Pre-Delinquency signals were lagging indicators.
By the time a portfolio dashboard refreshed, half of the at-risk accounts had already rolled into 30 DPD.
−22%
Personal loan NPA roll-rate (4.2% to 3.3%) over six months
View outcome below₹14crCollections previously slipping past 90 DPD, now recovered in 30–60 DPD
View outcome below1.9×Promise-to-pay rate on voice-led collections
View outcome below
- Client
- A multi-product mid-sized Indian NBFC
- Industry
- Diversified lending (personal, business, two-wheeler, MSME)
- AUM
- ₹3,200cr
- Branches
- 48
- Active borrowers
- 65,000
- Books covered
- Personal loans first, then MSME (in deployment now)
- Agents deployed
- Risk & Portfolio Ops Agents
- Deployment
- 12 weeks from kickoff to all workflows live
The situation
A strong committee, a monthly tempo, a daily book.
The risk function had three problems running in parallel: the early-warning signals were buried in MIS that landed three weeks late; the collections team was working alphabetical dialer queues with no recovery-probability ranking; and branch heads could not see their own P&L until day 22 of the following month.
The credit committee was strong, the team was experienced, but the operating tempo was monthly when the book was moving daily.
Any solution had to plug into the existing LMS and core systems. Nothing could leave the bank infrastructure.
Where the workflow broke down
Collections dialed alphabetically.
No model-ranked queue. Time was wasted on accounts that were either already lost or already going to pay anyway.
Branch P&L’s landed on day 22.
Branch heads were managing for the previous month while the current month was already half over.
Risk, collections, and branches worked off different versions of the truth.
Each team had its own MIS, refresh cadence, and definitions, so reviews became debates about whose number was right.
The deployment
Twelve weeks. On-premise. Read-only.
Twelve weeks from kickoff to all three workflows live. The platform runs on-premise; no data ever leaves the bank infrastructure.
Read-only across LMS, collections system, treasury, and GL. Engineering effort on the bank side: 22 hours.
Three workflows.
One connected platform.
Trigger
End-of-day batch, refreshed by 11 p.m.
Data sources
LMS, payment processor, behavioural signals (auto-debit bounces, partial pays), bureau pull (monthly refresh)
How it runs
Scores every active loan on probability of 30+ DPD in the next 60 days. Surfaces the three strongest leading indicators per at-risk loan. Segments the at-risk pool by product, region, vintage, sourcing channel.
Human in the loop
Credit head reviews the top 100 at-risk in the daily 9 a.m. routine. Collections head receives the daily roll-into-risk list.
Output
Risk cards in the credit Slack + daily PDF brief to the credit committee.
Before
Monthly portfolio review, three weeks lagged.
After
Daily risk lens with named leading indicators per account.
Trigger
Daily at 8 a.m., before the tele-collections team starts
Data sources
LMS, payment history, dialer outcomes, customer interaction history, bureau context
How it runs
Ranks all 30+ DPD accounts by predicted recovery probability × outstanding amount. Segments into four buckets: voice-callable, FOS visit, legal notice, routine reminder. Voice agent auto-dials the voice-callable segment, opens by identifying the bank, confirms the borrower, mentions the overdue EMI, attempts to secure a promise-to-pay, schedules a callback, or transfers to a live agent if escalated.
Human in the loop
Senior collections agents handle voice escalations and high-value accounts (>₹2L). Compliance reviews a 10% call sample weekly.
Output
Dialer queue pre-populated. LMS auto-updated with promise-to-pay. Call recordings retained for audit.
Before
Alphabetical dialer queue, random sequencing, ~12% promise-to-pay rate.
After
Ranked queue with voice agent on the medium tier, ~23% promise-to-pay rate. Senior agents now spend time only on conversations that need a human.
Trigger
Daily refresh
Data sources
GL, LMS, treasury, ops cost data
How it runs
Computes branch-level NIM, cost-to-income, GNPA, collections efficiency, and originations vs target. Refreshes daily, with comparison to prior month, quarter, and same period last year. Rolls up by region for zonal heads.
Human in the loop
Each branch head sees their own branch live. Regional heads see roll-ups. CFO sees consolidated.
Output
Web dashboard accessible inside the bank network. Monthly review meeting auto-prep deck.
Before
Day-22 MIS deck, three weeks lagged.
After
Live P&L from day one of the month. Branch heads run their own variance reviews on Monday morning.
Outcomes
−22%
NPA Roll-Rate on Personal Loans
4.2% to 3.3% over six months. Audited against pre-deployment baseline.
Before4.2%
After3.3%
- Period
- 6 months
Customer quote
“We finally see the book the way we always thought we did. The risk and collections teams stopped arguing about whose number was right because there is only one source now.”
Chief Risk Officer
From kickoff to production
Weeks 1–2
Scoping with credit, collections, and operations leadership. Three workflows and on-premise deployment confirmed.
Weeks 3–5
LMS and core systems read-only integration. On-premise install. Definition layer built with the credit risk team.
Weeks 6–7
Pre-delinquency early warning live.
Weeks 8–9
Collections workflow live, including voice agent (initial sample-audit period).
Week 12
Branch P&L live. All three workflows in production.
Common questions
Is the deployment really on-premise?
Yes. For regulated institutions Actioneer runs entirely inside the bank infrastructure. No data leaves the firewall.
How long does a deployment like this take?
Twelve weeks for this NBFC. Regulated on-premise deployments typically run ten to sixteen weeks; the variable is the number of source systems and the depth of the definition layer.
What controls govern the voice workflow?
A 10% call sample is reviewed weekly by compliance. The agent identifies itself, confirms identity, and never collects sensitive data over the call. Recordings are retained for audit.
Does the agent ever write to the LMS?
Only on scoped, audited paths. The collections workflow can write promise-to-pay back to the LMS. Everything else is read-only by default.
How are the numbers audited?
Against a pre-deployment baseline agreed with the credit committee. Every headline number ships with a traceable underlying query.
Can the same deployment cover MSME or two-wheeler books?
Yes. The context layer is product-aware. New books are scoped as incremental phases on top of the same platform.
