Assessing an AI System: Audits, Tests and Evidence

Three different questions, three different assessments#

When a client, a regulator or an executive committee asks “has your AI been audited?”, the question can cover three different realities:

Confusing these exercises leads to costly misunderstandings. An ISO/IEC 42001 certificate says nothing about the accuracy of a scoring model; an excellent precision score says nothing about how the organisation monitors that model over time.

The management system audit: ISO/IEC 42001 and 42006#

ISO/IEC 42001:2023 describes an AI management system: policies, roles, risk and impact assessment, controls, monitoring and improvement. The audit checks that this system exists, that it is applied and that it improves. It does not test algorithms.

Since July 2025, ISO/IEC 42006 has set AI-specific requirements for bodies that audit and certify against ISO/IEC 42001, in addition to ISO/IEC 17021-1: auditor competence, audit duration, certificate content. Only bodies accredited by a national accreditation body can issue a recognised certification; their accreditation can be checked in the accreditor's public register. Our guide on ISO 42001 certification sets out the process in detail.

Technical evaluation: bias, robustness, accuracy#

Technical evaluation covers a specific system in its intended use. Two ISO documents frame it:

• ISO/IEC TR 24027:2021 deals with bias in AI systems and AI-aided decision-making. Published in November 2021, it remains in force.
• ISO/IEC TS 12791:2024, published in October 2024, deals with the treatment of unwanted bias in classification and regression machine learning tasks.

A useful technical evaluation covers at least:

1. The data: representativeness of the population served, quality, possible biases. For high-risk systems, Article 10 of the AI Act requires training, validation and testing data that are relevant, representative and examined in view of possible biases.
2. Accuracy: metrics suited to the use, measured by subgroup and not just on average. Article 15 requires appropriate levels of accuracy, robustness and cybersecurity, with metrics declared in the instructions for use.
3. Robustness: behaviour when faced with unusual inputs, typing errors, mixed languages (Arabic, Darija, French) and manipulation attempts.
4. Real-world behaviour: gaps between testing and production, drift over time.

For a customer service chatbot, this means testing answers on real questions, in customers' real languages, and checking the cases where it must hand over (see our guide on AI in complaint handling).

The impact assessment: ISO/IEC 42005 and Article 27#

Published in May 2025, ISO/IEC 42005 provides guidelines for assessing the potential impacts of an AI system on individuals, groups and society, and for documenting that assessment. It is not a certifiable requirements standard: it complements ISO/IEC 42001.

In the AI Act, Article 27 requires a fundamental rights impact assessment before the first use of a high-risk system by certain deployers: public bodies, private entities providing public services, and deployers of systems for creditworthiness assessment or for pricing in life and health insurance (Annex III, points 5(b) and 5(c)). Article 27 was amended by Regulation 2026/1744; the details of the amendment must be read in the consolidated text before any decision.

Where the system processes personal data, the impact assessment ties in with the principles and formalities of Law 09-08 on personal data protection and, for European data, the GDPR's data protection impact assessment. Our privacy by design guide shows how to build these questions in from the design stage.

What the AI Act will require of high-risk systems#

Regulation 2026/1744, in force since 27 July 2026, pushed back the dates: 2 December 2027 for stand-alone Annex III systems, 2 August 2028 for Annex I systems.

Articles 27, 43 and 72 were amended by the Omnibus, and the Commission's service desk was not yet showing the amended text at the date of verification. Articles 9 to 17 also structure our guide on the AI system life cycle.

The NIST framework, useful outside the Union#

The NIST AI Risk Management Framework 1.0, published on 26 January 2023, organises risk management into four functions: govern, map, measure, manage. Its Generative AI Profile (NIST AI 600-1), published on 26 July 2024, applies these functions to the risks specific to generative models. Voluntary, this framework is often requested by North American clients and provides an analytical grid compatible with ISO/IEC 42001.

The Moroccan angle#

Morocco's AI-specific framework is still being built (see our guide on the legal framework for AI in Morocco). However, Law 09-08 applies as soon as the system processes personal data, and its Article 11 governs decisions based on automated processing. An institution that scores credit applications, sorts job applications or prioritises complaints must therefore assess its system against this rule, with the CNDP (Morocco's data protection authority) as the competent authority. AI providers themselves must be assessed as third parties (see our guide on third-party risk management).

Assessment plan and evidence file#

1. Define the system, its intended use, the people concerned and your role (provider or deployer).
2. Determine the level of risk under the AI Act and Law 09-08.
3. Carry out the impact assessment according to ISO/IEC 42005, with the business teams and the data protection officer.
4. Test the data, accuracy by subgroup, robustness and the cases where the system hands over to a human.
5. Check governance: roles, approval, logging, monitoring, incident management.
6. Document the results, known limitations and corrective measures.
7. Schedule reassessment at each substantial modification and at least once a year.

The evidence file brings together the system description, the impact assessment, dated test reports, approval decisions, monitoring logs and the incident register. It also serves to answer European clients' compliance questionnaires.