BSI A5: Germany’s AI assurance architecture and its gap
With A5, Germany’s BSI is building an assurance architecture for AI systems. Still a draft, comments closed on 31 August 2026, and ISO/IEC 42001 is absent.
In short
A5 is the AI Audit and Assurance Assessment Architecture of the German Federal Office for Information Security (BSI). It exists as a community draft, and the comment period ended on 31 August 2026. Methodically it builds on C5 and the assurance standard ISAE 3000. ISO/IEC 42001 is not mentioned in it.
Auf Deutsch lesen: deutsche Fassung
Update of 10 November 2026. The comment period expired on 31 August 2026. The BSI continues to list A5 as a community draft, and a final version has not been published. The assessment below still holds, the call to comment no longer does.
An assurance architecture for AI systems is currently taking shape in Germany, and most mid-sized companies have never heard of it. It is called A5, comes from the German Federal Office for Information Security and is available as a draft for comment. Anyone who is thinking today about how to demonstrate the trustworthiness of their AI at some later point should know it. Not because it already applies, but because it has a gap that sits exactly where mid-sized companies stand.
What A5 is, and what it is not yet
A5 stands for AI Audit and Assurance Assessment Architecture. A preliminary remark on the abbreviation, because the pointer to it often comes via British colleagues: BSI here is not the British Standards Institution, but the German Federal Office for Information Security (Bundesamt für Sicherheit in der Informationstechnik). A government authority, not a standards body. The same three letters, an entirely different role.
And the most important thing first: A5 is a community draft. A draft on which the professional community can give feedback. Today there is no certification to A5, no assurance report, no evidence that you could present to a customer. Anyone who sells you something else has not checked the status.
A draft status is still no reason to look away, and for a sober reason: it shows you the direction before it becomes binding. Anyone who reads today which evidence an assurance architecture will want to see later can align their records with it while that still costs nothing. Anyone who waits until the catalogue is finished collects evidence retrospectively, and retrospectively much evidence no longer exists. Logs that were never written cannot be made up for.
A draft is not a requirement. But a draft that comes from a federal authority and builds on an established audit catalogue is not a finger exercise either.
Where the methodology comes from, and why that explains a lot
A5 does not reinvent auditing. It adopts the methodology of C5, that is, the criteria catalogue the BSI has established for cloud services, and relies on the international assurance standard ISAE 3000. That is a decision with consequences, and it explains almost everything that follows.
ISAE 3000 is the world of assurance engagements, in which a practitioner issues an assurance report on a subject matter. In Germany that is traditionally the world of the public auditors. The ISO world works differently: there an accredited certification body certifies a management system under the rules of ISO/IEC 17021-1, supplemented for AI management systems by ISO/IEC 42006.
So there are two worlds of evidence, both legitimate and both attesting to something different. The one confirms a subject matter at a point in time or over a period. The other confirms that a system exists, is lived and is improved. Anyone who needs both is audited twice.
In practice the difference is more tangible than it sounds. An assurance report under the audit logic answers the question: did this system meet these criteria in this period? A certification answers: does this organisation have a procedure that keeps such questions answerable in the long term? The first is a snapshot with a seal, the second a commitment about the future. Customers mostly ask for the second and mean the first, and vice versa.
| Assessment under A5 (planned) | Certification to ISO/IEC 42001 | |
|---|---|---|
| Subject of the assessment | The AI system and its trustworthiness against a criteria catalogue. | The organisation’s management system for AI. |
| Methodical basis | The C5 scheme and the assurance standard ISAE 3000. | ISO/IEC 17021-1, supplemented for AI by ISO/IEC 42006. |
| Who assesses | A practitioner who performs an assurance engagement. | An accredited certification body with an audit team. |
| What you have at the end | An assurance report on a subject matter, relating to a point in time or a period. | A certificate for a defined scope, with surveillance audits in the cycle. |
| Availability today | Not yet. Community draft, comment period ended on 31 August 2026. | Available since 2023, certifiable worldwide. |
The last row gives the practical answer to the question I am currently asked frequently in conversations: should I wait with 42001 until A5 is finished? No. You cannot be assessed today against something that does not yet exist, and the groundwork is the same in both worlds. Records, roles, risk procedures and supplier control are needed by every assessment, whatever tradition it comes from.
The gap: ISO/IEC 42001 does not appear
I read the BSI’s A5 page with this in mind on 17 August 2026. The EU AI Act and the Cyber Resilience Act are named as the regulatory framework, C5 and ISAE 3000 as the methodology. ISO/IEC 42001, the international standard for AI management systems, is not mentioned.
That is not a reproach, a draft may have gaps, that is what the comment period is for. But it is a practical problem for everyone who has already set off. If you are building an AIMS to ISO/IEC 42001 today, you have an AI policy, a risk procedure, an impact assessment, supplier procedures and records. That is exactly the material from which an A5 assessment would draw its evidence. Without a mapping between the two you have to prepare this material twice, once for the standard and once for the assurance architecture.
A mapping of the A5 criteria to the clauses and Annex A controls of 42001 would therefore not be a luxury, but the difference between one body of evidence and two. That is exactly a point that can be raised as part of the comment process.
Behind this lies an experience from every integrated management system: duplicated evidence is not maintained twice. It is maintained once and forgotten once, and after a year nobody knows which of the two versions is current. Anyone who runs two evidence paths side by side without connecting them ends up with an additional contradiction that an auditor finds. That is not double security.
The same question, incidentally, also arises in the other direction. If an organisation has an assurance report under the C5 scheme for its cloud, how much of it can be used as evidence in a 42001 audit? These interface questions decide whether an assurance architecture lowers or raises the effort in the market. In the draft phase you can ask them, afterwards no longer.
Two modules, and who they are meant for
A5 has a modular structure. So far there are two modules: a horizontal base module on trustworthiness, designed to be independent of technology and application, and an operations module for the cloud infrastructure with a cross-link to C5.
Look closely at this split and you will see where you stand. A provider that develops an AI system and runs it in its own cloud finds itself in both modules. A mid-sized company that uses a purchased AI system finds itself mainly in the base module and otherwise not at all. For the role that the vast majority of companies are actually in, namely deployer of someone else’s system, the matching module is missing so far.
That matters because the obligations hang precisely on this role. In Article 25 and Article 26, the EU AI Act distinguishes cleanly between the responsibilities along the value chain and the obligations of deployers. An assurance architecture that primarily addresses providers and cloud operation leaves the larger group out for the time being.
That, too, is no reason for inaction, but an indication of what to prepare for. A deployer is measured by different things than a developer: by use in accordance with the intended purpose, by human oversight, by the instruction of its own people, by the monitoring of ongoing operation and by whether it notices when its provider changes something. These points are already in the EU AI Act and in ISO/IEC 42001 today. Anyone who puts them in order also puts in order what a later assessment will want to see.
What you should do now in concrete terms
- Know the status and classify it correctly: A5 is a draft. There is no A5 certification today, and nobody can demand one.
- Do not wait. The pragmatic route to evidence still runs through a management system to ISO/IEC 42001, supplemented by the obligations from the EU AI Act. Anyone who has that is better prepared for any assurance architecture than someone waiting for the finished catalogue.
- Determine your own role. Provider, deployer or distributor decides which part of a future assessment concerns you at all.
- If you are affected professionally: follow the final version. The comment period for the community draft expired on 31 August 2026. A draft on which only large providers have given feedback becomes a catalogue for large providers. So read the final version with that in mind.
The last point is the one I see underestimated most often. With standards and catalogues, the draft phase decides for whom they will be practicable in the end. Anyone who stays silent in this phase should not be surprised later that the finished catalogue presupposes conditions they do not have.
If you want to know where your use of AI stands today, before any catalogue is finished: the readiness check tells you in a few minutes where you stand, and the result holds regardless of which assurance architecture prevails in the end.
Go to the AI readiness checkMy assessment as an auditor, and it is an unexcited one: A5 is good news. A government body is building a methodology that sits on a proven catalogue instead of inventing something new. The fact that 42001 does not appear in it today is a gap in the draft and not a verdict on the standard. Until it is closed, what applied before applies in practice: anyone who manages, documents and evidences their AI properly is prepared. Anyone who waits for the finished catalogue loses time they will not have later.
Primary sources
Frequently asked questions
Can you be certified to BSI A5 today?+
No. A5 exists as a community draft, that is, a draft for comment. There is no A5 certification and no A5 assurance report, and nobody can demand one.
Is BSI here the British Standards Institution?+
No. What is meant is the German Federal Office for Information Security. The confusion is understandable when the pointer comes via British professional circles, but it leads to the wrong organisation.
Does A5 replace ISO/IEC 42001?+
No, the two attest to different things. A5 follows the assurance logic of ISAE 3000 and C5, while ISO/IEC 42001 is certified through accredited certification bodies. The current draft version does not mention 42001.
Until when can the draft be commented on?+
The comment period for the community draft ended on 31 August 2026. As of 10 November 2026 the BSI had not published a final version.
Author & expert review: Lars Zimmermann · ISO/IEC 42001 Senior Lead Auditor & Senior Lead Implementer · ISO/IEC 27001 Lead Auditor & Lead Implementer (PECB)
Last updated: 10 November 2026. Researched and reviewed to the best of our knowledge; not a substitute for individual legal advice.
Questions about your own case?
In a free 15-minute intro call we assess where you stand on ISO 42001, ISO 27001 and the EU AI Act, honestly and without a sales pitch.