The evidence a regulator, a board or a public buyer will ask for
Model inventory, risk classification, EU AI Act and ISO 42001 alignment, adversarial testing, and sovereign deployment where data cannot leave.
Model inventory
Operating model
A policy nobody builds against fails the first audit, and a committee that blocks everything pushes teams to shadow tools with no controls at all. We wire the controls into the delivery pipeline, so evidence accumulates while the work happens.
Tiers
4
EU AI Act risk tiers, each with a different obligation set
Deadline
2 Aug 2026
when most EU AI Act obligations apply in full
Frameworks
Exposure is usually narrower than feared but concentrated in one or two systems nobody considered. ISO/IEC 42001 certifies; NIST AI RMF structures the work without it.
Security
Indirect prompt injection through retrieved documents and tool output is the route that matters once a system reads your mail. Findings are reproducible and the suite runs in CI.
Sovereignty
Air-gapped operation with offline model artefacts and package mirrors, open-weight models on your own hardware, and residency shown in logs rather than asserted in a contract.
Frameworks
Regional law
Security testing
Controls
Sovereign delivery
Who we do this for
3–4 weeks
Model inventory, risk classification and regulatory gap analysis, scoped so your own teams can remediate without us.
Model inventory, obligation mapping, prioritised remediation plan
8–20 weeks
Operating model, policies and approval gates stood up, with controls and red-team suites wired into the pipeline.
Policy set, approval gates, red-team suite, evidence pack, model cards
Ongoing
Continuous monitoring, periodic red-teaming as models change, and support through audits and regulatory enquiries.
Horizon scanning, refreshed evidence, audit and certification support
Possibly. It reaches providers and deployers whose systems are placed on the EU market or whose outputs are used in the EU, so a Gulf or Turkish institution serving EU customers can be in scope for specific systems while the rest of the estate is not. We answer that per system, not as a blanket position.
If you have no models in production and no near-term plan to deploy any, a governance programme is premature: write a short acceptable-use policy, keep an inventory from day one, and come back when the first system nears production. If you already run model risk governance, you may need an AI extension to it.
Badly designed governance does, and it pushes teams into shadow tooling with no controls at all. Proportionate gates work the other way: minimal-risk systems ship with a lightweight record, and scrutiny concentrates where it is warranted. With controls inside the pipeline, most evidence is produced automatically.
Send us the constraint you are actually up against — budget, latency, regulator, deadline.
