AI agents for NBFCs, lenders and fintech operations
The short answer
Lending operations break down in the file, not the decision. Branemind builds agents that assemble and ground the credit file, tier it for review, and work collections inside consent and contact policy, while the credit decision and every reserved judgement stay with a named person.
Who we usually work with here
- Head of Credit
- Accountable for decision quality and turnaround.
- Collections Head
- Accountable for recovery rate and conduct risk.
- CTO or COO
- Accountable for systems, integration and operating cost.
- Compliance
- Accountable for auditability and customer conduct.
Where it actually breaks
Turnaround is spent reading, not deciding
Statements, financials, GST returns and KYC packs arrive as scans of uneven quality. Most of the elapsed time on a file is assembly and cross-checking, which is exactly the part that does not need a credit officer.
Two similar files get different scrutiny
Without a consistent assembly step, depth of review depends on who picked the file up and how busy they were.
Collections is either thin or indefensible
Manual outreach does not reach enough borrowers, and volume outreach without consent and frequency controls creates conduct exposure that outlasts the recovery.
The file cannot be reconstructed later
When an examiner asks why a decision was made eight months ago, the answer has to be reproducible, not remembered.
Constraints that shape every build here
- The credit decision must remain attributable to a named person
- Consent state, permitted channels, contact frequency and permitted hours are hard constraints
- Evidence has to survive an examination months after the fact
- Data residency and provider choice often need to be constrained before build
Where we would start
In order of how often it pays off, not in order of what is most interesting to build.
Underwriting file assembly
The highest-value first step, because it compresses the part of turnaround that is pure assembly without touching the decision itself.
Fintech AI AgentsCollections on WhatsApp
Consent-checked, policy-bounded outreach with human escalation on hardship signals. Measured on contact rate and dispute rate together, never contact rate alone.
WhatsApp Business AgentsFinance operations behind the lending book
Reconciliation of collections, disbursements and settlements, which is the same continuous-close problem in a regulated wrapper.
Finance Operations AutomationControls we hold ourselves to
- The credit decision is a human action, logged against a named officer
- Every claim in a summary cites the page it came from
- Consent, channel, frequency and hour checks block a send rather than reporting on it
- Model version, prompt version, evidence and human action retained together
Published work in this industry
Case study
An underwriting copilot that took decisions from days to minutes
A lending team was drowning in manual file review. We put an agent in the loop that reads every source, argues its own case, and hands the underwriter a decision worth signing.
- Median decision time
- 38s
- from 2.4 days
- Files auto-tiered
- 71%
- no analyst touch
- Override rate
- 4.1%
- underwriter disagrees
Case study
Collections on WhatsApp, run end-to-end by an agent
Recovery calls nobody answers, replaced by a conversation people actually finish. BM Relay negotiates, takes the payment, and knows exactly when to stop.
- Contact rate
- 94%
- vs 31% on voice dialer
- Promise-to-pay kept
- +38%
- Cost per recovery
- −62%
Questions we get asked
Can an AI agent approve a loan?
It should not. In our builds the agent assembles and grounds the file and recommends a review tier, and a credit officer decides. That is partly a controls position and partly practical: the audit trail has to name a person.
How do you keep collections outreach compliant?
Consent state, allowed channels, frequency caps and permitted contact hours are enforced in the send path as blocking checks. Dispute and hardship signals stop automation and route to a person immediately.