or, where management is the source of the concern, to the function responsible for compliance, risk or ethics, or to the governing body. An independent practitioner shall escalate to the client’s responsible senior person or to the contracting party.
Requirement 2 The shall record the escalation and the response. Where they are themselves the most senior person able to decide, they shall take or direct the response the matter requires within their authority, and shall record the decision and its basis. Having done so, and while continuing to apply 9.2 to their own contribution, they have met this duty.
Requirement 3 This standard does not require a report outside the organisation. A lawful report to a regulator or other competent authority is not a breach of this standard, and no shall treat it as one. A reporting duty that law places on the is not affected by this clause (see 2.5). Nothing in 9.2 or 9.3 requires the disclosure of information protected by professional privilege or secrecy (see 2.5).
Internal escalation is the limit of this duty. This is a policy choice of AIEI; it does not discourage a lawful report outside the organisation, which Requirement 3 protects from being treated as a breach. Where the organisation does nothing after a properly recorded escalation, that is an organisational failure, not a failure of the certified person. A who considers an external report should first take advice on the protection available in their jurisdiction and for the route they intend to use; in the EU, Directive (EU) 2019/1937 protects reporting persons under conditions set by national law, and many other jurisdictions offer less protection or none.
The record shows the escalation, the level it reached and the response, and that the continued to apply 9.2. The assessor judges the information and the options reasonably available at the time, not with hindsight.
Reporting outside the organisation, public disclosure, or ensuring that the recipient stops the harm.
Sources [AIA] Arts. 85, 87; [WBD] Arts. 6–7, 10, 21; [NSPE] II.1.a. The escalation sequence is an AIEI policy choice. [NSPE] II.1.a also calls for notice to such other authority as may be appropriate; this standard deliberately stops at escalation within the organisation.
10Relationship to other instruments
Requirement This standard establishes its own duties. The references give support or context; none of them is incorporated as an additional requirement. Conformity with this standard neither requires nor establishes conformity with any cited instrument, and shall not be presented as doing so. The one exception is section 12, which sets ISO/IEC 17024 as the reference for certification of persons against this standard. The definitions in section 11 govern interpretation of this standard; applicable legal definitions govern legal questions.
The OECD Recommendation and the NIST AI Risk Management Framework supply broader policy and risk-management context. NSPE supplies a -ethics comparator. EU legislation supplies jurisdiction-specific examples, not universal duties for every individual. ISO/IEC 42001, 23894 and 42005 address organisational management, risk and impact assessment. ISO/IEC 17024 addresses bodies certifying persons; citing it here does not validate this standard. AIEI AG 1000 sets requirements for organisations; this standard sets duties for persons. An organisation may use certification of its people against this standard as evidence of competence for roles defined under AIEI AG 1000 or under its own management system, but certification of a person is not evidence that the organisation conforms to any standard.
Requirement A shall identify the edition of this standard that it assesses. A revision of an external reference shall not silently change the requirements of that edition. Adoption of a substantive change requires a dated revision of this standard and a published transition arrangement. Changes in law apply independently of this standard. The citations in Annex A identify the versions used for this edition; local implementation and commencement must be checked before any legal decision is taken on them.
11Terms
These definitions set the scope of the professional duties in this standard. They are not substitutes for legal definitions, and 2.6 explains why they cannot be.
- Professional
- The individual to whom this standard applies, as described in 2.1. The term denotes a person carrying a duty under this standard. It is not a job title, a seniority level, a licence, or a claim about qualifications.
- Within their role
- The part of the work for which the person is responsible and which they have the practical access and authority to affect. Where a duty is limited to what is within a person’s role, the limit is on the action required, not on the duty to notice and to raise what falls outside it.
- AI system
- A machine-based system that, for explicit or implicit objectives, infers from the input it receives how to generate outputs, such as content, predictions, recommendations or decisions, that can influence physical or virtual environments. Systems differ in autonomy and in how far they adapt after deployment. Software that only executes fully specified procedural rules, without such inference, is outside this definition. The definition draws on [OECD], section I, and [AIA] Art. 3(1), with [ISO 22989] as terminology context, and is used for the purposes of this standard.
- AI output
- Content, a score, a recommendation, a decision or another result produced by an AI system, including an action initiated through a connected tool.
- AI agent
- An AI system that can take actions in other systems or in the physical world through connected tools, interfaces or devices, with or without human approval of each action.
- AI incident
- An event in which an AI system, or its use, caused harm to people, a breach of law, or a loss of confidential or personal information, or took an action outside its intended purpose that affected people or systems. An AI incident also includes a serious near miss: an event that created a credible risk of serious harm because a safeguard was missing, failed or was bypassed, even if another check or intervention prevented the harm. An expected output error detected and corrected by the intended checks before reliance or exposure is not an incident unless it reveals such a serious near miss.
- Affected person
- A person whose safety, rights, treatment or opportunities may be altered, directly or indirectly, by the work or its outputs, whether or not they use the system.
- Due care
- Reasonable attention and action to prevent or reduce foreseeable harm, within the person’s competence, role and practical ability. It is a professional expectation here, not a finding of legal liability.
- Material, significant
- Important enough to change a reasonable professional’s decision, safeguards or advice in the circumstances. A minor technical change can be material where its consequences are substantial.
- Consequential
- Describes work, a use or a decision whose foreseeable effects on people are substantial enough to warrant the fuller treatment that 2.3 requires. It is the same test as “material”, applied to the work rather than to a fact within it.
- Credible reason, credible evidence
- Specific information that a competent professional in the same situation would regard as sufficient grounds to act. It need not prove the concern; unsupported suspicion alone is insufficient.
- Serious harm
- Harm of substantial severity or scale to health, safety, rights, essential services, livelihoods, property or the environment, taking account of duration, reversibility and vulnerability. It includes cumulative effects that are serious in aggregate.
- Unsafe use
- A use with a reasonably foreseeable risk of serious harm for which adequate safeguards are absent, or a use that fails applicable safety requirements. A merely imperfect result is not automatically unsafe.
- Unjustified adverse difference
- A group disadvantage without an adequate, lawful and proportionate explanation in relation to the legitimate purpose and the affected interests. Statistical association alone is not a justification.
- Reliable primary record
- An original record, or a competent originating authority, relevant to the claim and to the current context. Institutional status alone does not make a source accurate. Verification under 6.2 may instead use testing or validation.
- Residual risk
- The risk that remains after safeguards have been applied.
- Adequately addressed
- A response that resolves the concern, or that controls the risk to a level the professional can justify on the evidence available. An acknowledgement, an assurance without evidence, or a deferral without a review date is not an adequate response.
- Scheme
- The set of arrangements by which a body assesses and certifies persons against this standard, including its role scope, entry expectations, assessment methods, decision rules and appeal arrangements. Section 12 states what a scheme shall do.
12Use in certification
12.1Application
Requirement This section applies to any claiming to assess conformity with this standard, including the scheme operated by AIEI. AIEI owns and operates the certification scheme for this standard: it publishes the standard and the information candidates need to prepare, develops and maintains the assessment scenarios and items, and issues certificates. The examination itself is delivered and invigilated by an independent specialised proctoring service. AIEI does not conduct the examination and cannot influence an individual candidate’s result. AIEI is seeking accreditation of its certification of persons against ISO/IEC 17024:2026, the relevant certification-body reference. Until accreditation is granted, citing it here is not a claim of accreditation, and no claim of conformity with it follows from this document.
12.2Scope of certification and job analysis
Requirement A shall define the role scope it certifies, the entry expectations and the assessment methods it uses. It shall derive its competence requirements from an analysis of the work the certified roles actually involve, and shall be able to produce that analysis. The “Sources” lines in this standard record where each duty came from; they are provenance, not a job analysis, and shall not be presented as one.
12.3Coverage and depth of assessment
Requirement 1 A shall assess all twenty-one duties and the application rules in sections 2, 3, 10 and 11. It shall state, for each duty, the depth at which it is assessed and why that depth is appropriate to the role scope. Where a duty has not arisen in a candidate’s work, the scheme shall use a relevant scenario or simulation. Lack of workplace opportunity shall not be treated as a failure, nor as evidence of competence. The certificate confirms that the candidate demonstrated, at the level of knowledge and judgment and by scenario-based assessment, the ability to apply all the duties; it does not confirm that the person performs each duty in their current work.
Requirement 2 A may define critical failures at the level of a group of related duties rather than separately for each duty, provided the grouping and its justification are published under 12.6.
12.4Integrity of assessment
Requirement 1 A shall state whether, and how, a candidate may use an during assessment, and shall be able to distinguish the candidate’s own judgment from an AI system’s contribution to a response. It shall manage the exposure of its scenario and item bank, and shall state its arrangements for verifying candidate identity and for invigilation.
Requirement 2 Where a uses AI in assessment, marking or certification decisions, it shall state that use to candidates, validate it before use and after change, keep a person responsible for every certification decision, and give a candidate affected by an AI-assisted adverse decision a review by a competent person. The duties in 6.2, 6.4, 7.1, 7.2 and 8.1 apply to the people who operate the scheme.
Requirement 3 A shall protect its scenario and item bank, candidate evidence and candidate personal data by documented information security controls proportionate to the risk, including access control, secure transfer and storage, and defined retention and disposal. It shall publish how it responds to requests by courts or authorities for assessment evidence, including notice to the candidate where lawful.
12.5Impartiality and scheme governance
Requirement A shall separate responsibility for developing this standard, for preparing candidates, and for deciding certification. It shall not make training provided by the scheme, or by AIEI, a condition of certification. It shall involve interested parties – including people working in the roles being certified, and people affected by – in developing and reviewing the scheme, and shall record how their views were addressed. It shall identify and manage conflicts of interest arising from its own funding and from its relationships with employers of candidates. This includes conflicts arising from AIEI’s role as owner of the standard and of the scheme. Authors of this standard, reviewers and anyone with access to assessment who seek certification shall be assessed against the same criteria as other candidates, by assessors independent of their involvement, and shall gain no advantage from prior access to items. The scheme shall publish how this is achieved before certification begins.
12.6What a scheme shall publish before it certifies anyone
Requirement 1 Before certification begins, a shall publish: the evidence it accepts and the pass criteria for every duty, linked to the binding requirements they assess; the critical failures that cannot be offset by other scores, and how they are grouped; and its arrangements for reassessment, recertification, complaints, suspension, withdrawal and independent appeal. It shall state the certificate’s role scope and its limits, and shall protect confidential assessment evidence.
Requirement 2 A candidate shall not be required to disclose employer or third-party information unlawfully. For every duty whose evidence would ordinarily be a workplace record, the shall specify at least one lawful alternative – such as a redacted extract, a supervised exercise, or a scenario response – and shall state which duties it assesses by scenario alone. For candidates bound by legal or medical privilege or a statutory duty of professional secrecy, duties ordinarily evidenced by workplace records shall be assessed by scenario or supervised exercise.
Requirement 3 A shall set a recertification period of not more than three years and published, measurable criteria for maintaining certification, including evidence of continued learning under 5.2 and a defined process for moving certificants to a new edition of this standard.
12.7AIEI’s position on support under 9.2 and 9.3
Requirement AIEI shall publish its position on what support, if any, it offers a person who acts under 9.2 or 9.3, before any uses this standard. Where no support is offered, the statement shall say so explicitly. A scheme shall make that statement available to every candidate before assessment.
The statement shall say whether AIEI offers consultation to certified acting under 9.2 or 9.3, its scope and how to obtain it, and that this standard confers no legal protection (2.6).
12.8Piloting and calibration
Requirement A shall pilot its assessments across the roles and risk contexts it intends to serve, calibrate its assessors against common worked examples, and record how it addresses disagreement between assessors, bias, and accessibility barriers. Assessment shall consider what the candidate could reasonably have known and done at the time, and not merely whether harm occurred or whether a concern was upheld. A scheme shall accept different decisions where each meets the requirements and is justified by the evidence and safeguards available; agreement with one preferred answer shall not be a pass condition.
12.9What a certificate does and does not assert
Requirement A correct scenario response demonstrates judgment in that assessment. It does not prove future conduct, and a shall not claim that it does. A complaint about a certificant’s conduct requires a separate, evidence-based process, with an opportunity to respond and an independent appeal; a scheme shall establish that process before it certifies anyone.
12.10Assessment evidence – informative examples
The following examples suggest evidence that would be suitable. They are informative: a may use equivalent methods, and the depth required in each case is set under 12.3.
| Requirements | Suitable assessment evidence |
|---|---|
| Sections 2–3 | Explanation of how the candidate scales a duty to the risk, their role and their actual authority; correct use of the defined terms in a worked example. Use of the organisation’s arrangements under 2.7. |
| 4.1–4.2 | Scenario identifying , competing interests, and a justified referral or decision. |
| 5.1 | Role-specific explanation of a system’s purpose and limits, and recognition of a competence limit. |
| 5.2 | Dated learning record and discussion of its practical relevance. This is the one duty normally evidenced by a record the candidate produces themselves. |
| 5.3 | Scenario on using AI tools with information of different categories, including a case where no rule applies. |
| 6.1 | Scenario or work sample addressing impact, permissions and protection of information. |
| 6.2 | Practical choice and application of checks; a decision taken under incomplete evidence. |
| 6.3–6.4 | Bias and oversight scenarios, including a case where the candidate has no lawful access to the data; practice evidence where available. |
| 6.5–6.6 | Recognition scenario: identifying signs of manipulation or misuse in a described situation and choosing the response; scenario on setting an ’s permissions, approval points and logging. |
| 7.1–7.4 | Review of claims, disclosure and synthetic content; response to a scrutiny request. Pass criteria are published per duty under 12.6 even where evidence is gathered across the group. |
| 8.1–8.2 | Redacted decision records, or an equivalent supervised exercise where no lawful record can be produced. |
| 8.3 | Incident scenario: recognising an , choosing the route and preserving the relevant information. |
| 9.1–9.3 | Conflict-of-interest and escalation scenarios, including safe transition and escalation within the organisation. Normally assessed by scenario alone. |
| Sections 10–11 | Correct treatment of a cited instrument that does not itself create a duty; correct application of a defined term against a near-miss fact pattern. |
Sources [ISO 17024], public scope only. The detailed scheme conditions above are AIEI requirements; clause-level conformity with ISO/IEC 17024:2026 requires separate verification.
Annex A Referenced instruments and research
Informative. Bibliographic details and cited editions checked against official public records on 20 October 2026. The research cited illustrates particular risks; it does not establish universal error rates, and it does not show that certification prevents harm.
[OECD] Organisation for Economic Co-operation and Development. Recommendation of the Council on Artificial Intelligence, OECD/LEGAL/0449, adopted 2019, amended 2023 and 2024. https://legalinstruments.oecd.org/en/instruments/OECD-LEGAL-0449
[NIST] National Institute of Standards and Technology. Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100-1, January 2023. This is the fixed version cited; later revision work does not change these references. https://doi.org/10.6028/NIST.AI.100-1
[NIST AML] National Institute of Standards and Technology. Adversarial Machine Learning: A Taxonomy and Terminology of Attacks and Mitigations, NIST AI 100-2 E2025, March 2025. https://csrc.nist.gov/pubs/ai/100/2/e2025/final
[NSPE] National Society of Engineers. Code of Ethics for Engineers, revision July 2019. The comparator applies to engineering practice. https://www.nspe.org/career-growth/nspe-code-ethics-engineers
[AIA] Regulation (EU) 2024/1689, Artificial Intelligence Act. EUR-Lex consolidated text dated 27 July 2026. Applicability varies by provision, actor and system; consult Arts. 2, 111 and 113. https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:02024R1689-20260727
[GDPR] Regulation (EU) 2016/679, General Data Protection Regulation. EUR-Lex consolidated text dated 4 May 2016. https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:02016R0679-20160504
[WBD] Directive (EU) 2019/1937 on the protection of persons who report breaches of Union law. EUR-Lex consolidated text dated 30 December 2024. The Directive creates rights for a reporting person, not a duty to report, and its protections depend on national implementation and on the route used. https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:02019L1937-20241230
[ISO 17024] ISO/IEC 17024:2026. Conformity assessment – General requirements for bodies operating certification of persons. Third edition, published March 2026; replaces the 2012 edition. The 2026 edition introduces definitions and requirements for the use of AI in certification processes, which 12.4 addresses. https://www.iso.org/standard/17024
[ISO 42001] ISO/IEC 42001:2023. Information technology – Artificial intelligence – Management system. https://www.iso.org/standard/42001
[ISO 23894] ISO/IEC 23894:2023. Information technology – Artificial intelligence – Guidance on risk management. https://www.iso.org/standard/77304.html
[ISO 42005] ISO/IEC 42005:2025. Information technology – Artificial intelligence (AI) – impact assessment. https://www.iso.org/standard/42005
[ISO 22989] ISO/IEC 22989:2022. Information technology – Artificial intelligence – Artificial intelligence concepts and terminology. Further reading, not the governing definition here. https://www.iso.org/standard/74296.html
[OWASP LLM] OWASP Foundation. OWASP Top 10 for Large Language Model Applications, 2025 edition. LLM06:2025 Excessive Agency. https://genai.owasp.org/llmrisk/llm062025-excessive-agency/
The ISO entries are catalogue- and scope-verified references. No inaccessible ISO clause is used as evidence for a specific duty.
[GENDER] Buolamwini, J. and Gebru, T. (2018). Gender Shades: Intersectional Accuracy Disparities in Commercial Gender Classification. Proceedings of Machine Learning Research, 81, 77–91. https://proceedings.mlr.press/v81/buolamwini18a.html
[BUCINCA] Buçinca, Z., Malaya, M. B. and Gajos, K. Z. (2021). To Trust or to Think: Cognitive Forcing Functions Can Reduce Overreliance on AI in AI-assisted Decision-making. Proceedings of the ACM on Human-Computer Interaction, 5, CSCW1, Article 188, 1–21. https://doi.org/10.1145/3449287
Annex B Reader’s guide by role
Informative. This annex adds no requirements and creates no exemptions. All twenty-one duties apply to everyone within the scope in 2.1 whenever their conditions arise; what changes with the role is how much each one asks, under 2.2. The guide below is only a starting point for reading – it says where people in each kind of role usually meet these duties first, and which ones they most often overlook.
The depth of checking follows the consequences. For an AI-assisted draft of a routine meeting agenda, checking it against the agreed topics may be enough, with the existing work record. For a tool that ranks applicants, the relevant duties call for evidence about suitability, differences in effect and effective review, with the fuller record required by 2.3. These examples illustrate the distinction; they do not assign fixed risk levels to tasks.
| If your work mainly involves… | Where you usually meet these duties first | Most often overlooked |
|---|---|---|
| Building or maintaining AI systems – development, data science, ML engineering, testing | 6.1 assessment and permissions; 6.2 verification; 6.3 differences in effect; 8.1–8.2 records | 4.1 people affected who are not users; 7.1 how your results are described once they leave you |
| Selecting, buying or integrating systems – procurement, vendor management, IT | 6.1 permissions and protection; 7.1 claims made about the system; 4.2 pressure and trade-offs | 6.3 asking the supplier for the examination you cannot run yourself; 7.2 whether a real review route exists before go-live |
| Deciding that a system will be used – management, service ownership, sponsorship | 4.2 competing duties and ; 6.4 whether oversight is real; 9.1 independence | 8.1 recording your own contribution to a collective decision; 2.2 seniority is not a substitute for the risk analysis |
| Acting on AI outputs about people – casework, clinical use, recruitment, credit, analysis, policing, prosecution, courts | 6.2 checking before you rely; 6.4 not approving as a formality; 7.2 explaining involvement | 5.1 knowing what the output actually means; 9.2 restricting your own reliance is within your control even when the system is not |
| Using AI tools in everyday work: any role, including drafting, research and analysis | 5.3 approved tools and information; 6.2 checking before you rely; 8.3 reporting what goes wrong | 6.1 licence terms and protection of what you upload; 7.1 passing on AI-generated claims as your own findings |
| Configuring or running AI agents: automation, operations, IT | 6.6 permissions, approval points and logging; 6.5 manipulation through content the agent reads | 8.1 accountability for actions taken on your behalf; 8.3 an agent acting outside its instructions is an incident |
| Advising, auditing or governing – legal, compliance, risk, audit, data protection | 2.6 what this standard does and does not do legally; 10 how the references work; 7.4 scrutiny | 4.1 your advice shapes who is exposed, so you are within scope; 6.2 verification applies to the outputs you rely on in advising |
The same evidence gap calls for different actions in different roles. A developer may test a model on the population missing from its validation. A procurement officer may request that evidence from the supplier and qualify their recommendation while it is missing. A manager may limit the proposed use and arrange a review. An operator may refer an affected case for independent assessment. Each acts , but none can treat the missing evidence as reassurance. A request without an adequate answer leaves the limits on reliance in 6.2 in place.
A note for readers coming to this from a legal or compliance background. This standard is not a legal instrument and does not interact with law in the way a code or a regulation does; 2.6 sets out the limits precisely, and section 10 governs how the citations in Annex A are to be read. Two points are worth carrying into any advice given on it. First, conformity with this standard does not discharge any legal obligation, including the AI-literacy obligation in Article 4 of the EU AI Act, and it must not be presented as doing so. Second, 9.3 limits the duty to escalation within the organisation. A lawful external report is not a breach of this standard, but the standard confers no protection on a person who makes one; protection depends entirely on the applicable law and on the route used, and 12.7 states AIEI’s own position.

