Can banks match fintechs in scaling agentic AI?
Fintechs are pulling ahead. Based on a global sample of 4,000 firms, they account for nearly 70% of AI initiatives. Banks are still running pilots, slowed by regulatory complexity, fragmented tech stacks, and organizational inertia.
The catch: the highest-growth opportunities are in agentic and revenue-generating use cases. Think AI-driven multi-asset trading, advanced predictive decision management, and decision agents that act within guardrails. That is where fintechs are already deploying at scale.
What banks are doing vs. what fintechs are shipping
Most incumbent programs are aimed at reliability and trust. Useful, but narrow: treasury automation, chat-based assistants, and advisory personalization. These risk commoditization and leave upside on the table.
Fintechs are leaning into action-oriented systems. Examples include AI-driven mobile trading with real-time, algorithmic decisions; agentic companions embedded in consumer flows; and rethinking end-to-end payment experiences.
Notable moves in market
- SMBC created a Singapore startup unit to test agentic AI faster, reducing internal friction and time-to-learn.
- DBS launched a gen AI assistant for corporate clients to handle round-the-clock banking queries.
- OCBC rolled out an AI stock advisory service trained on 4,000 equities.
- Ant International embedded an agentic travel companion (Alipay+ Voyager) to assist with itineraries, bookings, and in-merchant purchases.
- Stripe is exploring agentic AI to reshape the entire payment flow from shopping through checkout.
Agentic AI, in practical terms
Agentic systems don't just summarize. They plan, decide, and take bounded actions across tools and data sources, with policy and risk controls in place. For finance, that translates into better execution, faster decisions, and fewer manual handoffs.
How banks can close the gap-starting now
- Prioritize P&L-first use cases: cross-asset execution, risk-adjusted pricing, payment anomaly detection, cash forecasting, collections strategies. Tie each to a measurable revenue or cost metric.
- Create a ring-fenced build zone: a sandbox entity (like SMBC's approach) with model risk guardrails, synthetic data, and rapid approval paths for controlled pilots.
- Consolidate data for action: event streams, a shared feature store, and policy-controlled tool access for agents. Latency matters-target milliseconds where trading or fraud is involved.
- Bake in compliance from day one: model lifecycle governance, human-in-the-loop for high-impact actions, audit logs, and clear explainability thresholds. See SR 11-7 and MAS's FEAT principles.
- Productionize, don't just prototype: MLOps and agent orchestration, drift monitoring, action boundaries, fallbacks, and red-teaming for prompt or tool abuse.
- Rebalance teams for speed and safety: pair quant devs and ML engineers with product owners who own P&L, and embed model risk and compliance as partners, not gatekeepers.
- Partner where it accelerates outcomes: co-develop with fintechs, structure clear SLAs and revenue-share, and integrate via APIs to avoid long core-system overhauls.
- Measure what matters: spread improvement (bps), hit rate and slippage in trading, straight-through processing rates, dispute cycle time, fraud catch rate vs. customer friction, and opex per transaction.
Where agents can move the revenue needle
- Markets: idea-to-execution agents that scan signals, size risk, route orders, and monitor fills against benchmarks.
- Payments: autonomous retries, least-cost routing, and dispute summarization with evidence gathering.
- Wealth: portfolio rebalancing suggestions with pre-trade checks and client-ready rationales.
- Corporate banking: cash flow forecasting, working capital recommendations, and policy-aware self-serve servicing.
Operating model shifts that make scaling possible
- Guardrails over prohibition: action policies, approval thresholds, and intervention hooks instead of blanket bans.
- Product funding over projects: multi-quarter roadmaps with ROI targets and user adoption milestones.
- Data contracts: clear ownership, quality SLAs, and lineage for every feature used by an agent.
- Security by default: secrets isolation, least-privilege tool access, and continuous red-teaming.
Bottom line
Fintechs aren't just experimenting; they're operationalizing agentic AI where it drives revenue. Banks can catch up by focusing on production-grade, high-ROI use cases, building safe action loops, and aligning teams and governance around speed with control. The institutions that treat agents as decision and execution teammates-not demos-will see the P&L impact first.
Helpful resources
Your membership also unlocks:
