AI Voice Agents for Loan Collections in India: What to Automate, Where Humans Still Win, and How to Prove It (2026)
The short answer
AI voice agents can carry a real share of loan collections in India, but only where the evidence says they hold up: pre-due reminders, early and small-balance cases, and fast handover to skilled humans. A cheap AI that simply replaces collectors tends to collect less and leave borrowers less willing to pay next month. The strongest controlled study so far, run on about 22 million cases at a Chinese online lender, found AI callers trailed humans by up to 9 percentage points of the overdue balance and that promises made to AI were kept less often. [SRC001] [SRC002]
That study used pre-generative AI, so modern Hinglish voice agents can do better. But the lesson for an Indian lender or collections BPO holds: buy for human-level persuasion and prove it with a randomized pilot measured on kept promises and rupees recovered, not on cost per minute. This guide covers what to automate, what to keep human, the RBI and TRAI rules, and the KPIs. It is not legal advice. [SRC001] [SRC003] [SRC008]
What the best evidence actually found
Choi, Huang, Yang and Zhang (NBER Working Paper 33669, April 2025) used two experiments inside one lender. Debts of 300 yuan or less went to AI callers permanently, which allowed a regression discontinuity. Separately, each month a random 10% of larger debts was split: half called by AI through day 5 and humans after, half called by humans throughout. The paper is a working paper and not yet peer reviewed. [SRC001]
| Finding | Result | What it means for you |
|---|---|---|
| AI permanently on small debts | Gap peaks at about 9 points of the overdue balance near one month past due, still about 5 points at one year | Full replacement costs real recovery, even on small tickets |
| Credit quality | Gap larger for borrowers with lower credit scores | Riskier segments need more human skill, not less |
| AI for 5 days, then humans | Gap of 12 points by day 5, down to 2 points at day 10 and 0.8 points at day 30 | Fast handover recovers most of the loss |
| Lasting damage | The 0.8 point gap barely closed over 11 more months; AI-contacted borrowers were 1 to 2 points more likely to be late on each of the next 12 payments | Track next-cycle behaviour, not just this month's cure |
| First answered call | AI calls 31 seconds shorter, 21 points fewer promises to pay, 18 points fewer same-day payments; promises to AI kept less often | The weak point is commitment, not reach |
| Cost-effectiveness | Net of direct labour cost, AI matched humans only on smaller debts | Cheap per minute is not cheap per rupee recovered |
Two cautions. The AI was not a modern large-model agent, and the setting was China, not India. Use it as the baseline your vendor must beat, not as proof that AI cannot collect. [SRC002] [SRC001]
The disclosure problem every Indian lender has to design around
The same paper cites earlier work (Luo et al., 2019) in which an AI caller's success on a loan renewal offer fell by 80% when it disclosed up front that it was an AI. In India you cannot solve that by hiding the AI. RBI's FREE-AI committee report says consumers must be explicitly informed whenever they interact with AI systems, should always have the option to switch to a human, and that AI voice interactions should run only on verified 1601 series numbers. TRAI has separately moved BFSI transactional and service calls to the 1600 series. [SRC001] [SRC003] [SRC007]
So the design question is not "can the borrower tell" but "is the call worth staying on." That is a quality problem: natural Hinglish, low latency, accurate dues, real answers to questions, and a human one sentence away. It is exactly where a cheap script-reader loses and a premium agent earns its price. [SRC012] [SRC015]
What to automate, and what to keep human
| Stage | AI role | Human role | Why |
|---|---|---|---|
| Pre-due reminder | Lead | Exceptions only | Reminding and informing is the part the study says AI does nearly as well as humans |
| Bounce or day 0 to 5, small balance | Lead, with payment link on the call | Callback on any broken promise | Same-day payment beats a promise; fast handover closed most of the gap |
| Day 6 onward, larger balances | Assist: reach, verify, schedule | Lead the negotiation | The gap grew with days past due before handover |
| Low credit score segments | Assist | Lead | The AI gap was larger here |
| Hardship, disputes, distress, complaints | Detect and transfer at once | Own the case | RBI distress and grievance expectations; FREE-AI route to a human |
Indian vendors already market several of these controls: disclosure at the start, a right-party check before discussing dues, a WhatsApp payment link, no OTP or card details on the call, and disputes and hardship routed to the team. Ask to hear them on real calls, not slides. [SRC010] [SRC001] [SRC003]
Make promises feel binding
The study's sharpest result is that borrowers break promises made to AI more often. Five habits push against that: [SRC001] [SRC008]
- Ask for payment now. Send the UPI link during the call and stay on while it is used. Same-day payment was where AI lost 18 points. [SRC001] [SRC010]
- Make the promise specific. Amount, date and channel said back by the borrower, then confirmed on WhatsApp.
- Escalate a broken promise to a human. Do not let the AI loop on the same borrower.
- Hand over fast. The randomized arm recovered most of the gap once humans took over after day 5. [SRC001]
- Score the call, not just the outcome. Tone, accuracy of dues and every required statement, on every call. [SRC013]
The rules your AI collections calls must pass in 2027
RBI's 2026 Amendment Directions on recovery take effect on 1 January 2027 across banks, NBFCs, HFCs and others. The NBFC version requires logging the time and number of recovery calls, recording outbound and inbound calls with six-month retention, and calling only between 08:00 and 19:00. The 2022 circular already barred intimidation, persistent calling and misleading statements, and kept the lender responsible for its agents. An AI caller is still your agent. [SRC004] [SRC005] [SRC006]
Our RBI 2027 recovery rules checklist turns these into call QA checks, and BFSI call QA in India covers the wider rubric. [SRC014]
KPIs that prove it works
| KPI | Definition | Why it matters |
|---|---|---|
| Right party contact rate | Answered calls where the verified borrower was reached | Reach without verification is noise |
| Promise to pay rate | Right party contacts ending in a dated promise | Easy to inflate on its own |
| Kept promise rate | Promises paid in full by the promised date | The number the study says AI loses on |
| Same-day payment rate | Payments within 24 hours of the call | Commitment without a promise to break |
| Recovered share of overdue | Collected amount over overdue balance at 1, 3, 7 and 30 days | Comparable across vendors and buckets |
| Next-cycle on-time rate | Share paying the next EMI on time | Catches the lasting damage found in the study |
| Compliance and fatal-error rate | Calls breaking a required rule, on 100% of calls | One bad call can cost more than a month of savings |
Compare against a randomized control group split by days past due, balance, product and prior behaviour. Without one, seasonality and selection will flatter any vendor. [SRC008] [SRC001]
Vendor numbers vary and are not comparable. Skit.ai reports 50 to 55% right party contact, 8 to 10% promise to pay and 20 to 25% containment for one inbound US agency deployment. Treat any figure like that as a hypothesis to test on your own book. [SRC009]
Why the cheapest collections AI is the expensive one
Indian collections AI is marketed from as low as Rs 3 a minute in some languages. That price is real, but the study's math points the other way: a few points of the overdue balance lost on a lending book dwarfs any per-minute saving, and the AI only matched humans on cost-effectiveness for smaller debts. The premium bar is an agent that collects like your best human, at a price comparable to a human, while working every eligible borrower at 8 a.m. on day one. [SRC011] [SRC001] [SRC015]
A 30-day randomized pilot
- Pick one bucket, for example pre-due and day 0 to 5, and one product.
- Randomly split accounts into AI-first with human handover and human-only control.
- Get scripts, disclosure lines, hardship triggers and transfer rules approved by compliance before go-live. [SRC012] [SRC003]
- Score 100% of calls in both arms on the same scorecard, including Hinglish calls. [SRC013] [SRC016]
- Report the KPI table at 1, 3, 7 and 30 days, plus next-cycle on-time payment. [SRC008]
- Expand only where AI-first matches or beats control on recovered share and kept promises.
Questions to ask any collections voice AI vendor
- Can we run a randomized holdout, and will you report kept promises, not just promises?
- How does the agent disclose that it is AI, and how fast can a borrower reach a human?
- Does it dial only inside 08:00 to 19:00, cap attempts, and call from 1600-series numbers? [SRC005] [SRC007]
- Can it take payment on the call and confirm it on WhatsApp?
- How does it detect hardship, disputes and distress in Hindi, English and Hinglish?
- Are all calls recorded, retained six months and scored, or only sampled?
For a fuller bake-off, see how to choose a premium AI voice agent.
Where Qualia fits
Qualia Voice runs inbound and outbound collections calls in Hindi, English and Hinglish, with scripts your team approves before go-live, warm handoff to a human when confidence drops, and payment collection during the live call, on your existing telephony and CRM. CallPulse scores 100% of calls, AI and human, on 15 QA parameters including compliance and fatal-error checks, so the pilot table above comes from every call. [SRC012] [SRC013]
Bring one bucket and a month of call outcomes. We will help you design the randomized pilot before anyone talks price.
Related reading
FAQ
Do AI voice agents work for loan collections?
For reminders and early, small-balance cases, yes, but the best controlled evidence says a basic AI caller recovers less than a good human. In a study of a Chinese online lender, AI callers trailed humans by up to 9 percentage points of the overdue balance around one month past due, and the gap was larger for borrowers with lower credit scores. The AI in that study predates generative AI, so treat it as a floor, not a verdict. [SRC001] [SRC002]
Why do borrowers keep fewer promises made to an AI?
On the first answered call, borrowers who spoke to AI made 21 percentage points fewer promises to pay and 18 points fewer paid the same day, and promises made to AI were kept less often. The authors suggest AI is worse at extracting promises that feel binding. Design for that: ask for payment on the call, confirm a date, amount and channel, and send a payment link right away. [SRC001] [SRC010]
Do I have to tell borrowers they are talking to an AI in India?
RBI's FREE-AI committee report says consumers must be explicitly informed whenever they interact with AI systems and should always have the option to switch to a human, and that AI voice calls should use verified numbers. It is a committee report with recommendations, so check your own legal view, but plan on disclosure plus a one-step route to a person. [SRC003] [SRC007]
Which RBI rules apply to AI collections calls from 2027?
The 2026 Amendment Directions take effect on 1 January 2027. For NBFCs they require logging the time and number of recovery calls, recording outbound and inbound calls with six-month retention, and calling only between 08:00 and 19:00. The lender stays responsible for whoever makes the call, human or AI. [SRC004] [SRC005] [SRC006]
What KPIs should I use to judge an AI collections agent?
Right party contact rate, promise to pay rate, kept promise rate, same-day payment rate, and amount collected as a share of the overdue balance at 1, 3, 7 and 30 days, all against a randomized control group. Add next-cycle on-time payment, because borrowers first called by AI were 1 to 2 points more likely to be late on each of their next 12 payments. [SRC008] [SRC001]
Is the cheapest per-minute collections AI the best choice?
Not if it collects less. In the study, AI matched humans on cost-effectiveness only on smaller debts once direct labour cost was counted. A few rupees saved per minute is small next to a few points of the overdue book. Judge vendors on rupees recovered per rupee spent, measured on your own randomized pilot. [SRC001] [SRC011] [SRC015]
Sources
- How Good is AI at Twisting Arms? Experiments in Debt Collection (NBER Working Paper 33669) - National Bureau of Economic Research
- Can AI Replace Human Debt Collectors? - Yale Insights, Yale School of Management
- Framework for Responsible and Ethical Enablement of Artificial Intelligence (FREE-AI) Committee Report - Reserve Bank of India
- RBI Issues Amendment Directions on 'Conduct of Regulated Entities in Recovery of Loans and Engagement of Recovery Agents' (Press Release 2026-2027/827) - Reserve Bank of India
- Reserve Bank of India (Non-Banking Financial Companies - Responsible Business Conduct) Third Amendment Directions, 2026 (RBI/2026-27/230) - Reserve Bank of India
- Outsourcing of Financial Services: Responsibilities of regulated entities employing Recovery Agents (RBI/2022-23/108) - Reserve Bank of India
- TRAI issues Direction mandating phase-wise adoption of 1600-series by BFSI sector entities (Press Release No. 135/2025) - Telecom Regulatory Authority of India
- Debt Collection AI KPIs - CBOT
- Southwest Recovery Services case study (inbound and outbound) - Skit.ai
- AI Voice Agent for Debt Collection - Kallix
- Collections - Sarvam AI
- AI Voice Agents for BPO Call Centers | Qualia Bits - qualiabits.com
- Automated Call QA Software | Score 100% of Calls | CallPulse - qualiabits.com
- RBI's 2027 Loan Recovery Rules for Collections Calls: What Changes for Human and AI Voice Agents - Qualia Bits
- AI Voice Agent Cost per Resolved Call vs Human Agents (2026) - Qualia Bits
- Same Scorecard for Human and AI Voice Agents - Qualia Bits
Evidence map
- In the regression discontinuity sample, AI callers trailed humans by about 9 percentage points of the initial overdue balance near one month past due and about 5 points at one year, with a larger gap for lower credit scores.
Evidence: SRC001 - In the randomized experiment the gap reached 12 points by day 5 of AI calling, then narrowed to 2 points at day 10 and 0.8 points at day 30 after humans took over, and barely closed over the next 11 months.
Evidence: SRC001 - Borrowers first contacted by AI were 1 to 2 percentage points more likely to be late on each of their next 12 monthly payments.
Evidence: SRC001 - On the first answered call AI conversations were 31 seconds shorter, 21 points fewer borrowers promised to pay, 18 points fewer paid the same day, and promises made to AI were kept less often.
Evidence: SRC001, SRC002 - The study covered about 22 million cases from April 2021 to December 2022, before generative AI.
Evidence: SRC002 - RBI's FREE-AI report says consumers must be explicitly informed when interacting with AI, should always be able to switch to a human, and AI voice calls should use verified 1601 series numbers.
Evidence: SRC003 - RBI's 2026 recovery Amendment Directions take effect 1 January 2027 and, for NBFCs, require call logs, recording with six-month retention, and calls only between 08:00 and 19:00.
Evidence: SRC004, SRC005 - Kept promise to pay, not promise to pay, is the outcome to measure, with attribution windows and a randomized control group.
Evidence: SRC008 - Skit.ai reports 50 to 55% right party contact, 8 to 10% promise to pay and 20 to 25% containment for an inbound US collections deployment.
Evidence: SRC009 - Sarvam markets collections voice agents in 22 Indian languages with pricing from Rs 3 a minute in 11 languages.
Evidence: SRC011