TL;DR
BSI, the UK’s national standards body, launched three AI assessment and certification schemes on 14 October: independent performance testing for generative AI systems and large language models, product-level AI risk management assessment against ISO/IEC 23894 and the NIST framework, and a readiness scheme for high-risk biometric systems facing conformity assessment under the EU AI Act. They sit on top of BSI’s existing ISO/IEC 42001 certification. For UK companies selling AI into Europe, the biometric scheme is the practical one, since BSI says it is seeking Notified Body status for AI.
BSI pitches the set as assurance “across the entire AI stack”, from a single use case or agent to lifecycle certification of a whole system. Omkar Joshi, its managing director for AI and digital solutions, said the ambition was to be “a one-stop shop” for AI providers and for businesses that build AI into what they do.
The three schemes
The performance scheme puts a company’s derivative model through independent tests scored on quality, safety and whether it does the job claimed, and is aimed at organisations that have built on a foundation model and want independent evidence of how it behaves.
The risk management scheme is for model makers. It examines how an AI product’s risks are found, contained and governed, judged against established frameworks such as ISO/IEC 23894 and NIST’s AI RMF, and targets the layer where AI meets users and business processes.
The third scheme is a biometric AI pilot programme for organisations preparing for the conformity assessment the EU AI Act requires of high-risk biometric systems. BSI says it is developing the pilot as it moves towards Notified Body status under the Act, which would let it carry out those assessments itself.
The case BSI makes
The release leans on two pieces of research. FocalData found that 67% of business leaders would place more trust in AI tools validated by an outside organisation. BSI’s own study last year found only a quarter of leaders ran a formal programme of AI governance, and 28% knew which data sources trained or ran their models. BSI also notes that UKAS, ANAB and RvA each accredited it for ISO/IEC 42001 certification before any other body.
Looking forward
Third-party testing is becoming the common answer to a question regulators, buyers and the public are all asking: independent studies such as Full Fact’s chatbot benchmark and the Home Office’s own spot checks on its asylum summariser both turn on who checks the output. In our view the scheme that matters most commercially is the EU one. A UK company shipping biometric AI into the EU will need a Notified Body eventually, and a domestic route to one is worth more than another certificate on the wall. The open question is how BSI’s performance tests will be scoped, because “fitness for purpose” is only as strong as the criteria a client agrees to be measured against.