Which Recovery KPIs Matter Most for NBFCs?
Recovery KPIs measure whether collections activities contribute to successful borrower contact, repayment commitments, and actual payments.
Traditional collections teams already track many of these metrics. AI observability adds conversation-level evidence to explain changes in performance.
| KPI | Definition | How to Measure It |
|---|
| Right-party contact rate | Verified borrower contacts divided by connected calls | Check whether the borrower was appropriately verified before sensitive debt information was discussed |
| Promise-to-pay rate | Verified borrower contacts resulting in a recorded payment promise | Count promises with a completed confirmation or read-back, not merely a logged field |
| Promise-kept rate | Promises fulfilled by the agreed date divided by promises due in the measurement cohort | Join each promise to payment records and identify broken commitments |
| Roll rate | Accounts moving to the next delinquency bucket, such as 30 to 60 days past due | Measure from account records and analyze by campaign, agent version, or previous contact outcome |
| Cure rate | Delinquent accounts returning to current status | Measure from account records and link outcomes to previous collection interactions |
| Cost to collect | Operating cost per unit of currency recovered | Combine attributable telephony, model, platform, staffing, and collection costs with cash recovered |
| Collection efficiency rate | Amount collected divided by amount due for the defined portfolio or period | Calculated from payment and loan management systems |
1. Right-Party Contact Rate
Right-party contact (RPC) rate measures the percentage of connected calls that reach and appropriately verify the intended borrower.
Formula:
Right-Party Contact Rate = (Verified Borrower Contacts / Connected Calls) × 100
RPC is also a conduct metric.
If an AI voice agent counts a call as right-party contact when it actually spoke to a relative, the problem extends beyond inaccurate reporting. It can create a privacy or unauthorized disclosure risk if it reveals sensitive debt information.
For NBFCs, the verification process should follow approved authentication procedures and avoid unnecessarily exposing personal financial information.
2. Promise-to-Pay Rate
The promise-to-pay (PTP) rate measures how often verified borrower conversations result in a repayment commitment.
Formula:
Promise-to-Pay Rate = (Verified Contacts with a Valid PTP / Eligible Verified Contacts) × 100
An AI voice agent should not count a promise simply because a payment date appears in a conversation log.
The monitoring process should verify whether the borrower actually agreed to the amount and repayment date and whether the agent completed the required read-back.
3. Promise-Kept Rate
Promise-kept rate measures whether recorded repayment commitments were fulfilled.
Formula:
Promise-Kept Rate = (Promises Fulfilled on Time / Promises Due) × 100
The definition should specify whether partial payments qualify.
A strong promise-to-pay rate paired with a weak promise-kept rate may indicate that the agent is securing unrealistic commitments, failing to confirm repayment details, or missing hardship signals.
This is why you should always analyze the PTP rate and PTP-kept rate together.
4. Roll Rate
Roll rate measures the proportion of accounts moving from one delinquency bucket to another over a defined period.
For example, an NBFC may track how many accounts move from 30 days past due (DPD) into the 60 DPD bucket.
The account system provides the roll-rate outcome. Conversation observability helps teams investigate whether unsuccessful contacts, failed repayment arrangements, or incorrect information contributed to the deterioration.
5. Cure Rate
Cure rate measures the share of delinquent accounts that return to current status.
It is particularly useful for understanding whether early-stage collections interventions help restore repayment performance.
6. Cost to Collect
Cost to collect helps NBFCs determine whether AI-driven collections are economically efficient.
Formula:
Cost to Collect = Total Attributable Collections Cost / Amount Collected
Include relevant telephony, AI model, infrastructure, platform, and operational costs.
For accurate reporting, clearly define which collections costs and payments are attributed to the AI channel.
7. Collection Efficiency Rate
Collection efficiency measures how much of the amount due was actually recovered during the reporting period.
Formula:
Collection Efficiency Rate = (Amount Collected / Amount Due) × 100
For NBFCs, analyze collection efficiency across loan products, DPD buckets, customer segments, and AI voice agent campaigns.
These breakdowns help distinguish an agent-quality issue from differences in portfolio composition.
Which Regulatory Conduct KPIs Should NBFCs Track?
Recovery performance cannot be separated from borrower protection.
In India, NBFCs must consider applicable Reserve Bank of India requirements governing collections, outsourcing, customer confidentiality, and recovery-agent conduct.
The RBI's August 12, 2022 circular on recovery agents addresses intimidation, harassment, privacy intrusion, persistent calling, inappropriate messages, threatening calls, false representations, and calling borrowers before 8:00 a.m. or after 7:00 p.m. for recovery of overdue loans. Consider its scope and applicable exceptions, including the treatment of certain microfinance loans.
For NBFCs deploying AI voice agents, these requirements make collections conduct monitoring an important operational responsibility.
The following monitoring KPIs are recommended and mapped to relevant conduct risks. They are not presented as an RBI-prescribed list of KPI names.
| Conduct KPI | Definition | How to Measure It on Live Calls |
|---|
| Verification before disclosure | Calls where identity was appropriately confirmed before sensitive debt details were discussed | Check the conversation sequence and identify the turn where debt information was first disclosed |
| Third-party disclosures | Calls where unauthorized parties received debt-related information | Detect references to loan details when the intended borrower was not verified |
| Calling-hour compliance | Recovery calls placed within the applicable permitted calling window | Compare call timestamps against approved hours and applicable market rules |
| Harassment or threatening language | Calls containing intimidation, abusive language, threatening statements, or inappropriate pressure | Evaluate conversations against the NBFC's approved conduct policies |
| False legal claims | Calls containing unsupported statements about legal action, arrest, or other consequences | Check claims against approved scripts and permissible representations |
| Required disclosures given | Calls where required identity, purpose, and market-specific statements were appropriately communicated | Evaluate disclosure requirements against the applicable jurisdiction and call context |
| Complaints per 1,000 calls | Complaints identified during calls divided by total calls, multiplied by 1,000 | Detect complaint intent and verify appropriate routing |
| Hardship referrals | Calls containing hardship cues and the share receiving appropriate handling or referral | Detect hardship language and verify the subsequent response |
| Dispute handling | Calls where a borrower disputed the debt and the required dispute workflow was followed | Check acknowledgment, escalation, and applicable collection restrictions |
| Opt-outs honored | Calls containing valid stop or revocation requests that were handled according to applicable rules | Combine call analysis with subsequent contact logs |
| Contact-time requests honored | Borrower contact preferences recorded and respected where applicable | Check the conversation and later calling history |
Why Calling-Hour Compliance Matters
For recovery calls covered by the RBI's August 12, 2022 circular, calls before 8:00 a.m. and after 7:00 p.m. are prohibited.
An NBFC monitoring its AI collections agents should therefore validate call timestamps, borrower time zones where relevant, campaign configurations, and applicable product-specific rules.
For example, a call scheduled at 6:55 p.m. might continue beyond 7:00 p.m. The NBFC should define an operational policy for handling this situation in line with applicable requirements rather than assuming the scheduling system alone guarantees compliance.
Why Third-Party Disclosure Detection Matters
A borrower may not answer a collections call. Instead, a spouse, family member, colleague, or another individual may pick up.
An AI agent that reveals loan balances, overdue amounts, or other sensitive financial information before appropriate verification can create serious privacy and conduct concerns.
Monitoring should detect both explicit disclosures and indirect statements that may reveal sensitive information.
Why Harassment Detection Matters
Recovery agents must not use intimidation or harassment to collect.
AI observability can evaluate conversations for threatening phrases, inappropriate pressure, misleading representations, and repeated conduct patterns.
However, transcript-based detection alone is not sufficient to determine whether excessive calling occurred. You also need to review call logs to assess contact frequency over the relevant period.
Which Conduct KPIs Matter Across International Markets?
NBFCs and financial-services organizations operating in multiple jurisdictions may also need jurisdiction-specific monitoring.
The original collections KPI framework includes the following international considerations.
| KPI | Relevant International Frameworks or Considerations |
|---|
| Verification before disclosure | US Regulation F §1006.6(d); UK FCA CONC debt-collection requirements |
| Third-party disclosures | Jurisdiction-specific privacy and debt-collection restrictions |
| Required disclosures | US Regulation F §1006.18(e); applicable UK and other market requirements |
| Threats or false legal claims | US Regulation F §1006.18; applicable local prohibitions on misleading or threatening conduct |
| Complaints per 1,000 calls | Internal operational KPI; distinguish from UK FCA complaints reporting per 1,000 accounts |
| Hardship referrals | UK FCA CONC, applicable Australian hardship frameworks, and local customer-assistance requirements |
| Dispute handling | Jurisdiction-specific requirements for disputed debts |
| Opt-outs honored | US TCPA and other applicable communications-consent requirements |
| Contact-time requests honored | Applicable local contact restrictions and customer preferences |
Configure these controls by market, and review them against the current rules that apply to the lender, loan type, agent, and communication channel.
Some requirements need call logs as well as transcripts.
For example, US Regulation F includes a telephone-call-frequency presumption for more than seven calls within seven consecutive days under specified circumstances. Australian debt-collection guidance also addresses recommended contact frequency.
These are cross-call controls. You can't reliably measure them from a single conversation transcript.
Which Borrower Experience KPIs Matter?
A collections call can meet its immediate business objective while still producing a poor borrower experience.
For NBFCs, borrower experience matters because complaints, unresolved disputes, misunderstandings, and inappropriate handling of financial hardship can create long-term problems.
| KPI | Definition | How to Measure It |
|---|
| Sentiment trend | Change in borrower sentiment from the start to the end of a call | Per-turn sentiment and frustration analysis, with appropriate validation |
| Resolution | Calls ending with a clear, agreed next step | Check for a promise, plan, referral, callback, or recorded dispute |
| Repeat contact | Borrowers contacted again within a defined period, such as seven days, about the same issue | Join call logs by account and issue |
| Human requests | Calls where the borrower asked to speak with a person | Detect the request and check the escalation outcome |
| Vulnerability flags | Calls containing possible vulnerability disclosures or indicators | Detect relevant cues and route them through approved review procedures |
| Complaint resolution | Complaints that receive appropriate follow-up and resolution | Connect conversation records with the complaint management system |
Sentiment Trend
Changes in borrower sentiment may reveal rising frustration, confusion, or dissatisfaction.
However, sentiment is not a definitive measure of borrower intent, financial vulnerability, or whether a conversation was handled appropriately.
Use it alongside observed conversation behavior, complaint indicators, and documented outcomes.
Resolution and Repeat Contact
A collections call should end with a clear next step.
That might be a repayment promise, payment plan, hardship referral, callback, or dispute record.
Repeat contact provides useful context. If an AI agent marks a conversation as resolved but the borrower repeatedly contacts the NBFC about the same issue, the resolution KPI may overstate success.
Human Requests and Vulnerability Flags
Borrowers may request human assistance because they are confused, distressed, disputing the debt, or experiencing financial hardship.
Monitor whether the request was recognized, the handoff worked, and the human agent received relevant context.
The FCA's 2025 review of vulnerable customer outcomes highlighted limitations in relying solely on process-oriented call metrics.
Resolution, repeat contact, complaints, and appropriate escalation provide additional context on outcomes.
For NBFCs, vulnerability indicators should support appropriate assistance and human review, not automatically trigger adverse treatment.
Which AI-Specific KPIs Should NBFCs Track?
Traditional collections dashboards were not designed to explain every technical failure of an AI voice agent.
An AI agent can retrieve the wrong amount, delay its response, misunderstand an accent, repeat a question, or fail during a transfer.
These failures may not be visible in repayment reports until they affect performance.
| KPI | Definition | How to Measure It on Live Calls |
|---|
| Response latency | Time from the borrower finishing a turn to the agent starting to respond | Measure per turn and track median and tail latency |
| Dead air | Silences above a defined threshold, such as three seconds | Detect silence events from call audio or timing data |
| Interruptions | Cases where the agent talks over the borrower or fails to handle interruptions appropriately | Analyze overlapping speech and turn-taking |
| Invented or mismatched amounts | Amounts stated by the agent that do not match the authorized record or tool output | Compare stated values with account data |
| Escalation to human | Calls transferred to a human and whether the handoff succeeded | Combine conversation analysis with telephony and transfer records |
| Task completion | Calls where the configured objective was completed without violating defined rules | Evaluate against an approved scorecard |
| Judge-human agreement | Agreement between automated quality evaluations and human reviewers | Compare judgments on a human-reviewed sample |
| Transcript accuracy | Accuracy of recognized speech, especially financial amounts, names, and dates | Compare transcripts with sampled audio references |
| Tool and API failures | Failed account lookups, payment-data retrievals, or backend actions | Monitor tool calls, errors, and traces |
Response Latency and Dead Air
Latency matters because people expect conversations to flow naturally.
A cross-language study by Stivers and colleagues (2009) found that conversational turn-taking often involves relatively short gaps, although timing varies across languages and contexts.
Long pauses can prompt borrowers to repeat themselves, interrupt the agent, or assume the call has disconnected.
In collections conversations, these interruptions can also disrupt identity verification, repayment read-back, and confirmation.
Monitor both typical response times and unusually slow turns rather than relying only on average latency.
Invented or Mismatched Repayment Amounts
Incorrect repayment amounts are among the more consequential failure modes for AI collections agents.
An agent may state an amount not in the authorized loan record or produce an unsupported figure when account information is missing.
In internal testing with AI callers interacting with agents built for India, an empty amount field produced invented figures in 15 of 18 calls. Another observed failure involved an agent disclosing a borrower's loan to a relative.
These examples show why you should check AI agents' financial statements against authorized records rather than accept them solely because the conversation sounded confident.
The sample is small and scenario-specific. It should not be interpreted as an industry-wide failure rate.
Human Escalation
Escalation rate alone does not determine agent quality.
A low escalation rate might indicate that the AI agent handles conversations effectively. It might also indicate that the agent fails to transfer borrowers who explicitly request human assistance.
Track human requests, required escalation triggers, attempted transfers, successful transfers, and incomplete handoffs together.
Judge-Human Agreement
Judge-human agreement tells you how much confidence to place in automated conversation scoring.
Research on LLM-based evaluation shows that strong models can agree substantially with human preferences, but automated judges can also show position and verbosity biases.
For collections observability, evaluate agreement separately for each criterion.
An AI judge might perform reliably on a narrow question, such as whether the agent identified the organization it represented, but less consistently when evaluating whether hardship handling was appropriate.
A practical review process should compare automated judgments against human assessments, investigate disagreements, and refine evaluation criteria.
How Should NBFCs Set KPI Thresholds and Alerts?
Monitoring becomes operationally valuable when teams can identify changes that require attention.
Three principles help structure observability alerts for collections.
1. Zero-Tolerance KPIs Should Trigger Critical Review
Treat threats, unsupported legal claims, and unauthorized third-party disclosures as critical failures under an appropriately configured monitoring policy.
A detected event can mark the call for review and generate a timely alert.
Because automated classification can produce false positives or miss context, critical findings should support defined escalation and human investigation procedures.
2. Rate KPIs Should Warn on Meaningful Changes
Monitor response latency, dead air, sentiment, escalation rate, and other performance indicators for movement outside established operating ranges.
Alerts are especially useful after a prompt update, model change, telephony configuration change, or backend integration release.
For example, a sudden increase in dead air after deployment could point to a latency regression rather than a borrower-behavior change.
3. Business KPIs Need Pace Alerts
Track promises to pay and payments collected against daily or weekly targets.
When performance falls below expectations, the NBFC should be able to investigate whether the change is associated with contact rates, portfolio composition, agent behavior, or technical failures.
Vattara Observe's product framework includes call-level scoring, pass/warn/block classifications, metric-based alerts, and business objective tracking. The original product specification describes scores from 0 to 100, approximately 28 monitored metrics, Slack/email/Jira notifications, and pace alerts for objectives such as payments collected and promises to pay.
These capabilities let teams bring conversation quality and operational performance into the same monitoring workflow, depending on the features enabled in their deployment.
