At BlackRock I embedded generative AI into reconciliation prep, exception triage, and documentation — reusable prompt libraries and agent workflows that became the team's default starting point. This is the same workflow behind every tool on this site, and the one I'd bring to an AI Operations role.
Start from the workflow, not the technology. Who does this task, how often, where does it break, and what does "done well" look like in numbers? If I can't describe the current state in one paragraph, I'm not ready to automate it.
Write the requirements before touching a tool: users, inputs, outputs, edge cases, acceptance criteria. I authored business requirements for Aladdin platform enhancements and led UAT — the discipline transfers directly. AI builds faster, which makes a clear spec more important, not less.
Use AI for speed — drafting, scaffolding, iterating — while I own the decisions: scope, trade-offs, and what "good" looks like. Reusable prompts and workflows get saved back into a library so the next build starts further ahead.
AI output gets the UAT treatment: test it against real data, real edge cases, real users. In operations I ran tightened control frameworks for a reason — a tool that works in the demo and fails in production is worse than no tool.
Ship with the metrics defined up front — cycle time, error rate, adoption — then check the numbers and improve. The exception-reduction program I led was metrics-based from day one; that's how a 70% reduction gets proven instead of claimed.