The Fifteen-Minute Risk Officer

A credit application is a small work of autobiography. A business states what it does, what it earns and what it needs. Then someone must decide whether the story hangs together. At a large bank, that decision can wander through queues, committees and systems that remember the last century rather too fondly. At Coast, Anurag Puranik has tried to make it move in minutes.

Puranik is the Chief Risk Officer and a founding executive at the New York fintech, which provides expense cards and software to businesses with people and vehicles in the field. Roofers, plumbers, landscapers and fleet operators do not spend their days composing immaculate loan files. They buy fuel, pay tolls, replace parts and get back to the job. Their financial lives are legible, but they are rarely tidy.

This is precisely where Puranik has made his career: at the point where messy evidence becomes a consequential yes or no. He studied mechanical engineering at Visvesvaraya National Institute of Technology, then industrial engineering at Purdue. The route from machines to money is less eccentric than it first appears. Both disciplines ask how a system behaves under pressure, which parts carry load and where failure is likely to begin.

At Purdue's PRISM Center, an early colleague watched him translate requirements into models and remembered a habit of persisting when the conditions became difficult. Another recommendation from his modeling years praised his appetite for a steep learning curve and his quick turnaround on demanding work. Such testimonials are ordinarily the wallpaper of a professional profile: pleasant, beige and quickly passed. Here they reveal a useful continuity. Puranik has kept choosing rooms in which the answer is hidden inside imperfect data and the clock is unhelpfully loud.

He also earned the Chartered Financial Analyst charter in 2017, midway through his KeyBank tenure. The credential added the language of capital and portfolios to an education already concerned with optimization. By the time Coast came calling, he could see a lending system from several altitudes: the individual model, the operating queue, the portfolio and the profit-and-loss statement. Startup mythology often favors the dazzling generalist who arrives innocent of convention. Puranik brought the opposite advantage. He knew why the conventions existed, which made him better placed to decide which ones could safely be made faster.

A banker's patience, without the waiting

At Discover, PNC and KeyBank, Puranik accumulated more than a decade in consumer and commercial lending. At KeyBank he rose to Senior Vice President and Head of Credit Analytics, overseeing work tied to a $34 billion portfolio. The scale taught him what controls are for. It also taught him how slowly a control can change once it is wrapped in legacy systems and institutional caution.

His favorite contrast is almost comic. A policy adjustment that might have required a four-week review at a bank could be completed at Coast in 15 minutes. The point is not that a quarter-hour is morally superior to a month. It is that a young company can test, observe and adjust before a brittle rule becomes a permanent inconvenience. Good risk management is not a marble monument. It is a nervous system.

At Coast, he helped build more than a credit function. His remit has crossed finance, fraud, analytics and operations. That combination matters because the risk ledger and the company ledger eventually become the same book. One bad month of credit or fraud performance can turn an otherwise healthy profit-and-loss statement inside out. A risk chief who cannot see the economics is merely admiring the locks.

The operational gains are concrete. Coast set a goal of deciding applications in five minutes or less. An onboarding system automated most of the process, while difficult cases still reached an analyst. Puranik said that without this infrastructure, six people would be needed to process the volume handled by one. Later document-review automation pushed the routine work further down: what had taken an hour or more often required no manual work, while flagged cases could be reviewed in five to ten minutes.

When looking legitimate became cheap

Then the adversary acquired a copywriter, a web designer and a patient rehearsal partner, all inside the same text box. Generative AI has changed the texture of fraud. Puranik has described verification calls that suddenly sounded polished and knowledgeable. His team tested the obvious suspicion. Ask a chatbot to produce a script for the invented owner of a particular kind of business, and seconds later the caller can speak the dialect of legitimacy.

Websites changed too. The flimsy placeholder page, once a useful warning, can now arrive dressed in plausible copy and real-looking images. Bank statements can be templated with the kinds of transactions an underwriter expects to see. Coast once treated a payment to fleet software such as Fleetio as one hint that an applicant really operated vehicles. Once fraudsters learned the pattern, the hint became theatre.

The response is to stop trusting the shine. Puranik's team looks further down: when a domain was registered, what technology sits behind a site, how a document was produced, whether transaction behavior and stated intent agree. No single clue deserves the crown. The picture emerges from several signals that are expensive to coordinate and difficult to counterfeit together.

He frames the contest in economic terms. Fraudsters, too, have operating costs. They spend time, acquire tools and search for easy doors. A defender does not need to pursue every bad actor across the internet. The more practical ambition is to make an attack slow, troublesome and expensive enough that its expected return collapses. In his metaphor, the unlocked door invites a visitor; the heavy door with the complicated lock sends that visitor elsewhere.

The human at the end of the machine

There is a neat irony in Puranik's current project. After years spent removing manual labor from underwriting, he is adamant about preserving the human decision-maker. His distinction is between analysis and judgment. Software can gather evidence, parse transactions, compare documents, surface inconsistencies and execute a standard operating procedure. When a decision will harm a customer, trigger an adverse action or carry regulatory weight, a person should own it.

That is the banker's inheritance, carried into a company that prefers its clocks fast. For banks and credit unions experimenting with agentic AI, his advice is deliberately untheatrical: begin with a narrow use case, prove that it works, keep people in the loop and expand from there. Use agents to connect old systems before embarking on a heroic attempt to replace everything at once. Heroic replacements have a habit of becoming expensive museums.

At Coast, the stated 2026 plan has been to perfect three agentic use cases with governance built in. One would lower the cost of underwriting smaller businesses that are uneconomical to assess by hand. Another would assemble a 360-degree view across documents, websites, application intent, acquisition traffic and bank behavior. A third would automate repeatable downstream work in disputes and transaction monitoring. In every case, the machine prepares the room; the human conducts the difficult conversation.

Colleagues from earlier chapters describe Puranik as solution-driven, quick to learn and determined when a problem resists him. Those are useful qualities in a field where yesterday's signal can become today's disguise. They also help explain the arc from mechanical engineering to credit analytics. He is still examining systems under strain. The components now happen to include people, incentives and lies.

What the saved hour is for

It is tempting to finish with the impressive fractions: weeks reduced to minutes, six processors' work handled by one, most applications moving without manual review. Yet efficiency is the least interesting destination. The real question is what happens to the saved hour.

In Puranik's design, the saved hour returns to judgment. An analyst can spend less time rearranging statements in Excel and more time asking whether the evidence makes sense. A legitimate business receives an answer while it still needs one. A doubtful application reaches a person with its contradictions already marked. The machinery becomes less visible precisely so responsibility can become clearer.

That is an unfashionably modest ambition for AI, which is often introduced wearing a cape. Do the repetitive work. Find the hidden relationship. Bring the relevant facts forward. Then stop at the point where judgment has consequences. Puranik's fifteen-minute world is not a place without caution. It is a place where caution has learned to keep up.