EASE 9.0 · Innovation · Artificial IntelligenceA Model Scores. An Agent Decides.
A model scores. An AI agent decides, acts, and explains itself.
That distinction is the whole of the Innovation pillar.
A traditional machine-learning model scores an application and hands the score to a human who still has to decide. An AI agent decides end to end, enforces compliance as it goes, and explains its own reasoning autonomously, running on the bank's own core AI stack rather than a third-party black box.
Consider fraud detection. A model flags a transaction and queues it for a human analyst who reviews it hours later. An agent evaluates the same transaction against policy in real time: it holds the transaction, verifies the customer, and either releases or blocks it, all inside the transaction window and logged end to end.
Model Versus Agent
A model is a function: input in, score out. An agent perceives, decides against policy, acts, and learns from the outcome.
This Is Already Live, Not a Roadmap Slide
The distinction above is not theoretical, and a PSU bank does not have to look abroad to see it running. RBI's own Reserve Bank Innovation Hub built MuleHunter.AI precisely because a scoring model was not enough: trained on nineteen distinct patterns of mule-account behaviour, it does not flag one suspicious transaction for a human to review later, it perceives the pattern across a whole network of accounts and acts inside the window that matters, before stolen funds move on. It was piloted with two public-sector banks before wider rollout — detailed in full on the Cybersecurity & Fraud Resilience page. On the lending side, Perfios's CAM AI and Scienaptic AI's underwriting platform run the identical model-versus-agent distinction against Indian bank data specifically, reconciling bank statements, GST filings and ITRs and treating underwriting as continuous rather than a once-a-year event, covered in more depth on the Credit Risk & Resilience page. None of this is a global bank's roadmap slide; it is running inside the Indian banking system today, including at public-sector banks.
Sources: IndiaAI/Reserve Bank Innovation Hub coverage of MuleHunter.AI; Perfios press release, “Perfios launches CAM AI” (2026); Scienaptic AI quarterly credit-volume disclosure (2026).
The same pattern holds outside India, with the same caveats worth reading rather than skipping. Commonwealth Bank of Australia's fraud team built an agentic system that has contributed to developing or updating three-quarters of the bank's card-fraud rules; JPMorgan Chase reports roughly $2 billion a year in AI-driven savings across internal tools reaching 250,000 employees; DBS Bank rolled out agentic credit-assessment tools to roughly 1,500 relationship and credit-risk managers in August 2026, the same deployment detailed on the Credit Risk & Resilience page. What is more useful than the scale, though, is what these institutions say about its limits. DBS has said publicly that it is holding off on letting its AI agents run fully on their own because governance controls have not yet caught up with what the technology can do. Klarna is the cautionary data point on the other side: its AI agent handled 2.3 million chats a month in 2024, doing roughly the work of 700 support staff, but by mid-2025 the company had rehired human agents after service quality on complex cases fell short. A bank running MuleHunter.AI or CAM AI today is not behind that curve; a bank that skips the escalation ladder these examples argue for would be repeating Klarna's mistake, not DBS's caution.
Sources: CommBank newsroom, “CommBank develops AI agent that spots new fraud and helps build defences” (Apr 2026); CNBC, JPMorgan Chase AI coverage (Sep 2025); DBS Bank newsroom (Aug 2026); ComputerWeekly, “DBS holds off on letting AI agents run on their own as controls lag capability”; industry reporting on Klarna's 2025 AI-support reversal. Figures as reported by each company, not independently audited.
The Path to Progress
This category is different from the other fifteen: it is not a destination but the adoption sequence a bank runs, once, to make every other category possible.
- Phase 1: Bound the decision. Pick one well-understood, high-volume decision to automate first, not "AI everywhere," but one decision, cleanly scoped.
- Phase 2: Encode policy and the escalation ladder. Compliance and engineering co-author the rules and the five-rung ladder together, before a single production decision runs.
- Phase 3: Pilot with full audit sampling. Every escalation is reviewed by a human; every auto-decision is sampled. Trust is earned from a visible track record, not declared by mandate.
- Phase 4: Expand to adjacent decisions. Credit decisioning's infrastructure becomes fraud detection's infrastructure becomes collections' infrastructure, reused rather than rebuilt.
- Phase 5: Institutionalize governance. Named owners, continuous threshold tuning, and audit cadence replace an annual policy reset.
The Business Case
McKinsey's research on early bank deployments backs this up. The finding cuts both ways: real upside for adopters, and a real penalty for banks that wait.
The same research warns that most of those cost gains get competed away over time, the way digital banking's gains eventually were, and that banks which fail to adapt at all risk their share of a global banking profit pool McKinsey estimates could shrink by up to 10% over the next five to ten years. The gains compound because the same infrastructure that powers one agent (real-time data, policy engines, audit logging) is reusable across every one of the sixteen EASE 9.0 categories that follows.
Where AI Agents Fit
Role: Foundational Enabler
Every other page in this guide describes agents doing something specific — deciding credit, catching fraud, explaining a decline. This page is where that capability itself gets defined and governed. Take the Credit Decisioning Agent as the worked example, since it is the one most other categories borrow from. It perceives a live application alongside real-time bureau, Account Aggregator, and transaction data as it arrives, not a batch snapshot from the night before. It decides by weighing that data against ML credit models and rules-based policy the bank's own compliance team authored, including where the decision must escalate rather than proceed automatically. It acts by approving, declining, or routing the case to a human, with the reasoning attached rather than delivered separately. And it learns by feeding the outcome, whether the loan performed as scored, back into the threshold-tuning cycle, so next month's escalation ladder reflects this month's real results instead of an annual policy reset. Every named agent elsewhere in this guide (Fraud Detection, Compliance, Decisioning, Reasoning) runs the same four-part loop against different inputs and different policy. Get the escalation ladder and audit discipline wrong here, on this first agent, and every category downstream inherits the same governance gap.
Five Challenges to Making the Shift
- Trust & explainability. An agent must produce reasoning a bank examiner can follow, not only a data scientist.
- Legacy integration. Real-time agents need live hooks into systems built for nightly batch.
- Policy encoding. Turning written credit and compliance policy into executable rules requires compliance and engineering to co-author the logic together.
- Frontline trust. Credit officers and relationship managers must trust agent recommendations enough to stop reflexively overriding them.
- Governance capacity. Someone has to own ongoing agent monitoring, audit sampling, and threshold retuning.
The Skeptic's Answer
DBS is a top-five global bank on AI maturity and years into this work, yet it is still holding off on full agent autonomy; Klarna had to rehire humans after its AI-only support model broke down on complex cases. A PSU bank with a fraction of that maturity could reasonably ask why it should push further than either of them has.
That reading gets the lesson backwards. DBS and Klarna are not arguments against running agents; they are arguments against skipping the escalation ladder and the audit-sampling discipline this page's five-phase path is built around. Klarna's failure was letting an agent operate at full autonomy on complex cases it was never bounded to handle. DBS's caution is the same lesson, learned in advance rather than the hard way. A bank that bounds the decision, pilots with full audit sampling, and expands only on a visible track record is following DBS's own playbook — not skipping the step DBS is still taking.
Agents on Infrastructure You Control
The gap between “we use AI” and “we run AI agents” is the gap EASE 9.0's Innovation pillar is actually measuring. Appice's agents run on infrastructure the bank licenses and owns directly (on-premise, private cloud, or hybrid), bound to policy the bank defines, logged for audit, and open to override at the point a reviewer disagrees.