Compliance Automation for a Safety-Critical National Regulator
90% reduction in analyst review time on automated workflows.
Theme Risk, Compliance & Governance · Also Cost
In brief
Situation. A national regulatory agency, responsible for one of the most safety-critical sectors of its economy, employed deep specialist analysts whose judgment is genuinely irreplaceable.
Complication. Those specialists were spending a large fraction of their week not on substantive decisions, but on procedural review, the kind of work where the document either ticks the box or does not. Meanwhile the regulated industry was moving faster, and the backlog was growing.
Resolution. I led the engagement end-to-end across three blocks. First, data infrastructure modernisation and migration, with data governance built for a regulated environment. Second, cybersecurity hardening across the analytical and operational stack. Third, targeted AI integration using Google Gemini, scoped narrowly: not to make compliance decisions, but to prepare, summarise, and pre-classify the inputs that specialists then review. We applied EU AI Act-equivalent principles: explainability, auditability, clear separation between AI suggestion and human decision.
Impact. 90% reduction in analyst review time on automated workflows. Specialist capacity released for the high-judgment decisions only they can make. Infrastructure foundation now supports continued AI expansion across other compliance functions.
The longer story
If you talk to anyone who works in a serious regulator, they will eventually tell you the same thing: “My problem is not the hard cases. My problem is the easy cases drowning out the hard ones.”
This is the part the AI-replaces-humans narrative gets exactly backwards. Nobody wants a machine deciding nuclear safety questions. But everybody, including the regulators themselves, wants a machine doing the photocopying that surrounds the decision. We built precisely that: an AI that prepares, sorts, and pre-flags so that a specialist who used to spend Monday reading 200 pages to find the three that mattered, now opens Monday morning with those three pages already on top of the stack.
The specialist still decides. That is the entire point.
The deeper lesson is one that comes up in every regulated environment I work in: the political legitimacy of AI in government comes from where you draw the line. Draw it in the wrong place and the whole programme dies in a press conference. Draw it where the humans still own the verdict and the AI only handles the queue, and suddenly you have a productive, trusted partnership.
90% less time on the boring bits means more time on the bits that actually keep a country safe.