Financial KYC Automation
Automated KYC/AML workflow with AI-powered document verification and risk scoring.
Client
Digital lending platform
Industry
Financial Services (NBFC)
Region
South Asia
Duration
10 weeks build + regulatory approvals
A digital lending NBFC was losing 60% of applicants to drop-off because KYC took 2-5 days. We built an AI-powered onboarding pipeline - OCR + LLM document extraction, real-time eligibility scoring, conversational reapplication for failed verifications, and risk-prioritised review queues. Result: 5× faster onboarding (median 30 minutes instead of 2 days), 40% lower drop-off, and 70% of compliance reviewer capacity freed for the cases that need human judgement.
Headline results
The client is a digital NBFC offering personal and business loans across India, processing ~12,000 loan applications a month. Pre-automation, KYC and onboarding involved manual document review, multiple internal handoffs across credit and compliance, and a paper-trail audit log. The 2-5 day decision window was killing conversion - applicants would start with them, get delayed, get an instant offer from a competitor, and drop off. Compliance reviewers were drowning in low-risk applications, with no signal to focus on high-risk ones.
What we built and shipped
Vision + LLM document intelligence
Document upload triggers a vision + LLM pipeline that extracts identity fields, validates format, cross-references with our database, and surfaces tampering indicators - all in seconds, with regional-language IDs supported.
- Aadhaar, PAN, voter ID, passport, driving licence supported natively
- Regional language support: Hindi, Tamil, Telugu, Bengali, Marathi
- Tamper detection across image manipulation and template-based fraud
Real-time eligibility & risk scoring
Bureau data, internal signals, and ML-derived risk scores combine into a real-time decision that auto-approves clear cases, auto-rejects high-fraud-risk cases, and routes the middle band to compliance reviewers.
- Sub-30-second decision on green-path applications
- Composite score blending bureau, behavioural, and document-quality signals
- Explainable decisions with reason codes for every auto-rejection
Conversational reapplication AI
If documents fail verification, an AI agent walks the applicant through what to upload differently - instead of a generic error that causes drop-off. The agent uses simple language and shows good vs bad examples.
- Specific, actionable guidance per failure type
- Photo examples shown inline on WhatsApp or in-app
- Successful reapplication rate climbed from 12% to 58% post-launch
Risk-prioritised compliance review
Compliance reviewers see a queue sorted by risk score and complexity, with the AI-generated case summary pre-loaded - turning 30-minute reviews into 5-minute decisions on the cases that genuinely need a human.
- Queue auto-sorts by risk score, applicant LTV, and SLA pressure
- Pre-loaded case summary highlighting the specific risk indicators
- Reviewer time per case dropped from 30 minutes to 5 minutes on average
How it actually works
Document intelligence pipeline
Vision models for OCR and tamper detection feeding an LLM that interprets the extracted fields against KYC requirements. Confidence scores on every field - anything below threshold goes to reapplication, not rejection.
Decision engine
Deterministic policy rules wrap a probabilistic risk model - policy gates are non-negotiable, the ML score decides within the policy envelope. Every decision logs the rule set version and the input snapshot.
Audit-grade event log
Append-only event log of every read, write, and decision across the system - built for RBI inspection and quarterly audit reporting without manual evidence-gathering.
PII vault & access controls
Document images and extracted PII stored in an encrypted vault with role-based access. The decision engine sees scores and summaries; reviewers see full details only after access logging.
Phased delivery timeline
Regulatory + policy mapping
Worked with compliance to map every KYC rule, RBI master direction, and internal policy into explicit code. Built the policy engine that no ML can override.
Document pipeline + scoring build
Shipped document extraction, scoring, and decision engine in shadow mode - every decision compared against the existing manual decision on the same case. Refined until concordance crossed 97%.
Conversational reapplication + reviewer queue
Built the WhatsApp reapplication agent and the risk-prioritised reviewer queue. Reviewer team trained on the new dashboard and queue logic.
Regulatory sign-off + phased go-live
Regulatory approval after demonstrating audit-grade controls. Rolled out to 20% of applications, monitored daily, expanded to 100% over 4 weeks.
Before vs after
Same business, same team - measurably different operating model after the engagement.
- Median onboarding time2-5 days
- Application drop-off rate60%+
- Compliance reviewer time per case30 minutes
- Successful reapplication rate12%
- Regulatory audit prep time5-7 days per quarter
- First-pass approval rate on KYC65%
- Median onboarding time30 minutes (green path: 30 seconds)
- Application drop-off rateUnder 30%
- Compliance reviewer time per case5 minutes
- Successful reapplication rate58%
- Regulatory audit prep timeReal-time export, hours
- First-pass approval rate on KYC89%
What changed and by how much
Operational and revenue metrics tracked from go-live, measured against the pre-engagement baseline.
Composition of impact
Approximate breakdown of how this engagement contributed to the business outcome - the headline metric is a roll-up of these levers.
- Onboarding speed & conversion lift36%
- Compliance reviewer capacity freed22%
- Successful reapplication recovery18%
- Fraud risk reduction14%
- Audit & reporting automation10%
What we built it with
Document AI
- Custom vision models for OCR + tamper detection
- Anthropic Claude 3.5 for reasoning
- Indic-language OCR fine-tunes
Decision engine
- Custom policy rules layer
- ML risk model (XGBoost + behavioural features)
- Bureau API integrations (CIBIL, Equifax, CRIF, Experian)
Compliance & audit
- Audit-grade event store (immutable)
- RBI-aligned reporting hooks
- Field-level encryption + PII vault
Customer-facing
- WhatsApp Business API
- In-app SDK for live reapplication
- SMS fallback for low-data customers
What we de-risked along the way
Auto-approval of fraudulent applicants
Mitigation: Hard policy gates ML cannot override, plus a continuous backtest against confirmed-fraud cases. Suspicious patterns escalate to compliance regardless of score.
Regulatory non-compliance
Mitigation: Every decision logged with rule-set version and input snapshot, audit-grade evidence available in real time, quarterly external review of decision distributions.
Bias against under-represented applicants
Mitigation: Quarterly fairness audits across age, gender, geography, and language. Approval-rate deltas tracked; model retrained when drift exceeds threshold.
What we'd carry into the next build
Policy rules first, ML inside the envelope
Trying to encode KYC in an ML model alone is dangerous and not auditable. Policy rules as a hard outer envelope, with ML deciding within, is the only structure RBI auditors trust - and the only structure that ships.
Reapplication recovers more than acquisition
Improving the reapplication flow lifted overall approvals more than any front-of-funnel optimisation we could have done. Drop-off after a failed first attempt was the biggest leak.
Audit-grade logging from day one
Bolting on audit logs later would have meant rebuilding the system. Building the immutable event log on day one made the regulatory sign-off step painless - and made quarterly reporting an export, not a project.
Concordance, not accuracy, is the metric
We did not ask 'is the AI right?' - we asked 'does the AI decide the same way the manual reviewer would have?' Concordance at 97% was the bar regulators understood. Beating that on absolute accuracy was beside the point.
ROI & payback
Investment
Upper-six-figures one-time build + low-five-figures monthly run cost
Payback period
Inside 5 months on conversion lift and reviewer-cost reduction
Year-1 ROI
Estimated 8-12× ROI from incremental loan originations enabled by the conversion lift
“We went from 'onboarding is our bottleneck' to 'onboarding is our advantage.' Customers approve in minutes; reviewers focus on the cases that need them; and regulatory reporting is real-time instead of a weekly fire drill. The conversion lift alone paid for the project inside two quarters.”
Questions about this engagement
Is AI-driven KYC compliant with RBI master directions?+
Yes, when designed correctly. The pattern that works is: hard policy rules that no ML can override, audit-grade event logging on every decision, real-time access controls on PII, and quarterly fairness and concordance audits. We worked with the client's compliance team and external counsel through approval - it added weeks but is required.
How does AI handle Aadhaar and PAN verification?+
Document images are processed by vision models for OCR and tamper detection, then an LLM cross-references the extracted fields against UIDAI-allowed signals and the client's internal database. The actual verification API calls - UIDAI offline KYC, NSDL PAN, etc. - remain deterministic. The AI improves quality of extraction; it does not replace the regulatory APIs.
What happens when document quality is poor?+
Instead of rejecting outright, the AI initiates a conversational reapplication via WhatsApp - explaining specifically what was wrong (glare, partial scan, wrong document) with photo examples. Successful reapplication rate jumped from 12% to 58% with this approach.
How was bias managed?+
Quarterly fairness audits across age, gender, geography, and language - tracking approval rate deltas and intervening when drift exceeds threshold. The model is retrained against an explicitly balanced dataset whenever drift is flagged.
What was the payback period?+
Inside 5 months on conversion lift alone (more loans approved per inbound applicant). Reviewer-cost savings added a second payback path. Year-one ROI estimated at 8 to 12 times the build investment.
Let's buildReady to put AI to work in your business?
Book a free 30-minute strategy call. We'll map your highest-impact automation opportunities and give you a clear roadmap - no obligation.