We did a brutal capabilities assessment today. The gap is wider than I thought. We have strong data infrastructure, but weak ML Ops. We have talented data scientists, but they're siloed and lack productionalization skills. We're missing: 1) AI product managers who understand both banking and technology, 2) ML engineers who can deploy models at scale, 3) Ethics/governance specialists who can navigate AI regulations, 4) Change management experts who can drive adoption. The build vs. buy decision isn't binary—it's a portfolio approach. Core competitive differentiators (fraud detection, credit risk) we should build in-house with deep banking domain knowledge. Commoditized capabilities (document processing, chatbots) we should buy or use foundation models. The middle ground (personalization engines, recommendation systems) requires partnerships where we retain control of the data and models but leverage external expertise.

I've realized my most important job isn't picking the right AI technology—it's building a culture where AI experimentation is encouraged and failure is a learning opportunity. Right now, GTB's culture is risk-averse and perfectionist. That's served us well in traditional banking, but it's incompatible with AI innovation. Here's what I'm changing: 1) Public celebration of 'intelligent failures' where teams tried AI, learned fast, and pivoted, 2) Allocating 10% of each team's time to AI experimentation with no ROI requirements, 3) Executive visibility—I'm using AI daily and sharing my experiences (successes and failures), 4) Redefining success metrics from 'did it work perfectly?' to 'what did we learn and how fast?'. The most powerful cultural shift: framing AI as a tool that makes everyone's job more interesting by eliminating drudgery. When employees see AI as liberating, not threatening, adoption accelerates dramatically.

After reviewing case studies from JPMorgan and Goldman Sachs, I'm convinced our approach to AI strategy needs to be more aggressive. We've been too focused on cost reduction rather than revenue generation and customer experience transformation. The key insight: AI isn't just about automation—it's about creating entirely new business models. For example, AI-driven personalized wealth management could open up mass affluent segments we've struggled to serve profitably. I'm now thinking about AI strategy in three horizons: Horizon 1 (optimize existing operations), Horizon 2 (enhance current products), and Horizon 3 (create new AI-native offerings). Our current strategy is 90% Horizon 1. That needs to change.

The regulatory landscape is evolving faster than our internal policies. EU AI Act, NIST AI Risk Management Framework, OCC guidance on model risk management—we need to get ahead of this, not react. But here's what I learned: compliance is table stakes. True competitive advantage comes from going beyond compliance to build trust. Our framework needs four pillars: 1) Technical robustness (bias testing, explainability, security), 2) Transparency (clear communication about when customers interact with AI), 3) Human oversight (humans in the loop for consequential decisions), 4) Accountability (clear ownership and audit trails). The hardest part? Balancing innovation speed with responsible deployment. We can't afford to be reckless, but we also can't afford to be paralyzed. Solution: tiered risk framework where low-risk AI (chatbots) moves fast, high-risk AI (credit decisions) moves deliberately.

I finally understand why everyone keeps talking about Large Language Models. The breakthrough isn't that they're 'smart'—it's that they can understand context and generate human-like responses at scale. For banking, this means we can finally automate complex customer interactions that previously required experienced relationship managers. But here's the catch: LLMs are probabilistic, not deterministic. They can hallucinate. This has huge implications for how we deploy them in regulated environments. I now see why the Guardian oversight model is so critical—we need AI watching AI, with humans in the loop for high-stakes decisions. The technical architecture matters: we can't just plug ChatGPT into our systems and call it a day. We need fine-tuned models, RAG architectures for our proprietary data, and robust guardrails.

The town hall revealed what I suspected: our employees are terrified AI will eliminate their jobs. But the research is clear—AI augments, it doesn't replace, when implemented thoughtfully. The key is redesigning roles, not eliminating them. Take our loan officers: AI can handle credit scoring, document verification, and initial risk assessment in seconds. This doesn't make loan officers obsolete—it elevates them to relationship managers and complex case specialists. But this only works if we invest heavily in upskilling. I'm now convinced we need three workforce strategies running in parallel: 1) Upskill existing employees on AI collaboration, 2) Hire new 'hybrid' talent who understand both banking and AI, 3) Create clear career pathways that show how AI creates opportunities, not threats. The most powerful insight: employees who use AI daily become advocates. Fear comes from distance.

We have beautiful AI principles on paper. Now we need governance structures that actually work. I visited a competitor who nailed this—they have a three-layer governance model that's surprisingly simple: 1) Executive AI Steering Committee (strategy, budget, risk appetite), 2) AI Review Board (cross-functional approval for high-risk use cases), 3) Embedded AI Champions in each business unit (day-to-day execution). The key insight: governance can't be centralized in IT or a separate 'AI team.' It has to be distributed with clear accountability. We're implementing quarterly AI audits where the Guardian agent reviews all production AI systems for drift, bias, and compliance. Human experts review the Guardian's findings. This creates a continuous improvement loop. Another critical lesson: governance isn't bureaucracy if done right. Fast-track approval for low-risk use cases. Rigorous review for high-risk. Differentiate.