AI & agentic AI security architecture
You are being asked to approve AI systems faster than your control framework can assess them, and the failure modes do not look like the ones your existing standards were written for.
For banks, asset managers, custodians, the fintechs that sell to them and the law firms that advise them.
In regulated organisations, AI adoption often stalls at approval. Business teams move faster than the assurance function can assess, and the existing control framework rarely describes the failure modes that apply here: prompt injection, tool misuse, data leakage through context windows, model poisoning, and agents acting autonomously with more permission than they need.
Where I usually start
Threat modelling the specific system. An internal retrieval-augmented assistant and a customer-facing agent with tool access carry different risks. Treating them the same way produces controls that are too heavy in one place and too weak in another.
A policy connected to something that enforces it. An acceptable-use standard with nothing behind it adds exposure without reducing any, so the first question is what will enforce it.
Where your own applications call models, that is usually an AI gateway: one place to apply routing, logging, data-loss controls and model allowlisting. Where the AI arrives inside a product you have bought, such as Copilot or a vendor’s assistant, enforcement sits in tenant configuration, identity and access, data-loss rules and the contract.
Agentic permissioning. Non-human identity is where most existing IAM models struggle. Agents need scoped, revocable and auditable credentials, and conditional access frameworks that treat machine identities as first class.
The tools an agent can call. An agent’s risk sits less in the model than in what it can reach. Tool connectors and MCP servers give it mailboxes, document stores, ticketing and line-of-business systems, often through one service account with more permission than any single user has. The questions are what each tool can read and change, whether the call runs as the agent or as the user, and what stops an injected instruction becoming an action. An agent tool-access review answers them for one agent.
When the business has already bought it
Increasingly the tool arrives first. The business decides to onboard a product before security has seen it, sets a date, and then needs it through governance in a rush. Where authority is spread across many senior people, as it is in a partnership, there is rarely one person who can clear the path.
The reluctance of the security and risk functions is usually reasonable. They are not against the tool. They have objectives and roadmaps already committed, and rearranging them around an unplanned onboarding has a real cost.
That is the gap I fill. A business lead can engage me directly to do the security and risk work the review needs: threat-model what has actually been bought, design the controls around it, and assemble the evidence the reviewers ask for, without either function having to set aside its own objectives to produce it. The decision stays with them. What changes is that they are assessing a properly prepared case rather than a request to hurry, and a no comes with what it would take to get to yes.
Legal and contractual risk stays with your lawyers; I do not advise on it. Nor will I act as the independent reviewer on a system I have prepared for approval.
When you are the vendor
The same review looks different from the other side. An AI vendor selling into a bank meets the bank’s security and architecture review, which tests the product against the bank’s own standards and patterns. Before a proof of concept, that review mostly asks whether the right controls are possible, not for an audit of them. Vendors who arrive with prepared answers move through it. Vendors who find their gaps in the meeting wait for the next one.
I have written AI security standards and run approvals from the buyer side at a global custodian bank and a global asset manager, so I know what those reviewers ask for and what evidence satisfies them. The AI vendor assurance pack builds an answer sheet to the questions a major bank will ask, and works through it with you, so you know your gaps before the bank does. The answers are yours, and I won’t review the same product for a buyer.
Regulatory context
There are no AI-specific rules for UK financial services. The FCA and the PRA expect AI to be governed through the frameworks firms already apply, so the questions an AI system has to answer come from those:
- DORA, for EU financial entities and UK groups with EU subsidiaries. A model provider behind a business process is an ICT third-party dependency, with a place in the register of information and the contract terms of Articles 28 to 30.
- PRA SS2/21 and FCA SYSC 8 on outsourcing and third-party risk management, which a hosted model or AI product usually falls under.
- Operational resilience (PRA SS1/21, FCA PS21/3). An AI system supporting an important business service has to stay within that service’s impact tolerance.
- PRA SS1/23 on model risk management, which covers AI models. It applies directly to banks with internal-model approval for regulatory capital, and is read as good practice well beyond them.
- The EU AI Act, where a firm places an AI system on the EU market or its output is used in the EU. Credit scoring of individuals, and pricing for life and health insurance, are among its high-risk uses.
Your model provider is an ICT third party covers the third-party side in more detail.
How this differs from an AI policy document
Plenty of firms will sell you a policy. The harder work sits underneath it: what enforces the policy, what evidence it produces, and how a reviewer decides whether a given system passes.
Based in Edinburgh
I have worked in Edinburgh for seven years, including as a security architect at a global asset manager and a global custodian bank here. The practice works across the UK, mostly remotely, with time on site in Scotland wherever the work benefits from a room: workshops, design authority sessions, and risk committee conversations that go better face to face.