Banking's real AI blocker isn't the model. It's the overnight batch.
Banks have spent two years running AI pilots. In 2026 the conversation has shifted from what models can do to whether the bank’s architecture can keep up. Three deadlines and trends make the point.
T+1 is coming, and batch is the bottleneck
UK and EU markets move to T+1 settlement in October 2027, giving firms about a year to prove operational readiness. Shorter settlement windows, real-time payments and Pay by Bank growth all expose the same weakness: core processes that still depend on overnight batch cycles. Risk, liquidity and exception management must move toward intraday, event-driven processing.
AI governance is becoming a supervisory expectation
On 7 October, the Monetary Authority of Singapore issued its Guidelines on AI Risk Management for financial institutions, covering generative AI and AI agents. Board and senior management oversight and core AI risk management systems are expected by October 2027, and AI life cycle controls and capability by October 2028. Singapore often signals where other regulators go next, and the EU AI Act is already pushing in the same direction.
Resilience now assumes compromise
At Sibos 2026, a recurring theme was operating under an assumption of compromise, with cyber resilience working at machine speed. This week’s AI-discovered PHP flaws show why.
What this means for the enterprise architecture
- Decouple the core. Put an event streaming and API layer around legacy cores so new services, including AI agents, consume events rather than batch files. This is also the path to intraday processing for T+1.
- Build an AI inventory and control plane. Keep one register of models, agents and AI vendors with owners, risk ratings and evaluation results. It is the foundation for any supervisory review.
- Treat agents as regulated actors. Give each agent its own identity, scoped permissions and full audit logs. Human approval should stay in place for high-impact actions such as payments, credit decisions and customer communications.
- Design for third-party AI risk. Most banks will consume AI from a few large providers. Contracts, exit plans, model change notices and fallback options belong in the architecture, not only in procurement.
- Measure architecture maturity, not pilot count. The banks that pull ahead will be the ones whose platforms let them ship, monitor and retire AI capabilities safely and quickly.
The takeaway
The bank that wins with AI in 2027 will not necessarily have the best model. It will have real-time data, clear ownership and governance that a regulator can follow end to end.
KSL works with organisations on enterprise and AI architecture, from modernisation roadmaps to agent governance. Next week in this series, we look at another industry.