AI and agent governance advisory

We govern AI agents as accountable identities with limited, delegated authority, using our six-layer framework. Each agent is registered with a named owner, signs in with short-lived credentials, and can act for a person only within the permissions that person and the agent share. Every tool call passes a policy check, every action is written to a tamper-evident trail, and the agent's runtime is contained against hostile input. Each control names its owner and the record it produces, so your risk team can review the evidence.

AI agents should authenticate as workloads in their own right, with credentials that expire quickly and renew without human handling, and never through a person's account or a long-lived key. Suitable open standards include OAuth 2.0 client credentials, workload identity federation, mutual Transport Layer Security and short-lived signed tokens. For authorisation, an agent holds only what its task needs and, when it acts for a user, only the overlap of that user's rights and its own scope, which OAuth 2.0 Token Exchange, defined in Request for Comments (RFC) 8693, can carry. A policy check at every tool call, with human approval for irreversible actions, keeps those limits in force after sign-in.

Yes, in many cases. The EU AI Act applies to UK organisations that place AI systems or general-purpose AI models on the EU market, put AI systems into service in the EU, or provide or deploy AI systems whose output is used in the EU (Article 2). Being established outside the EU does not take an organisation out of scope; the activity decides it.

After the Digital Omnibus on AI (Regulation (EU) 2026/1744), obligations for high-risk AI systems apply from 2 December 2027 for systems listed in Annex III, and from 2 August 2028 for AI in products covered by the EU legislation in Annex I. Prohibited practices, general-purpose AI model obligations and most transparency duties already apply, on earlier dates.

We map your AI systems to the EU AI Act obligations that apply to them and write the controls you still need, with the evidence each control should produce. We can also review whether the running system matches that control set. Legal sign-off stays with your counsel, and building the controls is a separate AI engineering engagement if you want us to do it.

No. Certification to ISO/IEC 42001 is carried out by accredited certification bodies; in the UK, UKAS accredits them. We help you prepare: scoping the AI management system, mapping controls to the systems you actually run, and closing gaps before the certification audit. Mapping to ISO/IEC 42001 on its own is not certification.

ISO/IEC 42001 is a certifiable management-system standard for AI. The NIST AI RMF is a voluntary US framework organised around four functions: Govern, Map, Measure and Manage. They overlap well. We use ISO/IEC 42001 to structure roles, risk treatment and monitoring, and the NIST AI RMF as a shared vocabulary for risk, applying each only where it is relevant.

No. Any organisation putting a model in front of customers or staff has a reliability and security problem to manage. Regulation adds documentation and specific obligations, but the need for evaluation, access control and audit trails exists either way. A control set written for a real system is useful whether or not a regulator ever asks to see it.

Yes, as a follow-on AI engineering engagement. Governance work ends at the written control set and a review of the running system unless we agree implementation separately. Keeping the two apart means the advice can be reviewed independently of the build, and your own team can implement the controls if you prefer.

You receive a classification record for each AI system in scope, a written control set mapped to the frameworks that apply, an evaluation plan, and a review of whether the running system matches the control set, with gaps listed and owners named. These are documents and decisions your engineering, risk and compliance teams can act on directly.