For RBI-regulated entities DPDP · RBI · CERT-In, enforced at runtime

Run frontier AI agents.
Prove compliance
on every model call.

EAIP is the compliance-as-a-runtime control plane for AI agents in regulated India. Data residency, DPDP consent, PII redaction, and a hash-chained audit trail are enforced on every model call — not configured and hoped for. Deploy agents across credit, collections, fraud, and compliance, and show your board and the RBI exactly where every byte of borrower data went.

Inference pinned to India — Bedrock ap-south-1 · Vertex asia-south1 Fail-closed PII redaction membrane SHA-256 hash-chained audit, 7-year retention Per-call residency evidence
6 Lending agent departments pre-built, eval-set included
29 Workflow node types incl. consent gate, PII redact, HITL
24 Governed connectors CIBIL, Karza, Razorpay, Sahamati, DigiLocker…
7 yr Audit retention SHA-256 hash-chained, append-only
100% Calls through the membrane redaction + audit, fail-closed
Why now

Three regulators are asking
the same question.

When an AI agent touches a borrower's data — where did it go, who saw it, and can you prove it? The hyperscaler platforms hand you primitives and leave the compliance program to you. The deadlines below don't wait for that project to finish.

DPDP Act 2023 + Rules 2025

Penalties up to ₹250 crore

Consent Manager: Nov 13, 2026 · Full enforcement: May 13, 2027

Personal data moving from KYC journeys into prompts, embeddings, and model outputs needs consent lineage, retention control, and erasure — end to end. Most AI stacks cannot trace where one record lives across prompts, vector stores, and response logs.

RBI FREE-AI Framework

Board-level accountability

Released Aug 13, 2025 · Master Direction incorporation expected

A board-approved AI policy, independent model audits, explainability, and grievance redressal for AI-driven decisions — and the obligations flow down contractually to every AI vendor a regulated entity uses.

RBI Outsourcing Directions 2025

Non-compliant contracts are in violation now

Effective Nov 28, 2025 · Contract transition deadline Apr 10, 2026 — passed

Unqualified RBI and NBFC audit access to vendor systems. Critical data processed — not just stored — in India. Customer-controlled encryption keys. 6-hour incident notification from vendor to NBFC.

CERT-In Directions 2022

6-hour incident reporting clock

Active enforcement

Specified cyber incidents reported to CERT-In within 6 hours. System clocks synced to time.nplindia.in. Tamper-evident logging that an inspector can actually verify, not just be told about.

The enforcement is not hypothetical. The RBI levied ₹56 crore across 304 enforcement actions in FY 2024-25 — and "delegation of compliance decisions to third-party agents" is already a cited violation pattern. Ungoverned AI agents are how the next ones happen.

The compliance runtime

Four guarantees
no other platform makes.

These are structural properties of the architecture, not configuration you assemble and hope holds. Competing platforms ship one or two as options. None ships all four as defaults.

01 Compute-to-data

The model never sees raw rows

For agentic database work, row aggregation and profiling run inside the India-region data gateway. The model receives a masked statistical profile — never raw values, not even non-PII fields. A SQL guard enforces read-only access before any query reaches your database.

"What did the LLM see about my borrowers?" collapses to: "Nothing. Here is the profile it received."

02 Fail-closed membrane

Redaction that cannot fall open

Before any prompt, completion, or embedding request leaves the platform, a mandatory redaction step strips or tokenises Aadhaar (Verhoeff-checked), PAN, card numbers (Luhn-checked), UPI IDs, bank accounts, and CIBIL scores. If the redaction engine errors, the call is blocked — it never falls open.

The audit records entity types and counts redacted. Never the raw values.

03 Tenant-derived region enforcement

A residency floor consent can’t override

Inference region is derived from the tenant on the server — never trusted from a request. Cross-border model choice exists only behind a recorded organisation opt-in. And RBI-localised elements — Aadhaar, card and UPI data, bank and loan accounts, bureau scores — are hard-floored: they never cross a border in cleartext, consent or not.

Every call writes its resolved inference region to the audit trail as per-call residency evidence.

04 SHA-256 hash-chained trace store

An audit trail you can hand to an inspector

Every model call, tool call, and policy decision lands in an append-only store with a continuous SHA-256 hash chain — across batch boundaries and service restarts. Retained 7 years, well beyond CERT-In’s 180-day floor and aligned with PMLA-grade record-keeping. Time sourced from time.nplindia.in.

Chain verification is visible in-product. Tamper-evident means you can prove it, not just claim it.

Architecture

Sits over any model.
Under your obligations.

EAIP is not a model and not a chatbot. It is the governance layer between your agents and whichever models you choose — enforcing Indian statute on every call, whoever serves the tokens.

Your agents

Credit assessment · Collections · Fraud · Compliance monitoring · NPA filing · Customer operations

EAIP compliance runtime — every call passes through

Cedar policy ACL · PII redaction membrane · DPDP consent gates · Injection scanning · Token budgets · Hash-chained audit emit

Any model, India-pinned · bring your own keys

Gemini on Vertex AI asia-south1 (Mumbai) · Claude on AWS Bedrock APAC · India-sovereign models on the roadmap

Data plane in AWS ap-south-1 (Mumbai): database, auth, vault, vector memory, data gateway, and the trace store. Your model keys stay inside your residency boundary.

Workflows with compliance lanes

  • 29 node types: agents, RAG, DB queries, routing, loops
  • consent_gate — runs blocked without a matching DPDP consent record
  • pii_redact / pii_check — Aadhaar, PAN, account detection in-flow
  • hitl — SLA-gated human review for underwriting and IRAC sign-off
  • IST-native scheduled triggers and signed webhooks

Governed integrations

  • 24 connectors, including the India lending rails: CIBIL, Karza, Razorpay, Sahamati (Account Aggregator), DigiLocker
  • India-fintech connectors flagged india_only with strict scan profiles
  • MCP tool farm with fail-closed output scanning and SSRF re-validation
  • Database access is compute-to-data: read-only SQL, in-region aggregation

Spend you can defend at a board meeting

  • Per-session token budgets: off / warn / enforce
  • INR-denominated spend dashboard per tenant and per agent
  • Daily cost quota that fails closed outside development
  • Per-call cost attribution in the same audit trail as everything else
Statute → mechanism → evidence

Not certifications.
Runtime enforcement.

Every obligation maps to a named mechanism in the platform and an evidence artifact you can produce on demand. This is what an inspection-ready AI deployment looks like.

DPDP Act 2023 + Rules 2025

Runtime-enforced

Consent Manager: Nov 13, 2026 · Binding: May 13, 2027

  • consent_gate workflow node — runs fail without a matching consent record for the stated purpose (loan origination, KYC, credit assessment, collections…)
  • Consent records and erasure receipts with cascade across memory and traces
  • PII redaction membrane on every model call — data minimisation by default
  • Cross-border transfer only behind a recorded organisation opt-in

Evidence: Consent lineage per run · erasure verification query

RBI FREE-AI Framework (Aug 2025)

Runtime-enforced

Advisory today · Master Direction incorporation expected

  • Agent registry with approval workflow — a live AI inventory, not a spreadsheet
  • Explainable decisions: the reference credit agent cites rules and ships SHAP explanations
  • Per-agent cost and token budgets with enforce mode
  • Human-in-the-loop nodes for high-stakes lending decisions

Evidence: Board-reportable governance center · model inventory

RBI Outsourcing + Digital Lending Directions

Runtime-enforced

Outsourcing transition deadline: Apr 10, 2026 — passed

  • India-resident processing, not just storage — inference pinned to ap-south-1 / asia-south1
  • Region derived from the tenant server-side; never trusted from a request
  • Bring-your-own model keys — your provider relationship, your boundary
  • Architecture built for regulator and customer audit access

Evidence: Per-call inference_region in the audit trail

CERT-In Directions 2022

Runtime-enforced

Active enforcement

  • 6-hour incident reporting pipeline
  • NTP synchronised to time.nplindia.in (IST, no DST ambiguity)
  • SHA-256 hash-chained, append-only trace store
  • 7-year retention — beyond the 180-day CERT-In floor, aligned to PMLA-grade record-keeping

Evidence: Hash-chain continuity proof, verified in-product

Reference vertical

Cymbal Lend:
a full NBFC, pre-wired.

EAIP ships a reference lending vertical as six compliance-wired agent departments — deployable, not slideware. Each ships as a versioned agent spec with a system prompt, a policy matrix, and a golden evaluation set. The credit decision arrives with the explainability FREE-AI asks for, and the whole run lands in the audit chain.

Credit & Risk

Credit Uplift

3 hr → 30 s top-up decisions, every decision rule-cited with SHAP explainability

Compliance & Legal

DPDP / RBI Compliance

Real-time circular monitoring and audit-ready reports

Finance

NPA Filing

3 days → 30 min per RBI NPA/IRAC filing

Marketing

Campaign Intelligence

Days → minutes per campaign, with consent-aware audience filters

Engineering

DevEx / Infra

Blast-radius assessment and incident triage

Leadership

CXO Strategy

Multi-entity P&L and fleet-wide insight

DPDP consent gates

In every personal-data workflow

Underwriting workflows

Intake triage → parallel pod → HITL sign-off

India rails connected

CIBIL · Karza · Razorpay · Sahamati AA · DigiLocker

The honest comparison

A region dropdown is not
a compliance posture.

Every platform below gives a regulated Indian lender primitives and leaves the compliance program to the customer. EAIP ships the program — and runs on top of whichever cloud you already use.

Google — Gemini Enterprise

“Region selector with an India region.”

ML-processing residency is US/EU multi-region; for India, the best models and grounding fall back to the global region — Google’s own docs recommend global unless a regulator forces otherwise. For UPI, PAN, and bureau data, that is the opposite of what an NBFC needs.

Microsoft — Copilot Studio + Foundry

“Purview DLP and India processing on the roadmap.”

DLP is detect-and-block — the model still sees raw data when a rule doesn’t fire, and that is a reportable incident. India processing has been a shifting 2025–26 roadmap framed as an M365 feature, not an RBI residency guarantee.

AWS — Bedrock AgentCore

“GA in Mumbai (ap-south-1).”

A region is geography, not compliance: no DPDP consent primitive, no CERT-In scaffolding, and the documented Claude-in-India path is cross-region inference that can leave the country. You get a region; the compliance program is left to you.

India CX vendors — chat & voice platforms

“DPDP-aligned, certified, BFSI references.”

Compliance as certifications and checkboxes: no policy-as-code, no tenant-derived residency floor, no fail-closed scanning, no hash-chained audit an inspector can verify. Strong at conversation volume; absent at the governed agent layer.

EAIP

Compliance-as-a-runtime.

India-pinned inference with a tenant-derived hard floor, fail-closed redaction on every call, compute-to-data so the model never sees raw rows, and a hash-chained audit trail with per-call residency evidence — over Gemini, Claude, or whichever model serves you best.

Commercial model

We monetise the runtime,
not your tokens.

Predictable, INR-denominated, with hard caps — not a six-dimension consumption bill.

Open now

Starter

Free / self-serve, no card required

Sign up, verify your email, and get a governed tenant in ap-south-1 with the consent ledger, Cedar policy gate, and hash-chained audit trail already on.

  • Free starter tenant in AWS ap-south-1
  • Build agents with governed tools and connectors
  • Per-call residency record, live
  • DPDP consent ledger and guard policies on by default
  • Hash-chained audit trail from your first call
Get started free

Platform

Contact us / per org / month + usage

The governed runtime as a service. Bring your own model keys — we monetise the compliance runtime, not your tokens.

  • BYO model keys — your provider relationship
  • Transparent platform fee on governed usage
  • INR spend governance with hard caps
  • DPDP consent + erasure machinery
  • Trust & Governance Center
Talk to us

Enterprise / BYOC

Custom / for banks and NBFCs

Data plane in your VPC. Built for the audit-access and residency posture the RBI outsourcing directions expect of your vendors.

  • Data plane in your VPC, ap-south-1
  • Customer-held keys
  • Regulator and customer audit access
  • India compliance pack as the strict default
  • Dedicated onboarding engineer
Request a briefing
Self-serve · Free starter tenant

Your first governed agent, live today.

Sign up, verify your email, and get a free starter tenant running in ap-south-1 — with the consent ledger, Cedar policy gate, and hash-chained audit trail already on.

Email verification required · No card needed · Runs in AWS ap-south-1 (Mumbai)