Lenders evaluating or scaling AI voice collections need a clear view of what to measure and why. This FAQ is written for collections heads, risk teams, and operations leaders at Indian banks and NBFCs who want to move beyond gut-feel reviews and build a defensible, data-backed view of how AI-led outreach is performing against human-only collections.
1. What KPIs should we track when we start using AI voice calls for collections?
The core KPIs to track are contact rate, resolution rate, promise-to-pay (PTP) conversion, PTP-kept rate, and cost per account resolved. Contact rate tells you whether calls are actually reaching borrowers rather than going unanswered or to voicemail, which matters given how many Indian borrowers screen unknown numbers. Resolution rate measures the share of contacted accounts that result in a payment or a firm commitment, while PTP conversion and PTP-kept rate together tell you whether promises made on a call actually translate into money collected. Cost per account resolved lets you compare AI-assisted buckets against agent-only buckets on a like-for-like basis. Most lenders also track these by delinquency bucket (0-30, 31-60, 61-90 DPD) separately, since AI performance and appropriate tone differ meaningfully by how overdue an account is.
2. How do we measure whether AI collection calls are actually improving recovery rates?
The most reliable way is a controlled comparison: run AI-led outreach on one segment of a bucket and human-agent-led outreach on a statistically similar segment, then compare resolution rate and amount collected over the same window. This avoids the common trap of comparing AI performance against a stale historical baseline that doesn't account for seasonality or portfolio mix changes. Track the comparison over at least one full collection cycle, since early-stage AI adoption often shows a dip before agents and scripts are tuned. Many lenders also segment the comparison by DPD bucket and loan type, because AI-voice impact on early-bucket, low-friction reminders tends to show up faster than impact on legal-stage recovery, where human judgment still plays a larger role.
3. What is promise-to-pay (PTP) rate and why does it matter for AI collections?
Promise-to-pay rate is the percentage of contacted borrowers who commit to a payment date and amount during the call. It matters because it is the earliest reliable signal of whether a collections conversation is working before the actual payment settles days later. A high PTP rate with a low PTP-kept rate points to a different problem than a low PTP rate — the former suggests borrowers are agreeing just to end the call, which can happen with poorly tuned scripts, while the latter suggests the message isn't landing or the contact isn't reaching the right person. AI voice agents that ask for a specific date and amount, and repeat it back for confirmation, generally produce more reliable PTPs than open-ended reminder calls, which is worth watching closely when reviewing weekly reports.
4. Can AI accurately track PTP-kept rates without manual reconciliation?
Yes, when the AI voice platform is integrated with the loan management system (LMS) and payment gateway, PTP-kept rate can be tracked automatically by matching the promised date and amount against actual repayment records. This removes the manual reconciliation collections teams have historically done in spreadsheets, which is slow and error-prone at scale. Automated matching also allows lenders to flag partial payments, early payments, and broken promises in near real time, feeding that outcome back into the next call's script — for instance, escalating tone slightly or looping in a human agent sooner for a borrower who has broken two consecutive promises. This closed-loop tracking is one of the more tangible efficiency gains lenders report once AI and LMS data are properly connected.
5. What is a good contact rate to expect from AI outbound collection calls in India?
Contact rate depends heavily on the delinquency bucket, time of day, and quality of the phone number on file, so there is no single universal benchmark — what matters is tracking your own trend over time and comparing it against your existing agent-dialer contact rate on the same segment. AI dialers can improve contact rate modestly by calling within RBI-permitted hours and by intelligently retrying numbers at times when a borrower is statistically more likely to answer, based on past pickup patterns. Indian borrowers frequently do not answer unknown numbers, so caller ID registration and consistent, recognisable calling patterns also influence contact rate meaningfully. Lenders should track contact rate as a leading indicator, since resolution and PTP metrics are meaningless if the call never connects.
6. How should we measure cost per account collected when comparing AI to human agents?
Cost per account collected should include the AI platform's per-minute or per-call cost, any human escalation cost for the same account, and be divided by the number of accounts that reached resolution — not just accounts attempted. This is important because a fair comparison must account for the fact that AI typically handles high-volume, low-complexity reminder calls, while human agents are reserved for negotiation-heavy or legal-stage accounts. Blending these into one undifferentiated cost figure understates AI's efficiency on the segment it's actually built for. Most collections leaders find it more useful to report cost per resolved account by bucket, since early-bucket reminder calls at scale is where AI shows the clearest cost advantage over agent-only models.
7. What reporting cadence works best for tracking AI collections performance?
A daily operational dashboard combined with a weekly management review works well for most Indian lenders running AI voice collections at scale. Daily tracking should cover contact rate, PTP rate, and any calls flagged for escalation or compliance review, since these need same-day or next-day action. The weekly review is where resolution rate, amount collected, cost per account, and PTP-kept trends should be assessed against targets, ideally broken down by bucket and portfolio segment. Monthly or quarterly reviews are better suited for strategic questions — such as whether AI-led buckets are shifting roll-rates to later stages, or whether specific call scripts need retiring — since these effects take longer to show up clearly in the data.
8. Is it possible to measure the impact of AI collections on roll-rates, not just individual call outcomes?
Yes, and this is one of the more valuable but under-tracked metrics — roll-rate measures what share of accounts move from one delinquency bucket to the next (say 0-30 to 31-60 DPD) despite collection efforts, independent of any single call's outcome. Tracking roll-rate for AI-assisted segments versus comparable human-only segments over several months shows whether early intervention is actually preventing accounts from deteriorating, which is the real business objective behind most collections activity. A lender might see similar PTP rates between AI and human calling but a meaningfully better roll-rate for the AI-assisted segment simply because of more consistent, higher-frequency early touchpoints. This is a slower-moving metric, so it needs a longer observation window than weekly call-level KPIs.
9. What are the risks of over-indexing on call volume or resolution rate as the only success metrics?
Over-indexing on volume or resolution rate alone risks incentivising behaviour that technically improves the number but damages the borrower relationship or invites regulatory scrutiny — for example, excessive call frequency to the same borrower purely to inflate contact attempts. RBI's Fair Practices Code expects lenders to avoid harassment and to calibrate contact frequency and timing sensibly, so any KPI framework should include a complaint rate or borrower-sentiment metric alongside pure volume and recovery numbers. A collections program that shows strong resolution rates but a rising complaint or opt-out rate is trading short-term recovery for long-term reputational and compliance risk. The healthiest KPI frameworks pair recovery metrics with a few guardrail metrics that catch this kind of drift early.
10. How do we know if our AI voice collections program is ready to scale to more portfolios or buckets?
Readiness to scale is best judged by consistency of core metrics — contact rate, PTP-kept rate, and resolution rate — holding steady or improving across at least a few consecutive collection cycles on the initial pilot segment, not just a single strong month. It's also worth confirming that escalation handoffs to human agents are working smoothly, meaning the AI is correctly identifying disputes, hardship cases, or complex negotiations and routing them rather than attempting to resolve everything itself. A useful practical test is expanding to one adjacent bucket or portfolio type at a time and watching whether the same metric trends hold, rather than rolling out across the entire loan book at once. Lenders who scale gradually, checking metrics at each stage, tend to catch script or routing issues before they affect a large volume of borrowers.
Related Reading
Related reading
Talk to YuVerse
Want a clearer view of what your collections KPIs should look like with AI voice in the mix — talk to YuVerse.