Lending has traditionally followed a sequential process: sense, analyse, decide and act. Each step depends heavily on human intervention, with analysts collecting information, interpreting signals, conducting checks and moving files through the next stage.
The emergence of agentic AI is beginning to challenge that model by connecting these steps into a more continuous workflow. AI agents can pull data, reconcile information, initiate checks, identify exceptions and recommend the next action. However, according to Sushantam Mohan, Chief Data Officer, Vivriti Capital, the transformation is unlikely to begin with autonomous credit decisions.
Instead, adoption will move from the bottom of the value chain upwards, starting with workflows where the information is verifiable, volumes are high and the tolerance for errors is relatively greater.
“The first genuinely agentic workflows will be the high-volume, verifiable, high-error-tolerance ones: document collection and chasing, KYC/KYB, bank statement and GST reconciliation, tracking conditions precedent and subsequent, security perfection, covenant monitoring, and portfolio surveillance,” says Mohan.
For lenders, the approach provides a practical path towards agentic adoption. Rather than attempting to automate the most consequential decisions first, organisations can establish reliability and instrumentation in repetitive processes before gradually moving towards more complex workflows.
“Prove reliability and instrumentation there, then move up. Capture the lowest-hanging fruit before climbing up the tree,” he says.
Moving beyond the human handoff
The opportunity, Mohan argues, is not simply about making individual tasks faster. It is about removing the latency created when humans become the connection between different stages of a lending workflow.
“Today the human is the connection between sense, analyse, decide, and act. That is where the latency sits — work is batched, context is lost at handoffs, and the process only moves when someone picks up the file.”
An agent that can hold the context across these stages can make lending processes more continuous rather than episodic. It can monitor information without being constrained by working hours or the fatigue associated with repetitive tasks.
“That matters more than raw speed: monitoring quality degrades in humans precisely because it is repetitive, and it does not degrade in an agent,” Mohan says.
This shift becomes particularly significant as lenders look beyond point-in-time underwriting toward a more dynamic understanding of borrowers.
From a credit snapshot to a living credit file
The fundamentals of credit assessment, however, remain unchanged.
“The five Cs of credit are sacrosanct,” says Mohan. “You still need to understand who you are lending to, whether they can repay, what happens if they cannot, and whether you are being paid for the risk.”
What can change is the depth, speed and frequency with which these questions are answered.
Lenders already have access to significant volumes of financial and operational data. The challenge is often that information remains fragmented across systems, creating multiple unreconciled versions of the same borrower.
A human analyst also faces practical limits in bringing together large volumes of statements, filings and charge records within the time available for a credit decision.
Agentic AI could enable what Mohan describes as a “living credit file” — one that continuously updates as the borrower’s financial and operating environment changes.
“The change on offer is a living credit file: one that updates as bank flows, GST filings, bureau movements, buyer concentration and sector conditions move, rather than a file assembled once and then ageing quietly.”
This becomes particularly relevant for mid-market enterprises, where business performance can be uneven and the trajectory of the business can provide more insight than a single snapshot.
AI handles evidence; humans retain the decision
Despite the potential for greater autonomy, Mohan argues that lending should not become a domain where AI simply replaces human judgement.
The boundary should be explicit.
“The exclusions should be explicit and drawn on three lines: complexity, ticket size, and edge cases,” he says.
Final sanctions beyond delegated thresholds, exceptions, covenant waivers, restructuring and workout strategy, first-of-kind structures without comparable histories, promoter and management assessments, and adverse actions requiring a reasoned explanation should remain with experienced credit professionals.
The architectural principle, according to Mohan, is a separation between proposal and approval.
“Agents own evidence assembly, verification, reconciliation, scenario construction and a recommendation with both supporting and contradicting evidence. Humans own the decision.”
This also means that delegated authority cannot simply remain an informal process.
“The delegated authority matrix should be encoded in the system, deny-by-default, not left to convention.”
Mohan also argues for routing cases based on confidence and novelty so that unusual cases are surfaced rather than averaged into an automated flow. Agents should be required to challenge their own recommendations, while lenders should maintain records of both human overrides and instances where humans accept agent recommendations.
“Log overrides in both directions — a credit officer overruling the agent is as informative as the reverse, and over time that log is how you learn where the boundary should sit.”
The principle of accountability remains equally important. “Accountability stays named: a person, not a committee, owns each sanction.”
Continuous credit intelligence
Agentic AI could also change how lenders monitor portfolios after disbursement.
Instead of relying primarily on periodic reporting, lenders could move towards continuous, event-driven monitoring, where agents respond to changes as they happen.
“Data becomes continuous, not periodic.” Event-based to avoid heavy compute,” Mohan points out.
The rationale is straightforward: delays can diminish the value of information.
An agent can monitor a portfolio continuously and identify unusual behaviour after an event, rather than waiting for a periodic review.
For this model to work, however, the underlying data architecture needs to change. Data needs to flow as events occur rather than through monthly batches. Borrower-group entities need to be correctly linked so that total exposure is visible. Ratings and limits need to be capable of continuous updates rather than annual reviews.
Alerts also need to be carefully designed. If lenders are overwhelmed by false positives, they risk becoming desensitised to warnings. “Alerts must be tuned properly — if the system cries wolf, people stop listening. And every alert needs a defined action attached to it.”
Human oversight remains essential where conclusions are ambiguous or the value at risk is high.
Governance begins with observability
As agents begin taking actions rather than merely generating recommendations, AI governance also takes on a different dimension.
For Mohan, observability has to be built into the architecture from the beginning. “Observability has to be built in from day one, not added later. Governance then sits on top of it.”
Every agent action should leave a detailed and immutable record covering the data accessed, tools called, steps taken and model version used. This creates the ability to reconstruct an agent’s behaviour after the event.
The permissions architecture also needs to be tightly controlled. Agents should access only the data required for their specific purpose, while hard limits around exposure, tenor and sector should sit outside the model itself.
This ensures that an agent cannot exceed defined boundaries even if its underlying model produces an unexpected recommendation.
Observability also enables lenders to identify drift, escalate unusual cases and shut down problematic behaviour quickly.
Importantly, Mohan believes every agent should have a named owner. “Owners are like managers who track progress, assess performance, provide feedback and ‘coach/mentor’ agents so that their performance improves over a period of time.”
From cost reduction to better credit decisions
For Mohan, the long-term value of agentic AI in lending extends beyond reducing operational costs. “Cost savings cap out quickly. You can only cut so much, and a firm chasing only that will stall in a couple of years,” he adds.
The bigger opportunity is to improve the consistency and quality of credit decisions by giving every credit professional access to the depth of information and analysis typically associated with the strongest performers.
“The biggest bang for the buck is when these systems can make an average credit officer perform like your best one.”
This could have implications not only for productivity but also for the breadth of borrowers lenders can serve.
For mid-market companies, one-size-fits-all products can emerge because the cost of assessing a business properly is high relative to the ticket size. If AI can reduce both operational and credit assessment costs without compromising decision quality, lenders may be able to serve more of these businesses with greater precision.
Looking five years ahead, Mohan describes the objective as an institution capable of significant growth without a corresponding increase in headcount or deterioration in decision quality. “Five years out, an AI-native lender should be able to grow the book ten times without adding people in the same proportion and without decisions getting worse. That is the goal.”
The foundations for that future, however, need to be built now. Lenders need to clean up data and link entities correctly, establish consent infrastructure for ongoing data use, structure credit assessment reports in machine-readable formats, train agents and models, build observability into systems from the outset, and log decisions and overrides.
“Real business outcomes beyond mere cost savings lead to long-term sustainable value creation,” Mohan says.
For lending, the transition to agentic AI is therefore less about handing over decisions to machines and more about redesigning the operating architecture around continuous intelligence. The near-term opportunity lies in automating repetitive, verifiable workflows; the longer-term opportunity is to create a lending institution where information flows continuously, risks surface earlier, and human credit expertise is applied where judgement matters most.