Automating collections is one of the most valuable things AI can do for an NBFC. It is also one of the most regulated. The May 2026 draft guidelines on recovery agents set precise, enforceable requirements. Here is what your AI collections programme must do to stay inside them.
1. Calling hours must be enforced by the system
RBI's conduct rules prohibit recovery contact before 8:00 am and after 7:00 pm. For a human agent, this is a training and monitoring problem. For an AI agent, this must be a system-level control — the AI must check the time before initiating any call or sending any message, and must refuse if the window has passed.
The check must happen at the moment of action, not at the point of scheduling. A call scheduled at 6:00 pm for 7:45 pm needs to be blocked when the window closes, not sent.
2. Content must be checked against your conduct policy
RBI prohibits threats, misleading statements, and discussing a borrower's debt with family members or friends. Every message and call script your AI sends or follows must be checked against your conduct policy before it goes out.
This means: - Pre-approved message templates only — not AI-generated content that hasn't been reviewed - Script compliance checks before each call - Flagging any deviation from approved content for human review
Calls should be recorded and reviewable for your compliance team and, where required, for RBI.
3. Complaints reach a human, fast
Any signal of distress or complaint from a borrower must escalate to a human immediately. This includes explicit complaints, but also signals like a borrower mentioning a hardship, requesting more time, or expressing frustration.
Your AI must be able to detect these signals and stop the automated engagement. A human takes over from that point. The handoff must be logged.
4. Human escalation for material decisions
Settlement offers, restructuring proposals, and waiver decisions are material. They must go through a human approval path. Your AI can present options, gather information and log a borrower's preferences — it cannot make the call.
Pre-approved settlement ranges are acceptable. The AI selects from within that range. Anything outside it requires explicit human sign-off before the offer goes to the borrower.
5. Every action on the record
The May 2026 draft is clear that lenders remain accountable for the actions of their AI systems. This means your audit trail must show:
- Who was contacted, when, and on which channel - What was said or sent - Whether the action was within approved parameters - What the outcome was
If RBI asks why a borrower received a message at a particular time, you must be able to show the answer in seconds, not days.