AI agents Architecture Pricing Compliance Compare Book a demo
AI agents

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.

The eight

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."

01

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.

02

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.

03

Underwriting checklist

Runs your credit policy as a rule engine. Returns recommendations with policy match. Underwriter sees only judgment cases, never bookkeeping.

04

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.

05

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.

06

Audit and compliance

Tamper-evident audit logs. Pre-built reports for RBI, SRO, internal and statutory audits. Inspection-mode for regulators.

07

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.

08

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.

How agents are deployed

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.

01

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.

02

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.

03

Auto mode

Agent acts inside a defined policy envelope. Pre-approved for KYC, disbursement, bucket-1 collections under ₹X. Anything outside the envelope escalates.

04

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.

Models

The model layer is configurable.
You are not locked in.

LLMs

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.

Vision

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.

Voice

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.

Routing

Cost-aware

Same task — cheaper model first, escalate to bigger model on low-confidence. Per-tenant cost ceilings configurable in the dashboard.

Want to see the agents on your data?

30-minute walkthrough. Your data, not ours.