Eight specific operational tasks. Eight specific outcomes.
"AI agents" is a strong claim. This is the receipt — eight tasks the agents handle so your team can focus on judgment, not bookkeeping.
Each one replaces a specific role.
Not a generic copilot.
Every agent has a defined input, a defined output and a defined failure mode. None of them are "an LLM with a system prompt."
Lead capture
Captures leads from web, WhatsApp and partner channels. Conversational qualification on WhatsApp and voice. Hands warm leads to the RM with the full context attached.
KYC and verification
Document extraction with vision models. Bureau, GST, MCA, PAN cross-checks. WhatsApp loop until the file is complete. Flags inconsistencies to the underwriter, not the customer.
Underwriting checklist
Runs your credit policy as a rule engine. Returns recommendations with policy match. Underwriter sees only judgment cases, never bookkeeping.
Disbursement orchestration
Triggers NEFT / IMPS / RTGS / UPI to your payment partner. Reconciles, posts to the ledger, confirms to the borrower. Sub-minute end-to-end on the happy path.
Bucket-1 collections (THALA)
Voice, WhatsApp and SMS follow-ups on 1–30 dpd. Multilingual. Detects promise-to-pay, disputed-billing, hardship; routes to a human when the conversation calls for it.
Audit and compliance
Tamper-evident audit logs. Pre-built reports for RBI, SRO, internal and statutory audits. Inspection-mode for regulators.
Field productivity (KYEHR)
Attendance, route discipline, customer-visit completion. Outliers — over and under — surfaced automatically. Field officers know they are seen; managers know who needs help.
Cross-sell and renewals
Repayment behaviour and life events surface the right offer at the right time. Hands the lead to the relationship manager with the why pre-written.
Agents do bookkeeping.
Humans do judgment.
The handoff is explicit.
Four modes.
Earned, not assumed.
No agent goes from contract to auto-mode on day one. The path runs through shadow → co-pilot → auto, with continuous evaluation throughout.
Shadow mode
Each agent runs in shadow against your existing process for 4–6 weeks. Decisions are logged but not actioned. The variance from your team's outcomes is the calibration signal.
Co-pilot mode
Agent recommends, human approves. This is the steady state for underwriting, exception handling and high-ticket disbursements. Approval rate over time becomes a SLA.
Auto mode
Agent acts inside a defined policy envelope. Pre-approved for KYC, disbursement, bucket-1 collections under ₹X. Anything outside the envelope escalates.
Continuous evaluation
A 5% sample of agent decisions across all modes is reviewed weekly by a senior team member. The reviews feed back into prompt tuning, rule updates and human-judgment thresholds.
The model layer is configurable.
You are not locked in.
Frontier + open-weight
GPT-4 class for the highest-stakes reasoning; open-weight models (Llama, Qwen) for high-volume routine. Per-agent model selection, swappable.
Document extraction
In-house vision pipeline tuned for Indian KYC documents (PAN, Aadhaar masked, bank statements, GST returns, ITRs). Falls back to cloud vision APIs on edge cases.
Indic STT/TTS
11 Indian languages on inbound voice; outbound voice in 8. Native-quality Indic TTS — not a robotic voice the borrower hangs up on.
Cost-aware
Same task — cheaper model first, escalate to bigger model on low-confidence. Per-tenant cost ceilings configurable in the dashboard.