Translating global standards into assessable criteria for the AI value chain

The Appia Foundation is an international, openly governed collaboration that develops the specifications organizations use to demonstrate their AI systems meet the obligations that apply to them. It builds the practical connecting layer between the standards and regulations that set expectations and the assessments that verify them. It is hosted by the Linux Foundation’s Joint Development Foundation.

Standards for AI are published, regulation is moving to enforcement, and organizations are investing in responsible AI. What the ecosystem now needs is a consistent, practical way to show that a given AI system actually meets the obligations that apply to it, in a form others can recognize and rely on across the value chain and across borders. The Appia Foundation builds that layer.

Specifications: the assessable criteria that let parties demonstrate conformity with their applicable obligations. They are organized across two layers. The Requirements and Guidance layer covers what is required. The Assessment Enablement layer covers how those requirements are evaluated. The specifications build on foundational standards from ISO/IEC and other international and sectoral standards bodies, and translate them into assessable form.

Foundational standards, from ISO/IEC and other international and sectoral standards bodies, set requirements. Appia specifications build on those standards and translate them into assessable form: the criteria that make it possible to evaluate whether a system meets the standards and the obligations that reference them. The Foundation occupies the connecting layer between standards and assessment, rather than duplicating or competing with the bodies that define foundational standards. Some Appia specifications may themselves become foundational standards over time, through the normal processes of the bodies that adopt them.

No. Many initiatives set out principles, which are high level and not directly assessable, or contribute to the foundational standards and obligations that say what is required. What is now needed is the layer that makes those requirements checkable: the assessable criteria that let a party demonstrate a system actually meets them. The Appia Foundation occupies that connecting layer, between the standards and obligations that set expectations and the assessment that verifies them. It is not another set of rules; it is what makes the existing ones work in practice.

No. Conformity with Appia specifications is a technical result, not a legal status. It is the demonstration that a system meets specified criteria. Whether that demonstration satisfies a legal obligation is determined by the framework that imposes the obligation and the jurisdiction in which it applies. The Foundation produces conformity infrastructure; it does not confer compliance, and it does not claim regulatory authority.

The AI value chain involves different parties with different responsibilities. The specifications are modular so that a party demonstrates only what relates to its role, the part of the system it controls, and the context in which it operates. An organization composes the specifications relevant to its position rather than assessing against the entire set.

When an upstream provider has demonstrated conformity for the parts it controls, that evidence can be relied on by others further along the value chain. A party building on a platform does not have to re-establish what the platform provider already demonstrated. The evidence carries forward, while each party remains accountable for its own role and use. This is what makes conformity practical in a layered ecosystem.

The specifications provide the inputs from which conformity assessment schemes can be developed: modular, scalable, and interoperable criteria against which conformity can be assessed. Whether and how the Foundation develops or owns specific schemes is deliberately kept flexible at this stage, and sits with the Steering Committee. What the Foundation commits to is producing the assessable criteria that schemes are built from.

The specifications are written by the members, in working groups open to all members. No single company, sector, or country can build this layer on its own. The founding membership spans the AI value chain and three regions: organizations that put AI to work across automation, industry, energy, medical, networks, and payments (Ericsson, Mastercard, Mitsubishi Electric, Omron, Schneider Electric, Siemens); the model, platform, and compute providers they build on (Arm, Google, Microsoft, OpenAI); the organizations that test, assess, and certify AI systems (Nemko, Naaia); and insurers that rely on credible assessment data to underwrite AI systems (Armilla AI) — drawn from across Europe, North America, and Asia. The Foundation will also form an advisory board to bring academia, government, and civil society into the work.

Yes. The specifications are publicly available and designed for broad adoption. Their value depends on being widely shared; criteria that only some could see or use would rebuild the fragmentation they are meant to remove. Some specifications carry references to foundational standards that the relevant standards bodies publish under their own terms, which may require access to those standards to implement in full.

The most direct way is to join the Foundation as a member and contribute to the working groups where the specifications are written. The specifications are being drafted now, and the work done in the coming months will shape AI conformity for years; early participation is the chance to influence the criteria while they are still taking form, across the systems and markets a member works in. Membership is open across general and contributing tiers: general members shape direction across the full portfolio of work, and contributing members engage with the specific workstreams that matter most to them. The trust this work enables is meant to be held in common, by everyone who relies on AI, not only by those who build it.