Meet the bot collector: How AI is rewriting India’s debt recovery playbook

The borrower hadn’t answered the adviser’s call that morning. But he had responded to an automated interactive voice response (IVR), leaving a short message that said his daughter was in the hospital.

So, when Manoj called, he did not open with the overdue amount, and instead said, “Sir, how is your daughter doing?"

He noticed an immediate shift in the tone of the conversation. By the end of the call, the borrower committed to repaying his dues in the coming weeks.

In another recovery call that this writer listened to, an artificial intelligence (AI) agent with a female voice and warm demeanour opens the call in Hindi. “Sir, what seems to be the issue? Has there been a change in income?" it asks the borrower, who had missed an equated monthly instalment (EMI) payment. After a short silence, the borrower admits he has lost his job. As the call progresses, it is evident that the borrower, who began the call with hesitation, has now eased into the conversation after sensing that the recovery agent is trying to help him.

“Your credit score will be impacted. It may cause challenges if you need a personal loan in the future. You can make a partial payment, minimum 30%," the AI agent tells him. The tone changes from reassurance while discussing the job loss to firmness while discussing repayment. By the end of the call, the borrower agrees to make a partial payment.

Both these calls, one mediated by a human fed by real-time prompts and the other handled entirely by voice AI, capture the transformation underway in India’s debt recovery ecosystem. It is a seismic shift in a multimillion-dollar industry. According to a 2024 report by IMARC Group, the country’s debt collection software market was valued at $172.8 million in 2024.

From Excel to AI

For decades, debt recovery in India has been a largely informal and human-driven process. Banks, lenders and financial institutions outsourced the list of defaulting accounts to debt collection agencies or agents. Recovery agents equipped with Excel sheets would then call borrowers one by one, often juggling hundreds of accounts a month.

Sangeeta, an adviser with DPDzero, who had previously worked for a debt recovery agency, told Mint it was a tedious affair. “We got a sheet with 500 contacts, everything was manual," she recalls.

Agents relied on gut feel to decide whether to push harder, offer time or escalate. The process with Excel sheets and tired agents was also challenging to monitor at scale.

DPDzero co-founder and chief executive (CEO) Ananth Shroff said that when he visited third-party agencies handling collections for banks and non-banking financial companies (NBFCs), he found operations driven by pressure and minimal oversight, with outcomes depending on an agent’s temperament more than a structured process.

Today, technology companies such as DPDzero, Credgenics, Gnani, Spocto X and Rezolv are supplying software and AI-based collection tools to banks, NBFCs and digital lenders to replace many manual processes.

Rather than being driven by pressure and relying on luck, calls are timed by algorithm, repayment prompts are delivered in regional languages, and AI-driven agents or human advisers supported by them attempt to steer outcomes with consistency. Across these companies, data collection is going beyond regular conversations into full-fledged data-designed journeys.

Data-defined personas

Rather than gut feel, AI systems are driven by data. They analyse repayment behaviour, contact history, language and demographic data to classify borrowers into behavioural personas.

Shroff says early analysis revealed five distinct borrower personas: supportive, bargaining, circumstantial, escapist and intentional. Each persona is identified through patterns often missed by humans. Data points picked up during conversations, such as hesitation, word choice, broken promises, timing of responses and even silence help construct borrower personas.

“Supportive borrowers may simply be confused about a penalty," Shroff explains. “Circumstantial ones need time. Escapists will sound polite but disappear after promising payment."

Advisers that Mint spoke to say they can now see tone analysis, previous interactions, past promises and likely intent before saying hello, which according to them results in the conversation taking a positive turn.

DPDzero works with a mix of banks and fintech lenders, including RBL Bank, IndusInd Bank, Aditya Birla Capital, KreditBee and PaySense. The company secured $7 million in Series A funding in August after a reported sixfold jump in revenue over 18 months.

Noida-based Credgenics, a full-stack debt collection and resolution platform that counts institutions such as HDFC Bank, ICICI Bank and Aye Finance among its customers, uses a similar framework, starting it earlier in the repayment cycle. Its model predicts the communication mode—WhatsApp, IVR, SMS or a human call—through which a borrower is more likely to respond and adjusts the contact strategy accordingly.

Co-founder and CEO Rishabh Goel says the old model of calling everyone with the same frequency is now obsolete. “A high-risk borrower might need 10 touch points; a low-risk one maybe 1," he says.

The system remembers if a borrower typically pays on Fridays, responds best to morning reminders or only clicks payment links late at night. In the case of a broken promise, the tone of subsequent engagement automatically shifts, while in the case of a fulfilled promise, the borrower is moved back into a soft-touch flow.

For Credgenics, which raised $50 million in 2023, and posted ₹220 crore in revenue and ₹25 crore in profit before tax in 2023-24, AI-led channels have now replaced large parts of the traditional call-centre volume.

Meanwhile, Bengaluru-headquartered conversational AI company Gnani.ai has gone a step further, automating the call itself, thereby turning parts of India’s collection floors into machine-led operations. The company, which powers AI voice agents for more than 150 lenders, has built personas directly into its multilingual bots.

Ganesh Gopalan, co-founder and CEO at Gnani, says the voice bot modulates when a borrower mentions a medical emergency or switches immediately to Tamil, Kannada or Hindi when it detects those languages. These bots can be customized based on a financial institution’s requirements. “Some banks want the voice agent to sound very conservative, some want it to be emotionally warm," he says. “The persona changes based on who is on the other side."

Mimicking human empathy

The usage of AI in the debt recovery industry has found multifold advantages. With all calls now being recorded and analysed, there’s a massive amount of data that recovery companies can work with. It also helps achieve the scale a human call centre couldn’t—an AI voice agent can do 100 calls in 20 languages. However, the most interesting use case AI agents are being hailed for is their ability to empathize with borrowers—a human trait that was apparently missing in human agents.

At DPDzero, for instance, Shroff said the breakthrough came when they realised that most defaults weren’t driven by malice but by confusion, crisis or simple inattention. If someone sounds distressed, the system tells the adviser to slow down. Adviser Sangeeta, cited earlier, says the AI support warns her to slow down when it senses that the borrower is frustrated.

Shroff says borrowers in India are often handicapped by their lack of knowledge around financial systems and credit patterns. Many of India’s new borrowers, especially buy-now-pay-later users and first-time credit takers, fail to understand how the system works.

Some take a loan through an app and then delete the app when they can’t repay, believing that ends the problem. Others have never heard of a CIBIL score. So, advisers, apart from asking for a payment date, also explain credit scores, penalties and repayment options in simple terms and in the borrower’s own language.

“Once they understand the impact, a lot of so-called escapists become negotiators," says Shroff.

The flip side

While AI-driven collections promise precision, an AI system is only as reliable and accurate as the data that powers it. For instance, borrowers who have a fragmented credit history across multiple lenders or informal loans, or have gaps in data, may be misclassified by persona engines built purely on data. And for first-time borrowers or people whose repayment behaviour cannot be classified into the AI built personas, a human review is still critical.

In addition, while AI can detect tone and surface-level cues, for more complex emotions that accompany financial distress or cases where users jump between unrelated issues, human context switching and understanding is needed. In rare instances, if the pitch of a debtor’s voice is outside the AI’s training range, it may not be able to get a good read.

With companies replacing a chunk of the human workforce with AI, there has also been a growing sentiment against bots. Many customers complain of frustration at not being able to speak to human agents. In a situation as critical as debt recovery, any glitches or off-script responses would trigger irritation rather than trust for the end user.

Towards total automation

Gnani’s Gopalan calls what he’s running in his Bengaluru office a 30,000-seat contact centre with no humans. AI voices here complete the entirety of the process from early reminders to complex delinquency discussions.

The Gnani system is trying to achieve real-time adaptability. For instance, if someone claims they have already made a payment, it asks for a transaction ID and if the lender’s IT stack is connected, it is also able to check reconciliation instantly.

Through anonymized tests run by lenders, Gopalan said 85-89% of borrowers believed they were speaking to a human—a surprisingly high acceptance rate that shows how quickly automation is normalizing itself in Indian credit behaviour.

Gnani also offers its voice bot services to companies in the recovery space, such as Credgenics. While automation in entirety is a pipe dream, full-stack recovery providers, including Credgenics and DPDzero, are betting on ‘hybridization’. While a large portion of the debt recovery process, including early reminders, payment-link nudges and routine clarifications, are already being automated, full-stack providers insist that the more judgement-heavy cases still need a human.

Shroff says the goal is not to remove humans but to remove human error. “The machine handles volume and the humans handle nuance," he says.

Credgenics’s Goel also notes that automation now dominates the first 30-60 days of delinquency, where borrowers mostly need reminders, links or explanations. In these use cases, voice bots and personalized AI flows outperform human teams without training overheads or compliance risks. Beyond these buckets, the company’s AI tools support telecallers so that they don’t have to spend time to put context together manually.

Better data, better recovery

Based on the early data shared by founders, AI assistance in debt recovery is also helping improve recoveries across the ecosystem. Shroff said that when the company deployed an education-first AI flow for a student lending portfolio that had been delinquent for over 600 days, monthly repayments jumped eightfold. In mainstream retail portfolios, too, behavioural persona models and real-time prompts helped lenders reduce loss rates by 15-20%, he said.

Goel talks about a similar shift on the operational side at Credgenics, where telecallers who earlier were able to collect around ₹2 lakh a month now close as much as ₹5 lakh with AI-assisted guidance.

Gopalan claims Gnani’s AI agent outperformed both internal and outsourced human teams by about 4% within three months at an NBFC. And that across clients, the company’s systems helped enable over ₹40,000 crore in collections over a six-month period.

With AI tools now handling interactions that were once considered too messy, emotionally charged or linguistically diverse for automation, human intervention is slowly becoming a thing of the past.