Agentic AI in 70% of Banks: Stronger Fraud Defense, Faster Operations, Humans on the Final Call

Agentic AI in Banking Operations: What 70% of Institutions Are Doing Now

Agentic AI has moved from pilot hype to execution. A 2025 survey of 250 banking executives by MIT Technology Review and EY shows 70% of institutions are already using it through deployments (16%) or pilots (52%).

For operations leaders, the takeaway is simple: this isn't experimental anymore. It's a new operating model.

What Agentic AI Actually Does (vs. Traditional Automation)

Agentic systems don't just follow rules. They read unstructured data, interpret context, decide, and act within guardrails. That's a step change from RPA.

As Sameer Gupta of EY puts it: "With the maturing of agentic AI, it is becoming a lot more technologically possible for large-scale process automation that was not possible with rules-based approaches before. That moves the needle in terms of cost, efficiency, and customer experience impact."

• Customer ops: triage service requests, route with context, draft responses, and escalate with evidence.
• Lending: pre-assemble underwriting files, validate data across systems, and flag exceptions.
• Payments: adjust recurring bills to paycheck schedules within policy limits.
• Contracts: extract key terms, compare to standards, and generate variance summaries.

Where Banks See Results Today

Fraud detection is the top live use case: 56% report high capability. Systems monitor behavior, spot anomalies, and detect AI-enabled scams like deepfakes.

Security is next at 51% high capability. Agents quarantine infected machines, orchestrate identity checks, and respond to threats faster than human-only teams.

Customer experience improvements (41% high capability) include intelligent routing, personalized advice, and faster onboarding. DBS uses agentic AI to read and classify complex SWIFT messages, then surface action-ready summaries for human approval. See SWIFT standards for context on message complexity: SWIFT standards.

Treat It Like an Assistant, Not an Autopilot

Leaders are clear on the operating model. Ian Glasner of HSBC: "Think of agentic AI as like an intern helping you get all of the more simplistic tasks done, but the human is still there to oversee and take the final decision."

The survey backs this up. 95% say systems can advise and 92% can assist. Only 38% believe current tools can run with full autonomy. That means design for oversight, not full handoff.

The Hard Parts Ops Leaders Must Solve

• Governance (63% cite as biggest challenge): Regulations lag tech. Institutions are building their own guardrails. HSBC maintains inventories tied to business owners, documentation, and risk classes. DBS applies a "PURE" standard: Purposeful, Unsurprising, Respectful, Easy to explain-plus real-time metrics and kill switches for high-risk uses.
• Skills and culture (58%): This is org change, not just tooling. You need education on what agents can do, clear ROI stories for the C-suite, and feedback loops to prove value and improve behavior.
• Data quality and integration (54%): Agents need reliable APIs across hundreds of systems. Role-based access must be strict so agents don't learn or act outside scope. Security protocols should tighten, not loosen, with deployment.

For governance models and controls, the NIST AI Risk Management Framework is a useful reference: NIST AI RMF.

A Practical Playbook for Implementation

• 1) Start simple, then scale: Pick a high-volume process with defined rules and clear outputs. Examples: mortgage underwriting steps, small business loan processing, collections workflows, KYC checks. Automate a slice, prove ROI, then widen the scope.
• 2) Build governance into the platform: Maintain an inventory of AI systems, owners, risk tiers, and documentation. Standardize on common tooling for monitoring, audit trails, and access control. Keep humans in the loop by default.
• 3) Prioritize by measurable value: Buluswar at Citi recommends ranking processes by five factors:
  • How manual it is
  • Cost per transaction
  • Time to complete
  • Error frequency
  • Consequences of mistakes
  Focus where all five score high-mortgage underwriting often does.
• 4) Build platforms, not point solutions: A shared stack lets data scientists and engineers ship faster and safer, avoid one-off tools, and reuse monitoring and controls across use cases.

Metrics That Matter to Operations

• Fraud/security: precision/recall, false positive rates, mean time to detect/respond, identity verification pass rates
• Customer ops: average handle time, first contact resolution, SLA adherence, NPS/CSAT linked to agent-supported journeys
• Process health: straight-through processing rate, exception rate, rework rate, cycle time by step
• Quality and risk: human override rate, kill-switch activations, model drift indicators, audit trail completeness
• Financials: unit cost per process, avoided losses, capacity released (hours redeployed)

Why Early Movers Are Pulling Ahead

• Risk and security: Always-on monitoring and response at a scale teams can't match manually, with configurable detection per customer profile.
• Customer experience: Advisors can combine client context, macro signals, and portfolio data to deliver targeted guidance-less generic, more specific to the moment.
• Operational efficiency: Lower manual review volumes, fewer errors, and learning curves that improve over time. Underwriting timelines drop from weeks to days when you automate the prep and exception handling.

What's Next

Banks plan to deepen agentic AI in fraud (75% priority), security (64%), and customer experience (51%). As Nimish Panchmatia of DBS notes: "AI will apply to every part of the business: front office, middle office, back office."

This is sustained work, not a one-off project. As Panchmatia adds: "If done properly, there's significant value at the end of it. But you have to persevere." The institutions that start simple, scale with discipline, and keep humans in the loop will build advantages that compound.

Upskill Your Teams

If you're building AI-assisted operations, targeted upskilling shortens the learning curve. Explore role-based pathways here: Complete AI Training: Courses by Job.