Collections Head's Guide: Training and Quality-Checking 5,000 Recovery Agents on RBI Rules Without a Central QA Team
Quick answer
How collections heads at banks and NBFCs train and verify thousands of field recovery agents on RBI conduct rules: WhatsApp microlearning, AI roleplay calls, evidence.
The short answer
A collections head trains and quality-checks thousands of recovery agents by running three layers on the channel agents already use: short WhatsApp microlearning modules that teach the RBI conduct rules and internal policy in the agent's language; AI roleplay calls that phone each agent, play a borrower, and score whether the agent actually follows the rules under pressure; and trigger-based refreshers whenever a rule changes, a complaint lands or a score dips. Every module completion, quiz score and roleplay result is logged per agent, which is the evidence regulators and internal audit ask for. HDFC Bank trains 32,000 collection agents this way on Leap10x, through WhatsApp, with 95%+ engagement and no app or login.
The exposure: why collections training is different
Most frontline training is about competence. Collections training is about conduct, and conduct failures are public.
The RBI's directions on recovery practices constrain when agents may contact borrowers, what they may say, how they must identify themselves, how complaints are handled, and require lenders to ensure agents are trained and, in the case of direct recovery agents, certified. A single agent who calls a borrower at 10 pm, or uses language the borrower records and posts, becomes a regulatory matter, a reputational event and a board question within a day. We covered the rules themselves in RBI recovery agent training rules 2026 and the certification route in DRA certification and debt recovery agent training.
Three things make this hard to manage through training.
Scale. A mid-sized NBFC may have 2,000 to 5,000 agents; a large bank many times that, spread across in-house teams, collection agencies and business correspondents.
Distance. Agents are in the field, on two-wheelers, in borrowers' neighbourhoods. They are not in a branch, and many are not employees.
Verification. Knowing the rules and following them on a hostile call at 6 pm are different skills. Traditional QA (a central team listening to a sample of calls or phoning agents to test them) cannot cover 5,000 people more than once or twice a year.
Three populations, three problems
In-house field agents. On payroll, on the HRMS, but rarely at a desk. Problem: reach and refresh. They were trained at joining; the rules have changed twice since.
Agency agents. Employed by collection agencies you contract. Problem: visibility. The agency says they are trained; you have a certificate photo. When one misbehaves, the regulator holds you responsible.
Business correspondents and their staff. For banks, the BC network's field staff collect too. Problem: they are two steps removed and often trained on the BC's product, not on your conduct standards. See our KYC and AML training playbook for BCs and DSAs for the parallel compliance challenge.
A workable programme has to reach all three without depending on their employer's systems. The common denominator is that every one of them has a phone number and WhatsApp.
The layered model: teach, verify, refresh
Layer 1: Microlearning that reaches the field
Short modules, three to five minutes, delivered to the agent's WhatsApp at a scheduled time (start of day works). Each covers one rule or one situation: permitted calling hours; identification and disclosure script; what to do when a borrower disputes the debt; handling a third party who answers; recording and privacy; grievance redressal information; prohibited language and conduct.
Every module ends with a two- or three-question check. Completion and scores log per agent. Agency agents and BC staff are enrolled by phone number by their supervisor or by your regional team, with no dependence on the agency's HRMS; our guide to training workers without company email covers the enrolment mechanics.
Language matters more here than almost anywhere. Agents speak to borrowers in Marathi, Tamil, Bengali, Kannada, Hindi. Train them in that language, with audio, or the training is theatre. Leap10x translates and narrates in 70+ languages.
Layer 2: AI roleplay calls that verify conduct
This is the layer that was not possible three years ago. An AI voice agent phones the collections agent at a scheduled time, identifies itself as a training call, and plays a borrower: two EMIs overdue, anxious, then angry, then asking for the grievance number. The agent handles the call. The AI scores it against your rubric: did the agent identify themselves correctly, state the purpose, stay within permitted language, avoid threats or misrepresentation, provide the grievance redressal information, close appropriately. Compliance items are pass/fail. The transcript and score land in the dashboard.
Leap10x Converse does exactly this: it calls frontline workers, runs roleplays and knowledge checks in their own language, and scores comprehension automatically. It replaces the manual quality calls that would otherwise take a central team weeks to complete across a distributed workforce. What an AI voice roleplay call is explains the mechanics; AI roleplay for BFSI collections, sales and service has the collections-specific scenarios.
The scale changes what QA means. A central team of eight can verify a few hundred agents a quarter. An AI agent can verify every agent every month, and every new agent before their first live call.
Layer 3: Refresh on trigger
Static annual training decays; the cost of employee forgetting is well documented. Refreshers should be triggered, not calendared:
- Rule change: RBI issues a new direction; a three-minute module goes to every agent within 48 hours; a roleplay call follows within two weeks to verify.
- Complaint: a borrower complaint names an agent; that agent gets a targeted refresher and a roleplay call on the specific conduct.
- Score dip: an agent's roleplay score on a compliance item falls below threshold; automatic refresher and re-test.
- Portfolio event: a new product or a new bucket (e.g., agents moving from 30-day to 90-day delinquency) triggers the relevant modules.
The compliance refresher calendar is the baseline; triggers sit on top.
What to teach: the conduct curriculum
A core set of modules, each three to five minutes, each with a check. Sequence them into an onboarding journey for new agents and assign individually for refreshers.
- Who you are and who you represent: identification, authorisation letter, disclosure.
- When you may call and visit: permitted hours, frequency, borrower requests to stop.
- What you may and may not say: prohibited language, no threats, no misrepresentation of consequences.
- Privacy: no disclosure to third parties, no discussion of debt with family or employer, handling of documents and photos.
- Disputes and hardship: when a borrower disputes the debt or reports hardship, what to do and who to route to.
- Grievance redressal: the information every borrower must be able to receive, and how to give it.
- Recording and documentation: what to log after each contact, why it protects the agent.
- Digital lending and app-based recovery specifics, where applicable.
- Your own safety and de-escalation in the field.
- Internal policy on settlements, receipts and cash handling.
Map each module to the specific RBI direction or internal policy clause it implements, in the module metadata. When an auditor asks "how do you train on X", the answer is a module number and a completion report.
What to verify: the scorecard
The roleplay rubric should be short and mirror what your QA team scores on live-call audits. Six to eight items:
| Item | Type |
|---|---|
| Correct identification and disclosure at opening | Pass/fail |
| Purpose stated; borrower's identity confirmed before discussing account | Pass/fail |
| No prohibited language, threats or misrepresentation | Pass/fail |
| Grievance redressal information provided when asked or at close | Pass/fail |
| Handling of dispute or hardship per policy | Scaled, with evidence |
| De-escalation and tone | Scaled, with evidence |
| Accuracy of amounts, dates and options stated | Scaled |
| Appropriate close | Scaled |
Any pass/fail miss means the agent is not released (new joiners) or gets an immediate refresher and re-test (existing). Calibrate the AI's scoring against your QA team on 20 transcripts before rollout.
The evidence a regulator or internal audit will want
Per agent: identity, employer (in-house, agency name, BC name), modules completed with versions and dates, quiz scores, roleplay dates and scores, refreshers triggered and completed, certificate status (including DRA certification where applicable). Per portfolio: coverage percentage by population, outstanding list, rule-change response time (direction issued to 100% trained).
The system generates all of this if agents are trained and verified through it. Export to Excel or PDF; Enterprise plans on Leap10x add API export and custom reports. Our EHS-oriented piece on proving training completion to an auditor describes the five elements of defensible evidence; they apply identically here.
A 90-day rollout for a 5,000-agent network
Days 1–15: Curriculum and rubric. Build the ten modules from your existing policy documents (AI authoring from PDFs makes this days, not weeks). Translate into the agents' languages. Build the roleplay rubric with your QA team. Write five borrower scenarios.
Days 16–30: Pilot with one region, in-house agents only. 300 agents. Modules at day start over WhatsApp. Roleplay call to each within two weeks. Compare roleplay pass rates with the last live-call audit for the same region.
Days 31–60: Add agencies. Enrol agency agents by phone number via agency supervisors. Same modules, same roleplay. Give each agency a view of its own agents' completion and scores, and write the coverage expectation into the agency SLA.
Days 61–90: Full network, triggers live. All regions, BCs included. Rule-change trigger tested with a mock direction. Complaint trigger wired to your grievance system. First monthly compliance report to the risk committee: coverage, pass rates, outstanding, response times.
What this does not fix
Training does not fix incentive structures that reward aggressive recovery. It does not fix agencies you should not have contracted. It does not replace live-call monitoring where you have it; it makes the sample you monitor far better prepared. And an AI roleplay score is evidence of capability, not a guarantee of behaviour. Use it to decide who is ready and who needs coaching, and keep human judgement in the loop for disciplinary decisions.
Five borrower scenarios for roleplay, with scoring notes
A starting library. Each is written to be run as a voice call in the agent's working language, five to seven minutes, and to exercise specific scorecard items.
1. The borrower who wants to skip the formalities. Two EMIs overdue, friendly, says "I know who you are, just tell me the amount". Tests whether the agent still identifies themselves and the lender and confirms the borrower's identity before discussing the account. Scoring: identification and identity confirmation are pass/fail; skipping either fails the scenario regardless of how pleasant the call was.
2. The hardship case. Borrower has lost a job, is anxious, asks what happens if they cannot pay this month. Tests handling hardship per policy: acknowledging, not threatening, explaining the actual options and the process, routing to the right team if policy requires. Scoring: any language implying consequences beyond what policy states is a compliance fail; acknowledgement and correct routing are scaled with quoted evidence.
3. The third party who answers. The borrower's spouse or parent picks up and asks what the call is about. Tests privacy: the agent must not disclose the debt or its details to a third party, must ask for the borrower or leave a neutral callback message. Scoring: any disclosure of the debt to the third party is a fail. This scenario has the highest failure rate in most pilots and the highest regulatory exposure.
4. The angry repeat caller. Borrower has been contacted several times, is furious, threatens to complain and to post on social media, demands the grievance number. Tests de-escalation, permitted language under pressure, and providing grievance redressal information. Scoring: grievance information provided is pass/fail; de-escalation is scaled; any retaliatory or threatening language is a fail.
5. The settlement request. Borrower offers to pay a lump sum less than the outstanding and asks the agent to "adjust it". Tests knowledge of settlement policy, authority limits, receipt and documentation rules, and not promising what the agent cannot deliver. Scoring: any commitment outside the agent's authority is a fail; correct explanation of process is scaled.
Run all five for every new agent before release. Rotate two per month for tenured agents. Track the pass rate per scenario across the network: if scenario 3 fails at 40% network-wide, the problem is your training content or policy clarity, not 2,000 individual agents, and the fix is a module, not a memo. The 25 AI roleplay scenarios for frontline teams library has additional collections and service cases to extend from.
FAQ
Do agency agents need to be on our systems?
No. They are enrolled by phone number; their employer's HRMS is irrelevant. Their records sit under the agency's name in your reporting.
Can this satisfy the DRA certification requirement?
No. DRA certification runs through IIBF's accredited process. This programme prepares agents for it, keeps them current after it, and verifies conduct continuously. See the DRA certification guide.
Agents will not answer a training call in the field.
Pick-up rates are high when the call is scheduled (start of day, agreed slot) and short (five to eight minutes). Agents answer phones for a living.
Does the AI understand code-switching between Hindi and English?
Ask for a live demo in the exact language mix your agents use, with a native speaker from your team on the line. It is the right test for any vendor.
Related reading
- RBI Recovery Agent Training Rules 2026: Compliance Guide
- DRA Certification: Debt Recovery Agent Training in India
- NBFC Compliance Training on WhatsApp 2026
- IRDAI, SEBI and RBI Compliance Training for Field Staff
- KYC & AML Training for Bank Agents, DSAs and Business Correspondents
- Microfinance Field Officer Training
- Leap10x for BFSI
CTA: Bring your recovery conduct policy and one borrower scenario. We'll turn the policy into a WhatsApp module and place an AI roleplay call to you as the agent, in your language, during the demo. 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.
