Cutting manual compliance work by eighty per cent — without breaking the regulator's audit trail.

In regulated environments, "let's add AI" is a phrase that ends careers. Every automated decision has to be explainable, auditable and reversible. Every model output that touches a regulatory submission has to be traceable back to a human-accountable control.

We had several thousand active data products across six divisions, each requiring business rules, lineage documentation, and quality controls. Stewards spent the majority of their week on repetitive enrichment and rule-writing tasks — work that was high-volume, pattern-heavy, and ripe for automation, but that sat on the critical path for BCBS 239 submissions. Getting this wrong wasn't a feature regression. It was a regulatory finding.

• Identified the right slice of the work to automate. Not "everything stewards do" — specifically the high-volume, pattern-based work where AI augmentation could be controlled and audited. Business rule suggestion, metadata enrichment, lineage description, classification proposal.
• Designed an API-led workflow architecture with GenAI capabilities embedded as assistive services. The model proposes; the steward approves. Every AI output is captured with its prompt, model version, and human-accept/override decision — that's the audit trail.
• Defined the product acceptance criteria for AI use. Explainability thresholds. Confidence-score gating. Mandatory human review for anything touching regulatory data. Wrote the criteria the platform now uses to evaluate any future AI deployment.
• Sequenced rollout to manage risk. Started with the lowest-stakes enrichment tasks across two divisions, instrumented heavily, expanded to the harder rule-generation work only after we had three quarters of stable audit evidence.
• Tracked the operational shift. Measured time-on-task before and after, defect rates, audit findings, and steward satisfaction. The 80% reduction is measured against pre-deployment time-and-motion data, not estimated.

The reason this worked is that I refused to treat AI as a feature. It was a capability inside an existing data product, with the same product discipline applied to it as any other capability — acceptance criteria, rollout staging, instrumentation, retirement plan.

The hardest conversations weren't technical. They were with the second-line risk teams who quite reasonably wanted to know what happens when the model is wrong. The answer — built into the workflow — is that nothing reaches a regulatory submission without a steward's accept-decision, and the model is positioned as decision support, not decision-maker. That framing unblocked the deployment.

The 80% reduction is the headline. The work I'd point to is the governance pattern we now use to evaluate every new AI deployment in the data office.

The pattern has since become the reference architecture for AI use inside the Chief Data Office. The same explainability gating, audit-capture, and steward-in-the-loop pattern is now applied to every new model deployment we propose.