The EMI collections playbook for AI voice agents

Collections is a numbers game played on a human field. A lending book of any size generates thousands of EMIs due every month, and a predictable fraction of borrowers will miss a date, forget, run short, or genuinely struggle. The traditional response is a tele-calling floor: rows of agents dialing overdue accounts, reading roughly the same script, capturing promises on sticky notes and spreadsheets, and burning out on the monotony. It works, but it does not scale cleanly, quality drifts across agents and shifts, and compliance is only as good as the least careful caller on a bad day.

AI voice agents change the shape of this problem. They make the early, high-volume, low-emotion reminders consistent and cheap, capture promises as structured data instead of scribbles, send payment links on the same call, and enforce calling-hour and do-not-call rules by design rather than by hoping. Crucially, they free your human collectors to focus on the accounts where empathy and negotiation genuinely move the needle: hardship, disputes, and deep delinquency.

This is a practical playbook for running EMI collections with AI voice agents in the Indian BFSI context. It covers bucket-wise strategy, how to design persona and intensity, the mechanics of capturing a promise-to-pay and taking payment on the same call, compliance built in from the start, reconciliation, the metrics that actually matter, and a staged rollout so you can prove it before you scale it.

Think in buckets, not in a single script

The single biggest mistake in collections automation is treating every overdue account the same. A borrower who is two days from their due date and a borrower who is fifty days past due are in completely different psychological and financial situations. The tone, the objective, and the intensity of the call must change with the bucket. Design a distinct approach for each stage rather than one script with the amount swapped in.

Pre-due (before the due date)

The pre-due call is not collections, it is service. The goal is a friendly reminder that the EMI is coming up, confirmation that the borrower is aware, and a gentle offer to make paying easy. Tone is warm and helpful. Many missed payments are pure forgetfulness, and a well-timed pre-due nudge, two to three days ahead, prevents a large share of them from ever becoming overdue. This is the cheapest rupee you will ever collect because you spend it before there is a problem.

1 to 30 days past due

Early delinquency is still mostly forgetfulness and minor cash-flow timing, not distress. The call acknowledges the missed date without accusation, confirms the borrower intends to pay, and captures a specific promise-to-pay date. Tone is polite but clear. The objective is a concrete commitment and, ideally, immediate payment via a link sent on the call. Most of your recovery volume lives here, which makes it the sweet spot for AI: high volume, low emotion, repetitive, and enormously sensitive to consistency and speed.

31 to 60 days past due

By this bucket, easy forgetfulness is largely ruled out and you are into genuine reasons: cash shortage, a dispute, a change in circumstances. The call is firmer and more diagnostic. The agent should probe why payment has not happened, distinguish inability from unwillingness, and either capture a realistic promise-to-pay or flag the account for human follow-up. Intensity rises, but professionalism and compliance must stay rock solid, because this is where complaints and regulatory risk begin to concentrate.

Hardship and deep delinquency

Some borrowers genuinely cannot pay in full right now: job loss, medical emergency, business downturn. Treating them like ordinary defaulters is both cruel and counterproductive, because it hardens them and raises complaint and regulatory risk. The AI agent's role here is largely to identify hardship quickly, respond with empathy, and route to a human who can discuss restructuring, part-payment, or a settlement. The agent should never improvise financial concessions; it should recognize the situation and hand off cleanly with full context.

Designing persona and intensity

Intensity is a dial you turn up across the buckets, but the persona should stay recognizably the same professional, respectful voice throughout. Borrowers should feel they are dealing with a consistent institution, not a friendly bot that suddenly turns aggressive. The escalation is in firmness and specificity, not in rudeness or pressure tactics, which are both bad practice and a compliance hazard.

• Pre-due: warm, helpful, service-oriented. "Just a friendly reminder that your EMI of a certain amount is due on the fifth. Would you like me to send a payment link?"
• 1 to 30 days: polite and clear. "Your EMI was due on the fifth and we have not received it yet. Can you tell me by when you will be able to pay?"
• 31 to 60 days: firm and diagnostic. "Your account is now over a month overdue. I want to understand what is preventing payment so we can find a way forward. Is it a timing issue or something more?"
• Hardship: empathetic and supportive. "It sounds like you are going through a difficult time. Let me connect you with a colleague who can discuss options that might help."

A well-designed agent also mirrors the borrower's language. In India that means fluent Hinglish and code-switching: a borrower who replies in Hindi should be met in Hindi, one who mixes should be mixed with. Nothing undermines a collections call faster than an agent that sounds foreign, robotic, or unable to understand a natural mixed-language sentence about money.

Capturing the promise-to-pay properly

The promise-to-pay, or PTP, is the core artifact of a collections call, and capturing it as loose free text is a wasted opportunity. A PTP should be structured data: a specific amount, a specific date, and ideally a stated method. "He said he'll pay soon" is useless. "Promise to pay the full EMI amount by the twelfth via UPI" is actionable, trackable, and can trigger an automatic reminder on the eleventh.

Train the agent to always pin down specifics. If the borrower says "I'll pay next week," the agent should push gently for a date: "Can you give me a specific day, so I do not need to disturb you again before then?" Vague promises are where recovery leaks. A concrete, structured PTP written straight into the collections system lets you run a disciplined kept-versus-broken PTP process, which is one of the most powerful levers in all of collections.

• Always capture amount, date, and method as structured fields, not free text.
• Push politely for specificity: a named date beats "soon" every time.
• Write the PTP straight to the system of record so it can trigger a pre-PTP-date reminder automatically.
• Track kept versus broken PTPs by borrower so repeat breakers get escalated, not endlessly re-promised.

Taking payment on the same call

The moment a borrower agrees to pay is the moment of maximum intent, and every hour of delay erodes it. The highest-performing collections flows capture that intent immediately by sending a payment link on the same call, via SMS or WhatsApp, while the borrower is still engaged. "Great, I am sending you a secure payment link on WhatsApp right now, you will get it in a few seconds, please tap it and complete the payment."

This on-call, same-channel payment link turns a promise into a transaction before the borrower moves on to the next thing in their day. It also creates a natural close: the agent can wait briefly, confirm receipt, and even confirm completion if the payment reflects quickly. Compared to a promise to pay "sometime," a link tapped while still on the phone converts dramatically better. The agent should never handle card or account numbers verbally; it sends a secure link and lets the borrower transact through a proper payment page.

• Send the link the instant intent is expressed, on the channel the borrower prefers, usually WhatsApp or SMS.
• Keep the borrower on the line briefly to confirm they received it and, if possible, completed it.
• Never take card or bank details by voice; always route payment through a secure link and page.
• Log whether the link was sent, opened, and paid, so you can measure link-to-payment conversion and improve it.

Compliance by design

In BFSI collections, compliance is not a feature you add later, it is the foundation the whole program stands on. A single pattern of calls outside permitted hours, or repeated calls to a borrower who asked to be left alone, can create regulatory exposure and reputational damage far larger than any recovery gain. The advantage of AI agents is that compliance can be enforced by the system rather than left to individual caller discipline.

• Calling-hour windows: enforce TRAI-style permitted calling hours automatically, so no account is dialed outside the allowed window regardless of campaign timing or time zone quirks.
• Do-not-call and opt-out suppression: honor DNC registrations and any borrower request to stop calling, and suppress those numbers across all campaigns immediately.
• Data redaction: redact sensitive fields in transcripts and recordings, and never expose full account or card numbers in logs.
• Audit trail: keep an immutable record of every contact attempt, what was said, what was promised, and what consent or opt-out was registered, so any dispute or audit can be answered with evidence.
• Frequency caps: limit how many times a borrower is contacted in a day or week to avoid harassment patterns, and stop entirely once payment is confirmed.

The last point deserves emphasis. An agent that keeps calling a borrower who has already paid is not just annoying, it is a compliance and brand risk. The collections system must know the moment a payment reflects and suppress further contact immediately. This is why tight reconciliation, covered next, is part of compliance, not just accounting.

Reconciliation and the feedback loop

Collections automation lives or dies on reconciliation: matching payments received back to the accounts that were called and the promises that were captured. Without it, you cannot tell which calls drove which payments, you keep calling people who have paid, and you cannot measure kept-PTP rates. With it, the whole system becomes a learning loop.

Set up reconciliation so that when a payment reflects, the account is marked paid, any scheduled follow-up call is cancelled, and the associated PTP is marked kept. Payments that do not arrive by the promised date mark the PTP broken and trigger the next appropriate action. This closed loop is what lets you compute the metrics that matter and continuously improve which buckets, scripts, and timings actually convert.

The metrics that matter

Vanity metrics like calls made or minutes consumed tell you nothing about whether the program works. Collections has a specific, well-understood set of metrics, and an AI program should be judged on exactly the same ones a human floor is.

• Connect rate: the share of dial attempts that reach the borrower. Low connect rates point to bad numbers, wrong timing, or poor number reputation, not agent quality.
• PTP rate: of connected calls, the share that produce a concrete promise-to-pay. Measures how well the agent secures commitment.
• Kept-PTP rate: of promises made, the share actually honored by the promised date. This is the real quality signal, because promises that break are worthless.
• Resolution or cure rate: the share of overdue accounts that become current within the cycle. The outcome the whole program exists to produce.
• Cost-to-collect: total program cost divided by the amount recovered. This is where AI usually shines, because the per-contact cost is a fraction of a human floor's.

Watch these by bucket, not just in aggregate, because they behave very differently across pre-due, early, and deep delinquency. A rising broken-PTP rate in the 31-to-60 bucket, for instance, is an early warning that your diagnostic script is accepting soft promises it should be probing harder, or that those accounts genuinely need human negotiation rather than another automated reminder.

A staged rollout plan

Do not flip your entire book to AI collections on day one. Roll out in stages that let you prove results and tune the system on low-risk buckets before trusting it with sensitive ones. The natural sequence mirrors the buckets themselves, from safest to most delicate.

• Stage 1, pre-due reminders: start here. Lowest risk, highest goodwill, purely a service call. Prove connect rates, language quality, and payment-link mechanics with almost no downside.
• Stage 2, 1 to 30 days: extend to early delinquency once pre-due is solid. This is the high-volume core where AI delivers most of its value. Tune the PTP capture and same-call payment flow here.
• Stage 3, 31 to 60 days with human handoff: add the firmer diagnostic bucket, but with a clear rule that hardship or dispute signals escalate to a human. Let AI do the reach and triage, humans do the negotiation.
• Stage 4, optimization and scale: with data flowing, refine scripts by bucket, adjust calling times for connect rate, tighten the models to the right cost stack, and scale volume. Keep humans focused on hardship and deep delinquency where empathy pays.

At each stage, run the AI cohort against a comparable human-handled control group and compare on kept-PTP and cost-to-collect, not on gut feel. This gives you the evidence to expand with confidence and the early warning to fix problems before they scale.

Takeaways

AI voice collections work best when you stop thinking in one script and start thinking in buckets, each with its own tone, objective, and intensity, from warm pre-due reminders to empathetic hardship handoffs. Capture promises as structured data, take payment on the same call with an instant link, and let the system enforce calling-hour and do-not-call compliance rather than trusting it to human discipline.

Judge the program on kept-PTP rate and cost-to-collect, close the loop with tight reconciliation so you never chase someone who has already paid, and roll it out in stages from safest bucket to most sensitive. Done this way, AI handles the high-volume, repetitive early reminders consistently and cheaply, while your human collectors do what they are actually good at: the difficult, human conversations at the deep end of the book.