Agentic AI in Banking Will Follow Three Tracks - Fintechs Are Ahead
AI agents are already inside most financial institutions, but they're being used very differently. Incumbent banks are squeezing out back-office efficiencies. Digital-native fintechs are embedding agents across customer experiences and product design. The gap is widening, and it favors those who treat agentic AI as core infrastructure, not an add-on.
Recent analysis from Oliver Wyman maps where this is going: three distinct tracks - the Assistance Economy, Adaptive Customer Experiences, and Agentic Twins. Market leaders will industrialize all three.
Need to know
- More than half of financial services firms use AI agents to some degree: 57% in customer service, 48% in marketing, 43% in fraud management, and 40% for productivity and research.
- Only 32% of banks see strong returns from customer-facing AI, even though 99% prioritize it. Meanwhile, 68% say the biggest value comes from back-office efficiency.
- On leading benchmarks, models outperform industry experts on roughly 30% of finance tasks. Keep humans in the loop for complex processes.
From assistants to agents
We're moving from "helpers" to systems that plan, decide, and execute. Single-command workflows will replace step-by-step hand-holding. Think mortgage agents that source conveyancers, match rates to a customer's financial profile, coordinate third parties, submit applications, and present verified options for final approval - all in one flow.
This isn't just agent-to-client. It's also agent-to-agent collaboration across firms and platforms. That's where compounding value shows up.
Why the gap is widening
Banks are winning on productivity: KYC, compliance workflows, and marketing analytics see clear payoffs. The problem is misaligned investment. Customer-facing projects are prioritized on paper but don't deliver at scale yet.
Fintechs, built on modern stacks, bake agents into the entire value chain - from product design to delivery. They benefit from consistent patterns across services, so teams can ship features anywhere in the stack without friction.
- Stripe's Payments Foundation Model processes about $1.4T annually to detect fraud, optimize payment flows, and predict disputes.
- Arta Finance uses agent suites to build portfolios-of-one for thousands of investors.
The four enablers of mass adoption
- Infrastructure: Cloud platforms and integrated data pipelines that let agents run across workflows, at scale.
- Regulation: Clear rules and safe testbeds. The FCA's Digital Sandbox is a good example.
- Digital identity: Secure, reusable identity so agents can act on behalf of verified users and only interact with trusted counterparts.
- Trust: Confidence from customers and businesses, supported by higher accuracy, explainability, and auditability.
But enablers aren't enough. Mass adoption will follow breakout use cases that fix obvious pain and outperform current processes by a wide margin.
The three emerging economies
- The Assistance Economy: Agents deliver complete experiences end-to-end, across different engagement models.
- Adaptive Customer Experiences: Interfaces adjust in real time to preferences and context, changing how products are presented and consumed.
- Agentic Twins: User-owned agents hold permissions and act with delegated authority across financial relationships.
Early signal: Capital One's Chat Concierge helps customers buy cars - preference profiling, optioning, pricing, financing, and scheduling - with participating dealerships seeing up to 55% higher engagement.
Adaptive experiences: segment-of-one
Baseline personalization isn't enough. Only 21% of banking customers say they're fully satisfied with what they get today. Leaders will move to adaptive interfaces that assemble offers on the fly using context like search behavior and session signals.
As front-ends become dynamic and data-rich, lines between channel and product start to blur. The experience is the product.
Agentic twins: identity and authority in one place
Today, customer data is scattered across products and providers. Each version is incomplete and often outdated. That slows everything down.
Agentic twins will centralize consented data under user control. Owners set explicit permissions, and twins act within those guardrails - from routine tasks to more complex decisions. For providers, this replaces manual data chasing with permissioned collaboration, faster onboarding, and richer insights.
Eight moves for financial leaders
- Show up in the new chat interface: embed existing products in conversational channels and measure hard outcomes.
- Pilot highly personalized flows end-to-end; iterate based on conversion, NPS, and unit economics.
- Roll out adaptive interfaces with human validation; tune based on uptake and product satisfaction.
- Embed digital identity: integrate Open Banking data and identity rails; align with interoperability standards.
- Keep humans in the loop for complex decisions; define escalation paths and approval checkpoints.
- Set model risk and audit standards: data lineage, prompt/version control, and action logs for every agent step.
- Upgrade data foundations: unify pipelines, address silos and unstructured data, and extend via Open Finance partnerships.
- Build an analytics layer that combines customer and clickstream data with continuous feedback loops.
How to sequence execution
Start where ROI is obvious: back-office automation that frees cost, reduces errors, and cuts cycle times. Use those gains to fund controlled customer pilots that prove lift versus a baseline.
Treat Assistance, Adaptive, and Twins as capabilities to industrialize. Don't bolt tools onto legacy workflows; rebuild the core flows that matter.
Upskill the bench
Cross-functional squads win - product, risk, data, engineering, compliance. If your team needs a fast lift on practical AI for finance roles, these resources can help:
AI tools for finance | Courses by job
Your membership also unlocks:
