Measuring Collections Training: Completion to Complaints

Quick answer
No Indian study links agent training to recovery. How collections leaders can measure it honestly, from completion and call audits to promises and complaints.
A training head presents to the collections leadership team: 91% completion. The head of collections nods and asks one question: "And what did that do to my numbers?" The room goes quiet.
Here's the uncomfortable truth behind that silence. We looked for evidence on measuring collections training in India, and we found no published study, lender case study or independent report showing that agent training moves recovery. There are vendor claims. One cites a 20% improvement in call quality score in a testimonial (Convin). Another claims 15% to 30% better collection rates in six to twelve months from conversation intelligence, with no method or control (Gistly). Field experiments exist, but they test AI or personalised strategy rather than agent training (EMNLP 2023).
So if anyone tells you that their programme lifts recovery by a specific percentage, ask to see the comparison group. And that includes us. We sell training software, and we'd love to claim this. We can't.
What you can do is measure honestly. This post sets out what's worth measuring, how to define it, and how to test whether training is making a difference. It's for training heads and collections leaders building a case for 2027-28.
The metrics collections already uses
Start with what leadership already watches, because training will be judged against these whether you like it or not.
| Metric | Where it's used | What it means |
|---|---|---|
| Collection efficiency | Listed lender reporting, such as Mahindra Finance | Collected against what was due |
| Stage 2, Stage 3, credit cost | Listed lender reporting | Asset quality |
| Bucket roll rate | Internal | Share moving into a worse bucket |
| Cure rate | Internal | Share returning to current |
| Resolution rate and recovery yield against cost to collect | Job ads | Efficiency of the operation |
| Promise-to-pay conversion and kept rate | Job ads | Quality of conversations |
| Right-party contact | Job ads | Reaching the right person |
Definitions aren't standard, which causes real trouble. Cure rate may be measured on balances or on accounts (Fusion CX). Promise-to-pay could be measured against connects or against calls (Prodigal). Before you start, agree on one written definition per metric with your analytics team. Otherwise, you'll end up arguing about the denominator rather than the result.
Why recovery is a poor test of training
Collection efficiency moves for many reasons, and training is probably not the biggest of them.
| What moves recovery | Example |
|---|---|
| Portfolio mix | A new product, or a shift in ticket size |
| Sales and write-offs | Crisil notes about 7% of small finance bank portfolios were written off or sold, which helps headline numbers (Crisil) |
| The economy and seasons | Harvest cycles, festivals, floods |
| Allocation | Who gets which accounts |
| Agency changes | Moving accounts to a better agency |
| Strategy | A new dialler, a new voice bot, a new settlement policy |
| Incentives | A change in the payout grid |
Delinquency has been falling almost everywhere at June 2026 (CRIF). If you rolled out a new training programme this year and collections improved, you can't claim the credit from a simple before-and-after. A CFO who knows this will discount your slide, and they'd be right to.
That doesn't make recovery irrelevant. It just means you can't use it on its own. You need either a comparison group or measures closer to what training can actually influence.
A ladder of measures
Think of it as four rungs, moving from what's easy to attribute towards what matters most. It's similar in spirit to the Kirkpatrick model, which we've covered elsewhere.
| Rung | Question | Measures | How far to trust it as evidence of training |
|---|---|---|---|
| 1. Reach | Did they get it? | Coverage, completion by tenure, role and language | Direct, but it says nothing about learning |
| 2. Understanding | Do they know it? | Per-question quiz scores, spot checks | Direct, if the questions are good |
| 3. Behaviour | Do they do it? | Call audit scores, field observation, log accuracy, third-party calls | Strong. Closely tied to what you taught |
| 4. Outcomes | Did it matter? | Complaints, promise-kept rate, roll rate, cost to collect | Weaker. Needs a comparison group |
Most training teams report rung 1 and stop. Most business leaders ask about rung 4. The useful work happens at rungs 2 and 3, and you can do it with data you may already have.
Rung 1: Reach, done properly
Completion means little if the denominator's wrong. Report it with two supporting numbers.
- Coverage: of the people who had accounts or made calls this week, what share are on your training list?
- Freshness: of those marked active on your list, what share were seen working recently?
Our learner list post has the method. Also split completion by tenure. A healthy overall number can hide the fact that your newest agents, the ones most likely to make mistakes, are getting nothing.
Rung 2: Understanding
Look at results per question, not just an average score.
Say 88% of agents get the contact-hours question right, and only 41% get the "neighbour asks how much" question right. An average of 70% hides this completely. Per-question results tell you what to reteach, and they point to bad translations, too. If Tamil speakers get question 3 wrong much more often than others, check the Tamil version first.
Write questions on judgement, not recall. "The borrower asks you to call at 8 pm. What do you do?" tells you far more than "What are the permitted hours?"
A useful test at rung 2 is to re-test after 30 days. If scores drop sharply, the content didn't stick, and you may need spacing and reminders. Our post on tracking, measuring and intervening goes further.
Rung 3: Behaviour
This is the most valuable rung, and it's about to get easier. From 1 January 2027, lenders must log the time and number of calls and record content, with records kept for six months (para 454P for commercial banks, 100P for NBFCs) (RBI). Until 31 December 2026, existing paragraphs apply.
So you can tie measures directly to the lessons you've taught.
| Lesson | How to observe it |
|---|---|
| Contact hours | Share of calls and logged visits outside the window. A system query |
| Who hears about dues | Audit sample: does the agent discuss dues with a third party? |
| False statements, threats | Search recordings for words such as police, warrant, arrest; or sample by hand |
| Identifying yourself | Audit sample: name and company stated at the start? |
| Settlement promises | Audit sample, plus complaints that mention a promise |
| Repossession notice check | File audit: notice on file before each seizure |
| Log accuracy | Spot-check logs against recordings or field visits |
| Field conduct | Ride-alongs, observed visits, trace tests |
The audit needs a fixed scorecard, so you can compare month to month. Keep it to perhaps ten yes-or-no items. Have someone other than the agent's own team leader do at least some of it, as team leaders tend to score their own teams kindly.
The field is your blind spot. Recordings don't show a doorstep. Our agency checklist includes a trace test that may help.
One caution: last-week-of-month behaviour differs. Sample across the whole month, and look at the 28th separately.
Rung 4: Outcomes
These are the numbers that leaders care about most, and you've got to be careful with them.
Complaints are the most useful outcome for conduct training. Count those that mention threats, hours, third parties or false promises, per thousand accounts, by agency. We couldn't find an industry benchmark, so track against yourself. Be careful: if you make it easier to complain, the count may rise, which is good, not bad. RBI's Ombudsman reported 13,34,244 complaints in FY2024-25, up 13.55%, with NBFCs 14.80% of those at Ombudsman offices, but its web page shows no recovery-agent category (RBI). So you can't compare yourself with a national figure for this yet.
Promise-kept rate is the outcome that's closest to a collector's skill. If you've taught people to get a real promise, more of them should be kept. Agree on the definition first. It's also vulnerable to gaming, since agents can log fewer promises to look better, so watch the promise rate alongside it.
Roll rate and collection efficiency come last. Only use these with a comparison group.
Attrition at 30 and 90 days isn't a collections metric, but it's useful. If better onboarding helps people stay, that saves you hiring and certificate costs. Our budget worksheet shows how much this could be worth.
How to run a real comparison
This isn't as difficult as it sounds, and nobody in this market seems to have published one. Doing it would give you real evidence, and a real advantage.
Option 1: Matched territories
Pick two similar territories or agencies: similar product, ticket size, bucket mix and tenure. Roll out training in one, and hold the other for a few months. Compare the change in each one, not just the end point.
Option 2: Staggered rollout
If holding a group back forever feels wrong, as it should for mandatory content, roll out in waves. Wave 1 gets it in month 1, wave 2 in month 3, and wave 3 in month 5. Each wave serves as a comparison for the next. It's fair, and it's practical, because you were probably going to phase it anyway.
Option 3: Same people, different topics
Teach a single topic to everybody, and don't teach a similar one yet. If audit scores improve on the taught topic, and not on the untaught one, that's reasonable evidence the lesson did it.
| Design choice | Advice |
|---|---|
| Don't hold back compliance essentials | Teach required conduct content to everyone. Test the format or the add-ons |
| Decide measures in advance | Write them down before you begin, so you can't pick winners afterwards |
| Include leavers | Otherwise, you're just measuring who stayed |
| Allow time | At least two cycles. Longer for roll rates |
| Watch for spillover | Agents talk to each other, and this is especially true in the same agency |
| Have someone independent pull the data | It helps people to trust the result |
A worked example
Here's an illustration. All numbers are made up.
| Territory A (with lessons) | Territory B (comparison) | |
|---|---|---|
| Agents | 120 | 110 |
| Call audit, third-party item, before | 62% pass | 60% pass |
| Call audit, third-party item, after 3 months | 81% pass | 66% pass |
| Complaints per 1,000 accounts, before | 2.1 | 2.0 |
| Complaints per 1,000 accounts, after | 1.5 | 1.9 |
| Promise-kept rate, before | 54% | 55% |
| Promise-kept rate, after | 57% | 56% |
| Collection efficiency, change | +1.2 points | +0.9 points |
How would you read this? Behaviour improved clearly, compared with the other territory. Complaints look better, though the numbers are small. The promise-kept difference is modest, and could be noise. Collection efficiency barely differs, and you shouldn't claim it. That's an honest and credible story: "Conduct improved measurably. Recovery wasn't visibly affected in three months. We'll keep watching." A CFO will trust that far more than a claim of a 20% uplift.
Getting the data without a data team
Most training heads don't have an analyst. You can still get a long way with a spreadsheet, as long as you set it up once and keep it simple.
Start with one join: a list of agents with their agency, role, language and joining date, matched to lesson completions by mobile number. That gives you rung 1 and rung 2. Then ask your collections MIS person for one extract a month, by agent: calls made, calls outside the window, promises logged and promises kept. Add your call audit sheet, and you've got rung 3 and a start on rung 4.
Three practical tips:
- Fix the agent ID early. If the dialler uses one ID, the app another and the training tool a phone number, you'll spend every month matching names by hand. Build one mapping table and keep it current.
- Freeze your definitions in a one-page note. What counts as a promise? What's a conduct complaint? When someone leaves the team, the note stays.
- Report the same way every month. A plain table that stays the same will build more trust than a clever dashboard that changes each quarter.
If your MIS team's stretched, ask for less rather than waiting for everything. One territory, one lesson topic and three measures are enough to start.
What to put on one page
For your monthly report, here's a suggestion.
| Section | Metric |
|---|---|
| Reach | Coverage, freshness, completion overall and for under-30-day tenure |
| Understanding | Three weakest questions this month, and what you're doing about them |
| Behaviour | Call audit score, three weakest items; share of contact outside permitted hours |
| Certificate | Share of agents holding it, and those at risk |
| Outcomes | Conduct complaints per thousand accounts; promise-kept rate |
| Test in progress | What you're comparing, and the latest read |
Note what's not here: hours of training delivered, or number of courses. Those measure effort, not effect. Our ROI framework and post on measuring ROI for deskless workers offer other perspectives.
Common mistakes
- Reporting completion as success. It's a start.
- Before-and-after with no comparison. Too many other things are changing.
- Changing definitions midway. Write them down at the start.
- Team leaders scoring their own people. Sample independently.
- Measuring only tele-calling, as it's easier. Field conduct is where a lot of the risk lies.
- Treating a rise in complaints as failure. It may mean your process is finally trusted.
- Ignoring tenure. New agents and old hands behave differently.
- Promising recovery uplift to get budget. You'll be held to it.
Where Leap10x fits
I'm the founder of Leap10x, so here's where we help, and where we don't.
Leap10x MicroLearning supplies rungs 1 and 2: completion and per-question scores by agency, team, person and language, exportable as PDF or CSV, for lessons delivered on WhatsApp with no app or login. HDFC Bank has had "32,000 collection agents trained" on the platform, with "95%+ learner engagement". Those are reach and engagement numbers. They aren't a recovery result, and we don't claim one.
Rung 3 comes from your call recordings, your audits and your speech analytics, if you've got them. Rung 4 comes from your systems. If you run a comparison with us, we'd be glad to help you design it, and we'd rather you publish what you find, whichever way it goes.
FAQ
Does training improve recovery rates? We found no published Indian study showing that it does. It's a reasonable hypothesis, so test it with a comparison group.
What should we measure first? Coverage and completion by tenure, then per-question scores, then call audit items tied to specific lessons.
Which outcome is most directly linked to training? Conduct complaints and audit scores are easier to attribute than recovery. Promise-kept rate is a reasonable next step.
How do we test if training works, when everyone needs it? Use a staggered rollout, where each wave acts as a comparison for the next. Don't withhold required conduct content.
Are there industry benchmarks? We couldn't find reliable ones for training completion, complaints per thousand accounts or call audit scores. Track against yourself.
Answer the question honestly
When the head of collections asks what training did to the numbers, the best answer isn't a big percentage. It's this: "Here's what changed in behaviour, compared with a similar team. Here's what we can and can't say about recovery. Here's the test we're running next." Nobody else in the room will be able to say that.
CTA: Want help designing a two-territory comparison? Book a Demo.
Frontline training
Training software built for deskless teams
Onboarding, SOPs, safety refreshers, and compliance evidence for workers with no corporate email and no company device — tracked by site, role, and shift.


