Credgenics has launched a voice system that rings overdue borrowers itself instead of handing the list to a call centre. Based in Noida, India, it builds the collections software that Indian banks and non-bank lenders use to chase repayment. The product is called Prix AI, and the company announced it on 1 October; it places the call, holds the conversation, and decides during it whether the account can be settled there or needs a person.
What the agent does with a call is the part the company describes most precisely. It reads the borrower's tone and intent as the conversation runs, closes out routine cases on its own, and passes harder ones to a human collector along with everything said so far.
The handover rules are written into the agent itself. Once a borrower commits to a payment date, the system stops the reminder sequence for that account. Anything that turns into a dispute goes straight to a person rather than being argued by the machine.
Credgenics puts the response delay at under 800 milliseconds, uptime at 99.9% and language coverage at more than 12 Indian languages, with the agent able to switch language partway through a call if the borrower does. Those claims are specific enough to be checked later.
Behind the product sits the company's own record of the work. Credgenics says Prix AI draws on more than 1B collections conversations, across over 100M loan accounts that have passed through its platform. In internal testing, it says, the system matched or beat human collectors on accounts of similar risk.
Rishabh Goel, the co-founder and chief executive, put the pitch in one line. "Enterprises don't need more calls," he said. "They need more resolved accounts." Anand Agrawal, who co-founded the company and runs product and technology, framed it the same way: "The real question is what changes after the call ends."
The Pricing Carries The Argument
Credgenics charges for Prix AI on what gets recovered rather than on how many calls go out or how long they last. That is a real departure for a software vendor, because it moves the risk. A per-minute contract pays the supplier whether or not the borrower settles, while an outcome contract pays only when money arrives, which is the arrangement collection agencies have always worked under and software companies rarely have.
Voice agents are already sold three ways in India. Per-minute deals run roughly ₹5.52 to ₹9 a minute. A seat on a human-assisted platform costs ₹2,000 to ₹6,000 a month plus telephony, and per-outcome deals exist at ₹8 to ₹25 per resolved call.
So outcome pricing itself is not new. What is different is a collections specialist applying it to its own borrower data rather than a general voice vendor applying it to somebody else's.
The arrangement also ties the vendor's revenue to recovery rates, which is exactly the incentive that recovery conduct rules exist to restrain. Credgenics has not published how the agent's behaviour is bounded against that pull, and the design choices it has described, stopping reminders after a commitment and escalating disputes, point the other way.
India Limits When A Borrower Can Be Rung
An automated caller inherits every obligation the lender carries, and in India those are unusually explicit. The Reserve Bank of India, the country's banking regulator, told lenders in August 2022 that recovery agents may call only between 8am and 7pm. The same instruction barred intimidation or harassment of any kind, verbal or physical, along with threatening or anonymous calls and any false or misleading representation.
Those rules cover commercial banks, regional rural banks, cooperative banks, non-banking finance companies and asset reconstruction companies, which between them are most of Credgenics' customer base. The regulator placed the responsibility on the lender, not the agent, so a bank cannot outsource the liability along with the calling.
The fair practices code adds more, limiting attempts to two or three a day per borrower and requiring the caller to identify itself. An automated agent has to do that in the same way a person would.
The rules start before a call is even dialled: the telecom regulator separates transactional calls, such as instalment reminders, from promotional ones, and routes the two through different registered number ranges. The Digital Personal Data Protection Act of 2023 then requires consent tied to the purpose of each call, an audit trail, and the borrower's data to stay inside India.
None of that blocks what Credgenics has built, but it does mean the compliance layer is as much of the product as the conversation. The company has not said how attempt caps and caller identification are enforced inside the agent.
Speed Is The Harder Engineering Claim
Sub-800 millisecond response is the number worth watching, because it is well ahead of what the market generally reports. Indian voice deployments commonly answer in under two seconds, and the gap matters more on a phone call than it sounds. A pause of a second and a half tells the borrower they are talking to a machine, and the conversation changes the moment they know.
Recognition accuracy is the other half, and it is harder in India than almost anywhere. Telephone audio is compressed to 8 kilohertz, borrowers switch between Hindi and English inside a sentence, and background noise in smaller towns is constant.
Published benchmarks for the market put word accuracy on Hindi telephone audio at 92% to 96%, Hinglish code-switching at around a 6% error rate, noisier conditions nearer 7%, and Tamil, Telugu and Kannada between 88% and 93%. Credgenics has not published its own figures against those, and the claim of matching human collectors comes from testing the company ran itself.
The escalation design is where the accuracy question actually bites. An agent that mishears a dispute as a payment promise will switch off the reminders on an account that was never going to pay. That is the same question of where an agent's authority ends that runs through every autonomous deployment.
The Data Is Meant To Be The Moat
Credgenics is not competing on the voice technology itself, and it does not claim to be. Gnani, Ozonetel, Exotel, Knowlarity and Bolna all sell Indian enterprises voice agents, while Sarvam AI and ElevenLabs supply the speech models underneath several of them. Any of those can place a call in Hindi.
What Credgenics argues it has instead is the record of what works, drawn from the conversations already run through its platform. That is the familiar claim that owning the data is the advantage once the underlying models are available to everyone.
The figure supporting it needs reading carefully. 1B conversations across 100M loan accounts is the company's own count, and it is published without a definition. Nothing says what counts as a conversation, how many were calls rather than messages, or over what period they accumulated.
Goel, Agrawal and Mayank Khera founded Credgenics in 2018, and it raised $50M in a Series B led by WestBridge Capital in 2023 at a $340M valuation. ICICI Bank, HDFC Bank, Mahindra Finance, IIFL Finance and Hero Fincorp are among its customers.
Its revenue reached ₹100 crore, about $12.1M, in the 2023 financial year, and the platform handled a loan book it put at $47B. A voice product priced on recovery rather than usage is an attempt to earn a share of that book instead of a licence fee against it. The first evidence of whether that works will come from lenders reporting recovery rates, rather than from the launch.
