The seven pillars of responsible AI

AI recruitment systems fall under Annex III of the EU AI Act as high-risk, which triggers the full deployer obligations of Article 26, human oversight, data quality, monitoring, log retention, and a Fundamental Rights Impact Assessment under Article 27. These duties cannot be transferred to the software vendor.

A Fundamental Rights Impact Assessment (FRIA) under Art. 27 is conducted step by step: describe the system and its purpose, identify affected persons, assess the impact on each fundamental rights dimension, define mitigation measures, and document the residual risk before deployment.

Regulatory sandboxes under Art. 57-61 are controlled environments, supervised by the national authority, in which organizations can develop and test innovative AI systems with guidance and temporary relief from certain administrative requirements, without suspending the material safeguards or incident-reporting duties.

Joining a national AI regulatory sandbox under Art. 57-61 follows a structured path: prepare a project dossier, apply in one of the submission windows, sign a sandbox agreement with the supervisor, report progress and incidents during testing, and produce a final report that supports full compliance afterwards.

The EU AI Act and the GDPR create overlapping but distinct obligations for AI systems that process personal data. They align on data quality, impact assessments, transparency, and individual rights, but differ in scope, accountability roles, and incident-reporting timelines, so the efficient approach is integrated compliance, such as a combined DPIA/FRIA.

The OWASP Agentic Security Initiative Top 10 catalogs the security risks that autonomous AI introduces. It is written for security engineers, but the risks are governance problems, because they describe what an agent can be made to do rather than what it can be made to say. This article translates the agentic threat surface into the language of controls an AI Officer owns, and maps each risk to the responsible-AI pillar it stresses.

Ready-to-use transparency templates help deployers meet the EU AI Act information duties: a chatbot disclosure, an AI-generated-content label, and an Art. 26(11) notice for individuals subject to a high-risk system. The disclosure must be active and comprehensible at the moment of interaction.

Agentic AI does not need a new kind of governance. Autonomy widens the gap between what a system does and who is accountable for it, which makes the existing governance chain, control tracing to risk and forward to evidence, more important, not less. The actions are real and sometimes irreversible, so the stakes on each control rise.

Assessing the risk of an AI agent does not need a separate method. The steps stay the same: recognize the risk, assess how likely and how severe it is for your system, and control it. What changes is the input. An agent runs the process through recorded architecture decisions, about the model, the instruction, retrieved knowledge, tools, orchestration, memory, and autonomy, and each of those decisions, alone or in combination, creates the possibility of harm. The output of the assessment is a set of risk scenarios with a control objective for each.

AI certification is not one category. Three different objects are assessed, each by a different kind of assessor: a person, an organization's AI management system, and an AI system placed on the EU market. The first two can be certified. The third is subject to a legal conformity assessment, which produces a certificate on one of its two routes and none on the other. Identifying which object a credential covers is the first step to judging what it is worth.

AI governance is not a parallel structure that sits beside enterprise risk management. It belongs inside it. The seven pillars of responsible AI are the control structure the organization uses to govern each AI system; enterprise risk management is the machine that carries the residual risk those controls leave behind into the board's risk appetite, the risk register, and the assurance plan. The practical question is not whether to build AI governance or ERM, but how to slot the first into the second so that one accountable structure, not two competing ones, owns AI risk.

Control-level compliance means satisfying the EU AI Act through engineered, evidenced controls rather than policy documents. The technical articles translate directly into system controls: automatic, retained logs (Art. 12, 19), a stop function to a safe state (Art. 14(4)(e)), input masking before the model as a GDPR and Art. 26(4) control, configurable block policies (Art. 26), risk scoring and incident reporting within deadline (Art. 9, 73), and workspace isolation with role-based access (Art. 14, 26). Compliance at this level is an instrumented system, not a policy as PDF.

Whether you are a deployer or a provider under the EU AI Act decides which obligations and fines apply, so settle it per system first. Deployers use a system under their own authority; providers place it on the market under their own name, and you become a provider, with the heavier duties, the moment you substantially modify or rebrand a system (Art. 25).

Most organizations can describe their AI agents in detail: the architecture, the tools, the memory, the base model, the benchmarks, the internal safety testing. What far fewer can do is govern what those agents decide once they are running. Documentation answers the question "what is this agent and what can it do?" Governance answers a harder one: "can we trust the decisions it makes over time?" The two are routinely confused, and an agent that is thoroughly documented but ungoverned is exactly the kind of system that passes every review and then fails in production.

For SMEs, EU AI Act compliance is manageable but not optional: the Art. 5 prohibitions and Art. 4 literacy apply regardless of size, and SME deployers of high-risk AI carry the full Art. 26 obligations in proportionate form. Micro-enterprises gain administrative simplifications, not exemptions.

The EU AI Act, ISO/IEC 42001 and the NIST AI RMF answer different questions and are designed to be used together. The AI Act sets legal obligations for AI systems placed on the EU market. ISO/IEC 42001 specifies requirements for an AI management system that an organization can certify. The NIST AI RMF offers voluntary risk management outcomes and actions, and states that its functions are not a checklist and not a required sequence. This article explains how to use them together without merging them into a new framework.

The governance chain holds a single principle; the AI life cycle is the other axis of the model, time. A system is governed across a life from the first intake decision to the day it is switched off, in six stages: plan and design, data and develop, verify and validate, deploy, operate and monitor, and retire. Each stage has controls native to it and produces an anchor artifact, and the two axes meet at evidence. The stage that fails most often is operate and monitor.

An AI agent knows your business process through six building blocks: the model itself, the system instruction, retrieved knowledge (RAG), the tool definitions, the orchestration, and memory. Each block represents a design choice about what lives where and how much the model decides, and every choice shapes the risks the process carries and the control environment it will need.

ISO/IEC 42001:2023 is the first international management system standard for artificial intelligence, published in December 2023. It specifies requirements for establishing, implementing, maintaining, and continually improving an AI management system, and it is certifiable: an accredited body can audit an organization against it and issue a certificate. The certificate covers the management system, not any individual AI system, and it creates no presumption of conformity with the EU AI Act.

The NIST AI RMF, ISO/IEC 42001, and the OECD AI Principles are complementary, not competing: a voluntary risk-management process, a certifiable management system, and a values baseline. None replaces the EU AI Act's legal obligations; they help you operationalize them.

Under the EU AI Act, a provider develops an AI system or places it on the market under its own name; a deployer uses an AI system under its authority in a professional context. The roles carry very different obligations: providers bear the heavy technical duties (Art. 9-15, 19, 43, 48, 49, 72), while deployers bear a lighter operational set (Art. 26, 27). The same organization can be both at once, and determining your role per system is the first compliance step.

Responsible AI is the set of principles an AI system should meet; AI governance is the system that delivers those principles and proves it. One is what good looks like, the other is how you make it real and show that you did. Responsible AI is the goal; governance is how you reach it and evidence that you did.

Most organizations use far more AI than leadership has mapped, much of it switched on quietly inside the SaaS tools already in use. You cannot govern, classify, or defend what you have not inventoried, so the first governance step is always the same: build an AI inventory, then classify and control from there.

The AI Officer is the organization-wide director of responsible AI use, broader than a compliance role: it covers AI strategy, ethics, risk and literacy. The EU AI Act (Art. 26) makes the coordinating function necessary, but the need for an AI Officer extends beyond the law itself.

The control environment for an AI agent is the set of enforced controls and governance arrangements that keep the agent inside agreed boundaries, and it rests on one distinction. Deterministic components, the orchestration, the checks inside tools, the permissions, and policy rules, apply explicitly coded rules predictably and can therefore carry enforceable controls. Probabilistic components, everything interpreted or decided through the language model, can only be influenced and evaluated statistically. Every control from the business process therefore needs an enforcement point outside the model, and a control that exists only in prompt text is an instruction, not an enforced barrier.

The NIST AI Risk Management Framework is a voluntary framework for identifying, assessing, and treating AI risk, published by the US National Institute of Standards and Technology in January 2023. It organizes the work in four functions: govern, map, measure, and manage. It is a method, not a legal requirement and not a certifiable standard: no organization can be certified against it, and following it creates no presumption of conformity with the EU AI Act. Its value is the risk vocabulary and process it supplies inside a management system.

The OECD AI Principles are the first intergovernmental standard on AI, adopted in May 2019 and updated in May 2024, with 47 adhering jurisdictions including the European Union. They consist of five values-based principles for AI actors and five recommendations for policymakers. They are not law, not certifiable, and not a management system: they are the shared reference point that the EU AI Act, national AI strategies, and international frameworks build on. The OECD classification framework for AI systems, published in 2022, is the practical instrument that operationalizes them.

The EU AI Act assigns obligations on the assumption that the provider who builds a system and the deployer who uses it are distinct, stable roles. Agentic AI destabilises that assumption. A deployer who configures an agent with broad tool-calling rights, autonomous decision scope, or the ability to spawn sub-agents may be making changes substantial enough to carry provider-level obligations. Under autonomy, the question of who is answerable cannot be read off the contract. It has to be assessed against what the deployer actually configured the agent to do.

Under the EU AI Act, splitting an autonomous workflow across several agents does not split its regulatory classification. The Commission's draft guidelines on high-risk classification, published in May 2026, state that a complex system made up of several AI components, including an agentic stack of orchestrators and sub-agents, is assessed as a whole. An orchestrator coordinating sub-agents toward a high-risk decision is one high-risk system, and the full weight of the Act's high-risk obligations attaches to the stack, not to its parts.

AI risk management is the practice of knowing which risks your AI systems carry and keeping them within bounds you chose deliberately. The risks are not abstract: each of the seven pillars of responsible AI has concrete ways it fails in practice, from a model that quietly disadvantages one group to a system whose accuracy decays after deployment. Managing them follows one repeatable pattern: recognize the risk, assess how likely and how serious it is for your system, and control it with measures you can test.

An audit trail for EU AI Act compliance is the structured, retained record, combining the system logs (Art. 12) with the deployer's own oversight and monitoring documentation, that lets you demonstrate to a supervisor that a high-risk AI system was used lawfully.

EU AI Act obligations per department depend on the risk class of the AI system. HR selection and credit scoring are high-risk (Annex III) and carry the full Art. 26 obligations; marketing AI and chatbots usually fall under the transparency obligation of Art. 50. A per-system Art. 6 analysis determines the exact obligation.

The first steps to EU AI Act compliance for deployers are: build an AI inventory, classify each system against Art. 6, request the provider documentation, start AI-literacy training under Art. 4, and assign ownership. These steps create the foundation for the Art. 26 obligations.

Deployers using GPAI models like ChatGPT or Copilot are generally not subject to the provider obligations of Art. 52-55, but two frameworks do apply: the Art. 50 transparency obligations and a high-risk use-case analysis. If the way you deploy the GPAI creates a high-risk AI system under Annex III, the full Art. 26 deployer obligations apply.

Whether an AI system is high-risk depends on Art. 6: it is high-risk if it is a safety component under Annex I or falls within an Annex III use case (such as employment, credit, or essential services). The Art. 6.3 exception can apply where the system performs only a narrow, non-decisive task.

An SME that develops an AI system and makes it available to others is a provider under the EU AI Act and carries a substantially heavier burden than a deployer: for high-risk AI this includes a risk management system (Art. 9), data governance (Art. 10), technical documentation (Annex IV), conformity assessment (Art. 43), CE marking (Art. 48), and EU database registration (Art. 49).

Micro-enterprises (fewer than 10 employees and turnover up to €2 million) can use a simplified compliance pathway under the EU AI Act, mainly for the provider role: simplified technical documentation (Art. 11.3) and a proportionate quality management system (Art. 17.3). The material obligations, the Art. 5 prohibitions, human oversight, and incident reporting, still apply in full.

A supplier checklist for AI procurement verifies what a provider must deliver before you can comply as a deployer: the instructions for use (Art. 13.3), the conformity declaration, the risk classification, update notification, and cooperation in a supervisory investigation.

AI governance is the system through which an organization makes responsible use of its AI and can prove it. It is not an ethics statement or a one-time project but an operating discipline that runs for the full life of every AI system the organization builds, buys, or embeds, carrying each responsible-AI principle down a chain from the harm that would breach it, through the risk and the control that reduces it, to the evidence that the control works.

An AI policy is the governance instrument that translates the EU AI Act's obligations into organizational commitments and accountabilities. An EU AI Act-aligned policy covers at least the scope, AI principles, the governance structure, the AI inventory and classification, the Art. 5 prohibitions, the approval process, human oversight, literacy, and incident reporting.

Agentic AI is AI that carries out a chain of actions on its own rather than producing a single output for a human to review. That shift does not add a new responsible-AI principle; it changes how every existing principle has to be governed. The human checkpoint moves from inside each decision to around the whole system: setting the bounds the agent operates within, monitoring the chain as it runs, and holding the ability to intervene.

AI agents do not just answer, they take actions in your systems, amplifying both the value and every failure mode. Governing them means governing the actions, not only the decisions: action allowlists, approval gates for high-consequence steps, full logging, and a kill switch.

Human oversight means AI serves people rather than replacing their judgment. It keeps a competent person meaningfully in control of an AI system, with the authority and the information to intervene, and it keeps that control in proportion to what is at stake. The deeper idea behind it is human-centricity: AI should support human judgment, respect autonomy and dignity, and remain accountable to the people it affects, not only the people who use it. The practical core is choosing the right oversight pattern for the stakes, because oversight that is too light fails to catch harm and oversight that is too heavy fails to scale.

The safest way to deploy an agent is to grant it the least autonomy that lets it do its job, then widen that autonomy only as evidence of reliable behavior accumulates. Progressive autonomy is to agentic governance what the three control layers are to the seven pillars of responsible AI: the operating discipline that turns a pillar into a practice. This article sets out a maturity model for agent deployment along three dimensions, decision authority, process autonomy, and accountability, and the controls that should be in place at each level.

An oversight log is the contemporaneous record that proves human oversight of a high-risk AI system under Art. 26.2 of the EU AI Act. It must capture, per oversight event, who reviewed the AI output, what they decided and why, and it must be retained for at least six months under Art. 26.6.