Original analysis on AI governance and security.
Long-form, evergreen-leaning writing in Andrew's own voice — distinct from News (dated, external current-awareness) and Resources (annotated external links).
Can four people actually govern and steward a bank's AI?
The Thought Experiment is a UK-regulated retail bank scenario, built on the FCA's Mills Review: an evolving public investigation into whether a small human core can govern AI agents performing the rest of the operation, tested against a named organisational structure, a regulatory floor, and a ledger-integrity model.
Latest writing
Model Drift, Performance Regression and Versioning: Telling Them Apart
Three terms, three different things, and conflating them means monitoring the wrong one. Drift is the world moving. Regression is a measurement getting worse. Versioning is the artefact changing. For most LLM deployments the dominant cause of behaviour change is none of the classic ones. It is the model provider shipping an update you were not told about.
LLM Governance: What It Actually Means Operationally
Most AI governance programmes produce a policy, a committee and a register, then discover none of it touches the thing that actually changes: the assembly of model version, prompt, retrieval corpus and tool permissions that determines what the system does. LLM governance is an operations problem wearing a policy costume.
Prompt Injection and System Prompt Hardening: What Actually Works
System prompt hardening is worth doing and will not save you. The achievable goal is not a model that cannot be tricked, since natural language offers no way to build that. It is a system where being tricked does not matter much. OWASP's 2026 revision makes the same point: the biggest mover was not the attack, it was what the model was allowed to do afterwards.
Third-Party AI Risk: Model Provenance and Vendor Due Diligence
Most AI vendor questionnaires are security questionnaires with "AI" added to the headings. The useful version asks for the specific documentation the EU AI Act already obliges the provider to hold, and establishes in writing whether your own configuration has quietly made you the provider.
AI Governance in UK Financial Services: The Regimes That Already Apply
The UK has decided not to write an AI Act for financial services. That is not the relief it sounds like. Five existing regimes already bind AI in regulated firms, one of them has required a board-approved model risk appetite since May 2024, and UK lenders and insurers serving EU customers sit inside the EU AI Act whatever Westminster does.
AI Impact Assessments: DPIA, FRIA and AIA Compared
Three different assessments answer to the phrase "AI impact assessment", and only one of them may actually bind your organisation. The EU AI Act's fundamental rights impact assessment reaches a narrower set of deployers than most coverage suggests. For those inside it, completing the assessment is only half the obligation, because Article 27 also requires you to file it with a regulator.
Setting an AI Risk Appetite: A Practical Framework for Boards
Most organisations already have a risk appetite statement. Almost none of them survives contact with an AI system, because one sentence about "low appetite for operational risk" cannot tell you whether a 4% error rate is acceptable. A usable AI risk appetite is a matrix, and every threshold in it needs a trigger attached.
Undisclosed AI Interaction: The Article 50 Obligation That Was Not Delayed
The Digital Omnibus deferred the AI Act's high-risk obligations by more than a year. It left Article 50 alone. The transparency duties that took effect on 2 August 2026 apply now, to almost every organisation running a chatbot, a synthetic voice agent, or a generative content pipeline, and they carry fines of up to €15 million or 3% of worldwide turnover.