Practical writing on AI program leadership, governance, and execution inside regulated financial services institutions. Less theory, more patterns I'm actually seeing.
Every POC in the last three years has passed. None of the corresponding production deployments shipped on time. The POC isn't testing the right things — here's how to design one that actually tells you whether the production deployment will succeed.
Read articleThe prioritization workshop produced a ranked list of twenty-seven use cases. Number one was real-time fraud detection. Two years later it was still first — blocked on the same three things that were visible in the original workshop if anyone had looked closely. Here's what conventional prioritization frameworks consistently get wrong, and a corrected framework that actually predicts what ships.
Read articleThe board deck had a section called "AI Risk." It listed twelve categories. No dollar amounts. No mitigations. No ask. The board noted it and moved on. Six months later, the examiner arrived and asked to see the board's AI risk oversight documentation. That slide deck was the documentation.
Read articleThe CFO asked what the program had cost so far. The number was $800,000. Nobody mentioned the three data scientists who had quit, the vendor contract 60 days from termination, or the fact that the board had stopped asking about AI. Here's how to calculate the full cost — including the costs that never appear in the budget.
Read articleThe program had an excellent sponsor for the first six months. Then she got promoted. The steering committee continued meeting. Nothing moved. The difference between a sponsor who drives a program to production and one who doesn't isn't seniority or enthusiasm — it's a specific set of behaviors most sponsors are never told they need to exhibit.
Read articleThe bank hired three data scientists. They spent eight months building models that never reached production — not because the models were bad, but because nobody was leading the program work. Here's the right sequence for building an AI team at a mid-market bank, and why the role everyone hires first is the wrong starting point.
Read articleThe fraud model went into production at 89% precision. Eighteen months later it was at 71%. Nobody had noticed — not the data science team, not Model Risk, not the business line that owned the output. The first signal was a spike in customer complaints. Here's why drift is a more serious problem at a bank than almost anywhere else.
Read articleThe contract said eight months. It had been sixteen. The vendor was citing the bank's IT team. The bank was citing underdocumented API requirements from the vendor. The CFO had started asking her EA to pull the original contract. Here's the predictable sequence by which AI vendor relationships go wrong — and the three options when yours has.
Read articleThe board read a McKinsey report on AI in banking. Most of the use cases required ten years of data, a dedicated ML team, and regulatory infrastructure the bank was years from having. The meeting ended without a decision — not for lack of ambition, but because they were measuring themselves against the wrong institution.
Read articleThe vendor said "autonomous." The demo showed the AI drafting an email. That is not autonomy. Autonomy is when the system sends it. Here's what agentic AI actually means on a spectrum from AI with tools to AI that acts — and what governance infrastructure a regulated institution needs before deploying anything genuinely autonomous.
Read articleThe steering committee met monthly. The program lead presented a status deck. The committee asked questions. Nobody made a decision. The next meeting was scheduled. Here are the three design choices that determine whether a committee steers or just meets.
Read articleThe data scientist spent three weeks on the pilot dataset. It performed at 91% accuracy. In production, the same model performed at 67%. The difference was the data — here's what AI-ready data actually means, where mid-market banks have better data than they think, and where they have worse.
Read articleThe usage metrics came back at 23%. The model was working. The operations team had built a spreadsheet workaround during testing and never stopped using it. Here's the change management problem nobody talks about when deploying AI at a bank.
Read articleThe pilot was budgeted at $400,000. By month eight the ask was $1.8 million. Here are the five cost buckets that belong in every AI program budget, the line items that always get cut and always cause problems, and how to present the real number to a CFO anchored to the pilot cost.
Read articleThe board asked for an AI strategy. What came back was sixty slides and a mission statement. Nothing changed. Here's what a real strategy is, what a roadmap is, why they're different, and what order to build them in.
Read articleThe examiner scheduled a review of the bank's AI systems. The CRO asked for a list of everything in production. It took three weeks to compile and came back incomplete. Here's what examiners are actually looking for — and the three gaps most institutions have when they arrive.
Read articleEvery major AI vendor has a financial services pitch deck. The evaluation criteria that matter for a $10B bank are different from what works for JPMorgan — and most mid-market buyers use the wrong criteria.
Read articleThe board approved an AI initiative. The CTO is overloaded. The data science team can execute but can't drive cross-functional alignment. Nobody is driving. Here's what a fractional arrangement actually looks like, when it fits, and what to look for when hiring one.
Read articleThe deck had forty slides and an 87% accuracy improvement. The CFO asked what happens if it doesn't work, and nobody had an answer. Here's what a CFO actually needs to see before approving an AI program — and what to leave out.
Read articleMRM sent back thirty-seven questions. The team spent three months writing answers. MRM came back with twenty more. The path through this isn't more documentation — it's a different conversation, started at a different time.
Read articleThe pilot worked. The data scientist showed the chart. Everyone clapped. That was nine months ago. Here is a practical sequence for getting it unstuck — based on what actually moves these programs inside regulated financial institutions.
Read articleThe model isn't the problem. It almost never is. After watching dozens of AI initiatives stall inside large financial institutions, the patterns are remarkably consistent — and so are the things that actually get them moving again.
Read articleMost AI governance frameworks are designed by people who have never had to defend a model to a regulator. The ones that work treat governance as an enabler, not a gate — and start with the audit trail in mind.
Read articleIf any of this resonates with what's happening at your institution, I'd be glad to compare notes — even if it doesn't lead to an engagement.
Email me