DBS CEO on AI Agents: Resiliency Must Still Beat Speed as Models Rush to Production

For years, Singapore's banks have treated AI as a way to make internal processes faster — sharper credit decisions, better fraud detection, leaner back-office workflows. But as those same banks push AI agents closer to the systems that actually serve customers, a different kind of question surfaces: how fast can they safely move at all. The tension between shipping AI quickly and keeping the controls that banks have spent decades building is becoming the defining operational challenge of this phase of adoption.

At the Singapore Computer Society's Tech3 Forum, DBS Group CEO Tan Su Shan laid that tension out plainly. The pace at which new AI models and patches arrive, she said, is straining how companies normally put technology into production — and firms now face a trade-off between resiliency and speed that they rarely had to make before. She was direct about where DBS's own line sits: "Resiliency over speed. Resiliency over innovation." But she acknowledged that the era of autonomous agents may force companies to concede a little of that resiliency when the situation demands it.

The stakes are concrete for DBS, where employees now build and run their own agents on DBSGPT, an internal platform hosted on the bank's own premises that deliberately keeps APIs from pulling out its full data estate. Agents touching production systems face tighter limits, precisely because an agent can act across systems rather than just answer a prompt. Once it handles customer information or moves inside a live business process, a weak control stops being an inconvenience and becomes a genuine risk — which is why Tan stressed data ownership, usage rights and accountability as prerequisites before any agent talks to a customer.

The conversation also carries a sharp cybersecurity warning. Tan noted that autonomous agents can hunt and stitch together vulnerabilities far faster than a human attacker, effectively compressing the window a bank has to patch a weakness. That is why DBS still routes new software through its full development lifecycle, including user acceptance testing, before anything nears production. The open question she put to the industry is when a team is willing to trade a bit of that discipline for speed — and the answer, she argues, depends on what the AI is allowed to do, with research and writing tools warranting more flexibility than agents that can change records or run inside critical systems.

Why it matters for Singapore: DBS is not just another AI adopter — it is the clearest signal available of how a systemically important Singapore bank reconciles innovation with the stability regulators and customers depend on. As agentic AI spreads across the local financial sector, Tan's remarks give a useful yardstick for what responsible deployment looks like in practice: decide in advance which workflows can tolerate speed, and which must hold the line on resiliency. For Singapore's broader push to become a trusted AI hub, the message is that the country's most sophisticated institutions are treating AI governance as a first-order business decision, not an afterthought.