Collections and Recovery: How to Build It

How to use this page

This page walks you through building one thing: the system that contacts a borrower who has missed a payment, and decides what to say, when, and how many times.

Every other product guide on this site regulates a transaction. This one regulates a conversation. The rules govern the hour you may call, the words you may use, who you may mention it to, and what evidence you must keep — and they were written for a human agent at a time when the agent is increasingly a machine.

What you are actually building

Collections is a sequence of decisions, and the regulated part is the contact rather than the decision. What you are building:

What this is not:

• Not a dialler project. The hard part is deciding who not to call and when to stop.
• Not a place to deploy autonomy. See step 5.
• Not device locking. That category shrank sharply in August. Step 7.

The whole journey, in one table

Steps 1 to 3 — segment, gate, script

Step 1 — Segment by reason, not by days past due

Almost every collections system buckets accounts by how late they are. Days past due is a measure of how long, and tells you nothing about why — which is the only thing that determines what works.

The first bucket is usually the largest and the cheapest to fix, and treating it with the same intensity as the third is how collections programmes generate complaints while recovering money they would have recovered anyway.

Step 2 — Contact rules belong in code

A policy document saying we call between 8 and 7 is not a control. The control is a function that refuses to place the call.

What has to be enforced, not documented: the 8 AM to 7 PM window across every channel, in the borrower's local time; a frequency cap per day and per week; a rule that a promise to pay suppresses contact until the promised date; that a registered dispute suppresses collections contact entirely; and that nobody other than the borrower and a guarantor is contacted about the debt.

Step 3 — What gets said, and what may never be said

Script the permitted messages, and script them narrowly. More usefully, enumerate the prohibited ones, because that list is short, specific and the one that generates complaints:

• Any threat, including implied — arrest, criminal proceedings, seizure without an order, damage to a person's standing.
• Abusive, obscene or intimidating language.
• Disclosing the debt to anyone else, including by leaving a message with a family member or writing it on a visible envelope.
• Misrepresenting who you are or what authority you hold.
• Claiming legal consequences that do not exist. For an unsecured loan, no asset is seized without a court order.

Steps 2 to 5 — the gate, and the machine that speaks

Steps 2 to 5 — The gate, and the machine that speaks

from datetime import datetime, time, timedelta

# STEP 2. THE GATE. Nothing reaches a borrower except through this.
WINDOW = (time(8, 0), time(19, 0))      # Fair Practices Code: 8 AM to 7 PM

def may_contact(account, now_local, channel, history, flags):
    """Returns (allowed, reason). A policy document does not stop a call.
    This does."""
    if not (WINDOW[0] <= now_local.time() < WINDOW[1]):
        return False, "outside_permitted_hours"        # every channel, not just voice
    if flags["dispute_registered"]:
        return False, "dispute_open"
    if flags["promise_to_pay_until"] and now_local.date() <= flags["promise_to_pay_until"]:
        return False, "promise_to_pay_active"
    if history.contacts_today(account) >= flags["daily_cap"]:
        return False, "daily_cap_reached"
    if history.contacts_this_week(account) >= flags["weekly_cap"]:
        return False, "weekly_cap_reached"
    if flags["vulnerability_hold"]:
        return False, "vulnerability_hold"             # see step 5
    return True, "ok"

def permitted_recipients(account):
    # Nobody else. Not a relative, an employer, a neighbour or a colleague.
    return [account["borrower"]] + account.get("guarantors", [])

# STEP 5. IF A MACHINE DOES THE SPEAKING.
# There is no settled rulebook for an automated voice agent in collections.
# The conduct rules were written for a person, they apply to the lender
# regardless of who or what makes the call, and the gap is yours to fill.

AGENT_CONSTRAINTS = {
    "identifies_as_automated_up_front": True,
    "states_the_lender_by_name": True,
    "route_to_human_offered_every_turn": True,
    "may_negotiate": False,          # it presents options; it does not deal
    "may_threaten": False,           # not expressible, not merely disallowed
    "max_turns_before_human": 6,
    "recording_disclosed_before_it_starts": True,
}

# The thing an automated caller cannot do is notice.
DISTRESS_MARKERS = ("hospital", "died", "passed away", "lost my job",
                    "cannot cope", "harassing me", "legal notice")

def supervise(turn, agent_constraints):
    if any(m in turn["transcript"].lower() for m in DISTRESS_MARKERS):
        # Not a branch in the script. Leave the automation entirely.
        return {"action": "hand_to_human_now", "suppress_further_automation": True,
                "set_vulnerability_hold": True}
    if turn["index"] >= agent_constraints["max_turns_before_human"]:
        return {"action": "offer_human"}
    return {"action": "continue"}

# WHAT TO CHECK
# [ ] the gate is the ONLY path to a borrower. A campaign tool that can
#     bypass it is the whole control gone
# [ ] the hour test uses the BORROWER's local time, not your server's
# [ ] the window applies to SMS and messaging apps too. An automated message
#     at 10pm is the same violation as a call
# [ ] a promise to pay suppresses contact. Calling someone who has already
#     told you when they will pay is the most common avoidable complaint
# [ ] distress markers exit the automation rather than branch inside it
# [ ] the agent cannot express a threat. Not "is instructed not to" -- the
#     phrasing is not in its permitted output at all
# [ ] every suppression is logged with its reason. "We did not call" is
#     evidence only if it is recorded

THE finding, and it is the reason this page is different from the other fifteen: the conduct rules describe how a person should behave, they bind the lender regardless of who or what places the call, and nothing yet describes how an automated caller should behave. The gap is yours to fill and it will be filled later, by somebody else, retrospectively.

An automated dialler complies with the 8 AM to 7 PM window trivially — better than a human team does. What it cannot do is notice. A person hearing "my father is in hospital" stops. A script hears an unmatched intent and asks about the payment again, and the second question is the one that becomes a complaint, a screenshot and an Ombudsman case.

Two further positions worth taking before anyone requires them. Say it is automated, up front. There is no Indian rule presently compelling that disclosure in collections, and a borrower who discovers mid-call that they have been arguing with a machine about their debt is a complaint you created. And never let it negotiate. It may present options your policy has already approved; it may not agree terms, because an agreement reached with a machine is an argument about what was agreed.

Steps 6 to 8 — resolve, escalate, answer for it

Step 6 — Offer a way out

The purpose of collections is recovery, and the highest-recovery action for the cannot pay segment is almost always a restructure rather than pressure. Build the options into the first contact rather than holding them back as a concession: a revised schedule, a short pause, a part-payment arrangement, a settlement where the account warrants it.

Make them self-serve. A borrower who can restructure at eleven at night without speaking to anyone will do it. The same borrower, asked to call during office hours to discuss it, frequently does not.

from datetime import date, timedelta

# STEP 6. THE WAY OUT, OFFERED EARLY RATHER THAN CONCEDED LATE.
def resolution_options(account, policy):
    """Present what your policy already approves. Do not hold options back as
    a negotiating position -- the segment that needs them is the segment that
    does not respond to pressure."""
    opts = []
    if policy["allow_reschedule"]:
        opts.append({"type": "reschedule", "self_serve": True})
    if account["hardship_flag"] and policy["allow_pause"]:
        opts.append({"type": "pause", "months": policy["max_pause_months"],
                     "self_serve": True})
    if account["days_past_due"] > policy["settlement_after_days"]:
        opts.append({"type": "settlement", "self_serve": False})   # needs approval
    # Self-serve at 11pm converts. The same option behind a daytime phone
    # call frequently does not.
    return opts

# STEP 7. ESCALATION. EACH STEP HAS A PRECONDITION, NOT A DAY COUNT.
def may_escalate(account, step, agents, now):
    if step == "field_visit":
        # Fourth Amendment Directions, from 1 Jan 2027: notice first.
        return account["visit_notice_sent_on"] is not None and \
               now.date() >= account["visit_notice_sent_on"] + timedelta(days=1)
    if step == "assign_agency":
        # Certified, background-verified, publicly listed, borrower informed.
        a = agents.assigned(account)
        return all([a["iibf_certified"], a["background_verified"],
                    a["published_on_website"], account["borrower_informed_in_writing"]])
    if step == "restrict_device":
        # Baseline is prohibition. Only a device this loan financed.
        return account["device_was_financed_by_this_loan"] and \
               account["days_past_due"] >= policy_min_days() and \
               account["graduated_restriction_only"]
    return False

# STEP 8. THE EVIDENCE, AND THE CLOCK THAT RUNS WITHOUT YOU.
def complaint_sla(received_on):
    """RB-IOS 2026: no reply within 30 days, or an unsatisfactory one, and the
    borrower may go to the Ombudsman. Set your own clock shorter."""
    return {"internal_target": received_on + timedelta(days=14),
            "ombudsman_eligible_from": received_on + timedelta(days=30)}

def evidence_pack(account, store):
    # What a complaint is answered with. Assemble it now, not later.
    return {
        "contact_log": store.contacts(account),        # incl. SUPPRESSED, with reason
        "recordings": store.recordings(account),       # prior intimation given
        "transcripts": store.transcripts(account),
        "who_called": store.actor(account),            # human or agent, named
        "options_offered": store.offers(account),
        "notices_sent": store.notices(account),
    }

# WHAT TO CHECK
# [ ] report complaint AGE, not complaint volume. The oldest open complaint
#     predicts escalation; the count does not
# [ ] log suppressed contacts with their reason. "We did not call" is
#     evidence only if it was recorded at the time
# [ ] every escalation step checks a PRECONDITION, not a day counter
# [ ] the actor on every contact is identifiable -- which human, or which
#     automated agent and which version of it
# [ ] test retrieval of a recording from six months ago before you need one
# [ ] run the SLA clock in IST against calendar days, not working days. The
#     Ombudsman's 30 days do not pause for your weekend

Step 7 — Escalation, and what changed in August

The Fourth Amendment Directions of 6 August 2026, in force from 1 January 2027 and issued under sections 21 and 35A of the Banking Regulation Act, convert a scattered set of expectations into a detailed and enforceable code. What they add:

Step 8 — Record it, and answer for it

Two clocks matter. Internally, a complaint must be capable of being answered with what was actually said — which means the recording, the transcript, the contact log with its suppression reasons, and the identity of whoever or whatever made the call.

Externally, the Reserve Bank Integrated Ombudsman Scheme, 2026, effective 1 July 2026, sets the path: the complainant approaches the regulated entity first, and may escalate to the Ombudsman if there is no reply within 30 days or the reply is unsatisfactory, within 90 days of that period expiring.

What it costs

Three versions you could build

Reminders and self-serve

Build: failed-mandate detection → a reminder inside permitted hours → a one-tap payment link → self-serve restructure.

You get: most of the recoverable balance, almost no conduct exposure, and no recovery agents. For many lenders this is the whole product, and the reason it is not built first is that it does not look like collections.

Segmented treatment with human calling

Build: the above, plus segmentation by reason → the contact gate enforcing hours, caps, promises and disputes → scripted treatments with a prohibited list → human calling for the cannot pay segment → complaints to a named owner.

Trade: people cost money and exercise judgement. The judgement is what you are buying, not the throughput.

Automated voice at scale

Build: the above, plus an automated caller that identifies itself, cannot express a threat, cannot negotiate, offers a human every turn, and exits entirely on any distress marker — with transcript supervision and a human handover capacity sized for the exits.

It breaks when: the exit is built as a branch inside the conversation rather than an exit from it. A machine that acknowledges a bereavement and then asks about the payment again has produced the worst available outcome, and it will be a screenshot before it is a metric.

What goes wrong

Where to go next

Sources

Every figure, rule and date on this page, and where to check it. Entries are typed so you can see which are primary-sourced and which are industry reporting.