AI Risk Control Strategy for Global Banking Operations
I first noticed something off in 2022. A tier-one bank’s AI reported zero issues across a full quarter of transaction monitoring. The risk committee reviewed the numbers, saw the clean run, and moved on. False positives were down. Processing time had dropped by sixty percent. By every measure they had, the model worked exactly as designed. A mid-level analyst thought something felt wrong. Not a hunch without reason. She had fifteen years in correspondent banking. Fifteen years watching money move through nested accounts, cross-border flows, paper entities that seemed solid one day and vanished the next. She read those rhythms like a cardiologist reads an ECG, not just the spikes, but the silences that shouldn’t be that quiet., – The Situation She flagged a counterparty. When I asked what had triggered her concern, she hesitated. The honest answer? She couldn’t fully explain it. The flows looked normal. The entity had paperwork. The model had processed and cleared it without a hitch. No single data point stood out. What she had was a shape. A pattern she’d seen before, not identical, but close enough that fifteen years of experience made her notice. The counterparty was structuring: deliberately breaking transactions into pieces to stay below automated detection limits. The method wasn’t new, but the setup, the jurisdictions, the counterparty type, the timing, fell outside the model’s training data. It had never encountered this exact arrangement, so it said nothing. Here’s what haunts me about that year. The model wasn’t broken. It did exactly what it was built to do, precisely. The risk committee wasn’t careless. They reviewed the outputs they were given. The governance process looked correct from every angle it could see. That’s the problem. The system had no visible failure mode for the people responsible for spotting failures. It had no red light. It only had the absence of one, and the organisation had, over time and without anyone saying it aloud, learned to treat that silence as safety., – What Actually Failed The first failure wasn’t the model. It was the way we thought about what the AI was doing. Some AI risk governance treats model metrics as a stand-in for real-world coverage. False positives. Processing speed. Accuracy on test data. These numbers matter. They tell you something real. But they don’t tell you what the model has never seen. A model trained on past transactions will catch patterns it knows. It won’t notice when the world changes, when a new structuring trick emerges, when a rarely used jurisdiction becomes a conduit. It can’t warn you about its own blind spots. That’s not a flaw in design. It’s how these systems learn. The second failure was subtler. When AI metrics look good for long stretches, organisations adjust their human oversight accordingly. Senior analysts spend less time on cleared transactions. Review processes thin out around the automated layer. That makes sense, you wouldn’t manually check every calculation a spreadsheet makes. But the analogy is wrong. A spreadsheet follows rules consistently. An AI model learns patterns from data, and the patterns it never saw are the ones it will never find. The oversight we’re cutting back is exactly the oversight we need to catch what the model misses. The third failure is the one that stays with me. It’s what happened to the analyst’s instinct inside the organisation before that moment. She had flagged things before that didn’t turn into confirmed issues. That’s how real pattern recognition works, not every signal leads to a finding. In some places, repeated flags without outcomes become a professional liability. Analysts learn to adjust their instincts to match what the model approves. The pressure, unspoken but real, pushes toward alignment with the machine. When the machine says nothing is wrong, insisting something is feels risky. Organisations can quietly erode that confidence over time without meaning to., – What This Means for Your Organisation If you run AI-assisted transaction monitoring, or any AI-assisted risk function in a regulated setting, the question isn’t whether your model performs well. It’s whether your oversight is built around what the model cannot see, or whether it’s been quietly reshaped around what the model can process. Those are not the same structures. The institutions getting this right aren’t choosing between AI capability and human judgment. They’re being precise about what each can actually do. AI handles volume and applies learned patterns at a scale no team can match. Experienced analysts spot anomalies outside those patterns, not because they’re better than the model, but because they’re different from it in the ways that count. The gap between the model’s world and reality isn’t something you close by improving the model. You close it by staffing it., – The best AI risk control system I’ve seen wasn’t the one with the strongest model. It was the one that knew, without doubt, where the model ended and what had to happen next. A system that understands its limits is still a system. One that doesn’t is a liability wearing impressive metrics.
AI Implementation Success: An OCC Compliance Story
When the OCC Said Yes: What a §17f-1 Fix Taught Me About AI in Regulated Finance Regulators do not applaud. They document, they question, they reserve judgement, and occasionally-very occasionally-they express satisfaction. That last phrase, in OCC examination language, is roughly equivalent to a standing ovation from a Scandinavian audience. Last year, when I heard it, I did not celebrate immediately. I went back through the file to check whether we had missed something. We had not. But the reason we had not is more instructive than the outcome itself. — The Situation The custody bank came to me with a §17f-1 problem that had quietly compounded for longer than anyone wanted to admit. Fourteen custodial accounts. Manual reconciliation spread across three jurisdictions-the US, Luxembourg, and a Cayman structure that generated its own particular brand of administrative joy. The average lag between identifying a securities fail and reporting it to the OCC examiner’s desk was eleven days. Eleven days is not a compliance gap. It is a liability with a bow on it. Here is the moment I do not enjoy recounting. In the first working session with the internal operations team, I asked to see the reconciliation workflow. What I expected was a documented process with some inefficiencies. What I found was a spreadsheet. A colour-coded, lovingly maintained, deeply human spreadsheet-owned by one person, checked on her schedule, dependent entirely on her being in the office on a Friday afternoon and not having a migraine. She was excellent at her job. She was also the single point of failure for a regulated function that the OCC takes seriously enough to have its own numbered rule. I had seen versions of this before-in Bahrain, in Mumbai, in London-but something about seeing it at a US custody bank in 2024 still caught me. The gap between what institutions tell regulators about their controls and what actually runs their controls is, in my experience, almost always a person with a spreadsheet and good intentions. We needed to close that gap with something more durable than intention. — The Build The surveillance layer we constructed pulled directly from the core custody ledger-integrated via structured API connections into the bank’s existing custody management infrastructure, which in this case sat on a platform familiar to most mid-tier US custodians. The exception logic ran continuously, not on a schedule. Every identified fail triggered an automated escalation path, timestamped at the moment of detection. SAR-adjacent flagging narratives were auto-drafted and queued for human review before anyone had to open a ticket or send a message. The tooling itself was not exotic. Structured data pipelines, rule-based exception engines layered with a classification model, and a reporting stack that wrote directly to the audit trail in a format the OCC’s examination teams could read without interpretation. Platforms like Nasdaq’s Surveillance infrastructure, Broadridge’s reconciliation and regulatory reporting suite, and AxiomSL (now part of Adenza / Nasdaq) exist precisely for this class of problem. The architecture principles are well understood. What is less understood is why so many institutions still do not implement them until an examiner forces the question. When the OCC review team arrived, they saw real-time audit trails. Timestamped escalation paths. Zero documentation gaps between identification and reporting. The eleven-day lag was gone. The process was no longer dependent on a person remembering to check something. They expressed satisfaction. — Three Things That Were Actually True **First:** the technology was not the differentiator. Every vendor in that room had technology. The differentiator was that the bank finally had a single source of truth-one that did not require a human to remember, to be available, or to interpret ambiguous data under time pressure. The OCC was not impressed by AI. They were impressed by accuracy. AI made accuracy repeatable. That is a meaningfully different claim, and most sales decks in this space get it backwards. **Second:** the eleven-day lag was a symptom, not the disease. The real problem was that no one had priced the exposure correctly. Eleven days of unreported lost or stolen securities is eleven days of regulatory, reputational, and counterparty risk sitting off the risk register. Until you can see the gap in real time, you cannot price it. Until you cannot price it, you will not fix it. Visibility is not a compliance nicety-it is the precondition for every risk decision that follows. This connects to something I wrote about separately: the structural danger hiding inside AI-native clearing approvals, where boards are signing off on automated systems they do not fully understand, compounding the very exposures they believe they are managing. **Third-and this is the one most organisations get wrong-the examiner is not your adversary.** Build for the examiner who assumes the worst. Give them audit trails so clean they have nothing left to question. The institutions that treat regulatory examination as an adversarial event spend enormous energy managing the optics of their controls. The institutions that treat it as a transparency exercise spend that same energy making their controls actually work. One of those strategies scales. The other one ends badly in year three. — What This Means for Your Organisation If your reconciliation workflow depends on a person, a schedule, or a spreadsheet-however capable the person, however reliable the schedule-you are one absence, one error, or one examiner visit away from an eleven-day problem of your own. The technology to close that gap is not emerging. It is available, it is implementable, and in most custody environments it is not even particularly expensive relative to the exposure it eliminates. The question is not whether you can afford to build the surveillance layer. The question is whether you can afford to keep explaining to an examiner why you have not. I have written before about what it looks like when a team holds together under that kind of pressure-the thirty-six-hour stretches, the decisions made at 3am that determine whether the morning looks manageable. None of that replaces the upstream work of building systems