AI in Collections After RBI's 2027 Rules

Collections is one of the first places lenders want to use AI agents. The volume is high, the calls are repetitive, and the cost of a missed follow-up is easy to measure. From 1 January 2027, RBI's revised conduct rules on loan recovery will also apply to every call, message and visit, whoever or whatever makes it.

My view is that collections is one of the hardest tests of autonomy in lending, and the reason is not the technology. A spoken sentence cannot be taken back. When an AI agent says the wrong thing to a borrower who has just lost a parent, there is no undo button. That changes how the system has to be designed.

This piece is for CTOs, heads of collections and risk leaders at NBFCs and banks. It sets out what the new rules mean for technology, applies a simple readiness test to a collection call, and shows where AI can help safely first.

The short version

RBI's recovery conduct amendments, issued in August 2026 and effective from January 2027, make collections a governed, recorded and audited activity. An AI agent that contacts borrowers inherits every one of those duties. Because a conversation cannot be reversed, autonomy in collections needs controls before the action, not after it.

• Map every recovery rule to a system control, not a script line.
• Score your collection process with Autonomy Readiness before deploying agents.
• Enforce contact windows in the dialler, not in the prompt.
• Detect disputes, hardship and bereavement early and hand off.
• Record, store and retrieve every AI conversation like a human one.
• Start with assistance and quality review before autonomous calling.

This article sits inside the ai and enterprise solutions architecture cluster, where the wider argument is set out in full.

What the recovery rules change for technology

On 6 August 2026, RBI issued amendments to its responsible business conduct directions for NBFCs, alongside parallel versions for banks and other lenders, all effective from 1 January 2027. Published analyses of the NBFC text describe a board-approved collection and recovery policy, due diligence on recovery agencies, certification for recovery agents, and closer control over how and when borrowers are contacted. Reports on the package also describe a contact window of 8 a.m. to 7 p.m., call recording with six-month retention, and tight limits on device-locking technology.

Read those obligations as a technology architect and most of them become system requirements. A contact window is a dialler rule. Recording and retention are storage and retrieval requirements. A policy on conduct is a set of constraints any automated agent must follow and prove it followed.

Check the final RBI text and your compliance team's reading before building anything. This article is general information, not legal advice.

Why collections is a hard test for autonomy

Most lending processes have an undo path. A wrong posting can be reversed. A wrong limit can be corrected. A wrong sentence to a distressed borrower cannot.

Collections also carries more variation than it looks. The script covers a borrower who forgot a due date. Real calls include borrowers disputing the amount, reporting a death in the family, claiming fraud, asking not to be called, losing a job, or speaking a language the script does not cover. In most operations today, the human agent handles those moments without anyone writing a rule for them.

That is the point I keep returning to with automation. In most enterprise processes, humans were not "in the loop". They were the loop, quietly absorbing the variation nobody documented. A copilot keeps that buffer in place. An autonomous agent removes it, and whatever the human was absorbing becomes the agent's problem.

Autonomy Readiness applied to a collection call

I use a four-factor test for whether a process can safely run without a human:

[Autonomy Readiness](/frameworks/autonomy-readiness) = Specification × Exception Coverage × Observability × Reversibility

The factors multiply. A zero on any one makes autonomy unsafe, not merely imperfect. Applied to an outbound collection call:

• Specification. Are the rules written precisely enough for a machine: contact windows, frequency limits, permitted language, what may and may not be said about consequences? Most collection policies are written for trained people, not systems.
• Exception coverage. Does the design detect and route disputes, hardship, bereavement, fraud claims and requests to stop contact? This is where most pilots are thin.
• Observability. Is every conversation recorded, transcribed, stored for the required period and searchable by borrower and date? Could you show a supervisor exactly what the agent said?
• Reversibility. Close to zero for a spoken sentence. That is the factor you cannot fix after the fact.

Because reversibility is inherently low, the other three have to carry more weight, and controls must act before the agent speaks: eligibility checks before dialling, hard-coded time windows, and an early handoff when the conversation leaves the specified path.

A second measure helps size the risk. The Variant Coverage Ratio is the share of actual call volume covered by paths the agent has been explicitly designed to handle. If your transcripts show that a large share of calls include a dispute or hardship signal, and the agent handles neither, its real coverage is far lower than the pilot suggested.

Should your organisation do this now?

• Yes, if you have recorded, transcribed calls to measure real variation.
• Not yet, if your collection policy exists only as training material.
• Instead, first: rewrite the policy as testable rules with owners.
• Measure it by: Variant Coverage Ratio on a sample of recent calls.

Where AI helps first

Autonomous calling is the most visible use of AI in collections and the one with the least room for error. There are safer places to start that still move the numbers.

Quality review. If every call is recorded, AI can review far more of them than a human quality team, flagging tone, off-script statements and missed hardship signals. This improves conduct across human and machine agents.

Agent assistance. A copilot that suggests the next step, surfaces the borrower's history and warns about contact limits keeps a person responsible for what is said.

Scheduling and eligibility. Deciding who may be contacted, when, and through which channel is rule-driven and reversible before any contact happens. It is also where contact-window violations are prevented.

Narrow autonomous flows. Reminders before a due date, to borrowers with no dispute or hardship flags, within the permitted window, with an immediate route to a human. Expand from there only as coverage is measured.

Should your organisation do this now?

• Yes, if you can already enforce contact rules in the dialling platform.
• Not yet, if recordings are incomplete or hard to retrieve.
• Instead, first: fix recording coverage and retrieval.
• Measure it by: share of contacts recorded and retrievable within a day.

Before you approve it

Checklist:

• Recovery rules mapped to named system controls.
• Autonomy Readiness scored for each proposed AI use.
• Contact windows enforced in the platform, not in prompts.
• Dispute, hardship, bereavement and stop-contact detection tested.
• Human handoff available within the same conversation.
• Recording, retention and retrieval for AI conversations.
• Board-approved policy updated to cover automated contact.

Questions to ask:

• Your team: what share of last month's calls contained a dispute or hardship signal?
• Your team: can the agent contact a borrower outside the permitted window under any configuration?
• Your vendor: where are recordings and transcripts stored, and for how long?
• Your vendor: how does the agent detect that it has left its specified path?
• Your compliance head: does our recovery policy cover automated agents explicitly?
• Your board: who is accountable if an AI agent breaches conduct rules?

How to measure it

• Variant Coverage Ratio. Share of call volume the agent is designed to handle. Baseline: sample of recent recorded calls. Owner: head of collections. Review: monthly. Leading.
• Out-of-window attempts. Contacts attempted outside permitted hours. Baseline: current dialler logs. Owner: collections technology lead. Review: weekly. Leading.
• Handoff rate. Share of AI conversations handed to a person. Baseline: pilot. Owner: collections operations. Review: weekly. Leading.
• Missed exception rate. Hardship or dispute signals the agent failed to route, found in QA samples. Baseline: pilot QA. Owner: quality lead. Review: monthly. Leading.
• Recording retrievability. Share of contacts retrievable within one working day. Baseline: current. Owner: IT. Review: monthly. Leading.
• Complaints per thousand contacts. Conduct outcome. Baseline: last two quarters. Owner: grievance officer. Review: monthly. Lagging.

Mistakes that cost the most

Putting compliance rules in the prompt

A model instruction is not a control.

• Why it happens: it is the fastest way to ship a pilot.
• Prevention: enforce time windows and eligibility in the platform.
• Early warning: any contact logged outside permitted hours.

Measuring pilots on easy calls

Pilot samples often exclude difficult borrowers.

• Why it happens: early success is rewarded.
• Prevention: measure coverage on a random sample of real calls.
• Early warning: performance drops sharply at scale.

No exit from the script

The agent keeps pressing when the borrower reports hardship.

• Why it happens: exception detection was not specified.
• Prevention: test detection and handoff before launch.
• Early warning: complaints mentioning insensitivity.

Treating AI calls differently from human calls

Recording and retention gaps appear.

• Why it happens: AI conversations are stored by the vendor.
• Prevention: same retention and retrieval standard for both.
• Early warning: transcripts you cannot produce on request.

Unclear accountability

Nobody owns conduct outcomes for automated contact.

• Why it happens: the project sits between IT and collections.
• Prevention: a named business owner accountable to the board.
• Early warning: conduct issues escalated to the vendor first.

Frequently asked questions

When do RBI's new recovery rules apply?

The amendments to the responsible business conduct directions were issued on 6 August 2026 and take effect from 1 January 2027. Existing instructions apply until then.

Do the rules apply to AI agents?

The rules govern how lenders and their agents recover dues. An AI system contacting borrowers on a lender's behalf is part of that recovery process, so treat it as bound by the same conduct requirements. Confirm with compliance.

What contact hours apply?

Reports on the new directions describe a permitted window of 8 a.m. to 7 p.m., with limited exceptions. Build the window into the dialling platform and check the final text for exceptions.

Must collection calls be recorded?

Analyses of the directions describe call recording with a six-month retention period. Apply the same standard to AI conversations and confirm the requirement against the RBI text.

What is Autonomy Readiness?

A four-factor test for whether a process can safely run without a human: Specification × Exception Coverage × Observability × Reversibility. A zero on any factor makes autonomy unsafe.

What is the Variant Coverage Ratio?

The share of a process's actual transaction volume covered by the paths an automation has been explicitly specified to handle. It shows how much of real work the automation can actually do.

Is it safer to start with an AI copilot?

Usually, yes. A copilot keeps a person responsible for what is said while still improving consistency and speed, and it produces the data needed to judge later autonomy.

Can AI help with collections compliance itself?

Yes. Reviewing recorded calls for off-script statements, missed hardship signals and contact-rule breaches is a strong early use, because it improves conduct without adding new contact risk.

What to do next

Pull a random sample of last month's recorded calls and measure how many contain a dispute, hardship or stop-contact signal. That number sets the real ceiling for automation. Then read the Autonomy Readiness framework and the enterprise AI architecture topic page.

Sources

• Corporate Professionals, When the recovery call comes: RBI rewrites the rules of engagement for NBFCs, August 2026, analysing RBI notification DOR.MCS.REC.No.199, 6 August 2026. corporateprofessionals.com
• Tax Update India, RBI's new loan recovery rules from January 1, 2027, August 2026. taxupdate.in
• Agrud Partners, RBI loan recovery rules 2026: business conduct directions, 2026. agrudpartners.com

Last reviewed: 15 September 2026.

Views are my own and do not represent my employer.

This article is general information, not legal advice.