EASE 9.0 · Risk & Resilience · Credit Risk & Resilience

Old Banking or New Banking

Real-time credit decisioning at machine speed.
How AI agents enable simultaneous decisioning at scale while maintaining governance and RBI compliance.

PSU banks process millions of credit applications annually. Traditional decisioning takes days, often weeks. A customer applies for a loan on Monday. A credit officer reviews it Wednesday. A committee approves or denies it Friday. The application may be excellent, but by Friday the customer has moved on. A fintech has already said yes.

Meanwhile, the competitive landscape has shifted. A fintech approves loans in 24 hours. Some competitors approve in minutes. Speed has become the baseline for customer acquisition. PSU banks that cannot match this speed lose market share. They lose the customers they most want: young, digitally savvy, credit-worthy.

But speed creates fear. Fast decisions are careless decisions. Banks that approve quickly are banks that approve recklessly. Right? This is the assumption that has paralyzed many PSU banks. They believe they must choose: approve slowly and safely, or approve quickly and take bad credit.

The choice is no longer speed vs. safety. It is old banking or new banking.

The Path to Progress

Real-time credit decisioning happens as a sequence of five moves, not a single system swap, each one a genuine prerequisite for the next.

The Business Case

Notice what stays constant in that comparison: total monthly spend does not change. What changes is what that spend buys — 10x the approval volume, at a 90% lower cost per approval, with a lower NPL rate rather than a higher one. Agents also reduce human bias. A credit officer might unconsciously favor certain customer profiles; an agent applies consistent logic to all customers.

The illustrative scenario above is directionally consistent with a real deployment, though the two figures below measure different things and should not be read as one derived from the other: at one PSU bank, campaign-triggered credit offers moved from a 3–5 day response cycle to same-day delivery for the majority of pre-qualified cases, while average turnaround time across the full portfolio, including cases that still require escalation, fell 40%. That number reflects how fast a qualified offer reaches the customer once the bank's own eligibility rules clear it, not a claim about autonomous underwriting authority; the shadow-scoring and cutover sequence below is what actually governs when, and whether, an agent earns sanction authority.

Source (deployment figures): Appice client engagement data, EASE 9.0 & AI client briefing, 2026. Indicative of results reported across these campaign-management engagements, not independently audited; bank name withheld at client request.

Where AI Agents Fit

The deployment cited above is a campaign-management engagement, not an instance of the Credit Decisioning Agent described below; this section uses this guide's shared agent vocabulary to describe where the category is headed, not a claim about that specific deployment or about what Appice itself has built. A model that returns a probability of default is not new — Indian banks have run those since the mid-2010s, and a credit head who has sat through a decade of scorecard vendor pitches is right to be unimpressed by "ML credit models with rules-based policy checks" on its own. What has actually changed is what happens around the score. In August 2026, DBS Bank rolled out agentic AI tools to roughly 1,500 relationship and credit-risk managers worldwide, after a 150-person pilot: the agents pull annual reports, industry research and internal records and assemble a review-ready credit memo across more than 70 distinct sub-tasks, work that had been eating up to 40% of a relationship manager's time. The banker still owns the decision; the agent owns the assembly, and DBS is targeting at least a 30% cut in preparation time as a result. Domestic platforms are doing the equivalent for Indian lending: Perfios's CAM AI reconciles bank statements, GST filings and ITRs to cut underwriting turnaround by as much as 85%, and Scienaptic AI's platform treats underwriting as continuous rather than a one-time event. The same agent that approved a loan at origination keeps watching the live portfolio afterward, flagging early-warning signals instead of waiting for the next annual review cycle.

A Credit Decisioning Agent built on that same premise does not merely return a score. It assembles the case file, applies compliance checks inside the decision rather than after it, and keeps re-scoring the loan after disbursement. A Reasoning Agent runs alongside it, generating the explainable, auditable trace an examiner or the bank's own risk team can pull for any decision, at any point in the loan's life, not only at sanction. That is also what a credit committee now has to govern: not merely a scorecard's PSI and KS statistics, but which agent has permission to touch which stage of the file, and what triggers an automatic escalation once the loan is live.

Sources: DBS Bank newsroom, “DBS scales agentic AI to transform way of working for corporate bankers” (Aug 2026); Perfios press release, “Perfios launches CAM AI” (2026); Scienaptic AI quarterly credit-volume disclosure (2026). Third-party figures as reported by each company, not independently audited.

The Skeptic's Answer

A credit head who has already lived through one black-box model that looked fine in back-testing will hear that a campaign-triggered offer moved from a 3–5 day response cycle to same-day delivery and translate it immediately: faster decisions running at a scale the system has not yet proven across a full credit cycle. That skepticism is earned, not paranoid — and worth answering directly: that number describes campaign speed, not autonomous underwriting authority.

Sequencing answers that, not a better back-test: shadow-scoring the agent against live human decisions before it ever holds sanction authority, cutting over one segment at a time rather than the whole book, and re-scoring every loan continuously after disbursement rather than trusting the score frozen at origination. A credit committee's real job isn't judging whether the agent is accurate on day one; it's judging whether it stays accountable on day four hundred. That is what the escalation ladder and the Reasoning Agent's audit trail are actually for.

Why Sovereignty Matters

Many banks consider outsourcing credit decisioning to a vendor who hosts the engine and returns decisions — operationally simple, strategically risky. Appice's decisioning engine runs on the bank's own infrastructure. The bank owns the data, defines policy, and controls the agents. When RBI introduces a new requirement, the bank updates the agent within hours, not a vendor's next release cycle.