Frequently asked questions

How EAIP runs AI agents under DPDP, RBI and CERT-In obligations — with the mechanism behind every claim.

What is EAIP?

EAIP (Enterprise Agent Intelligence Platform) is a platform for running AI agents inside regulated enterprises. It combines an agent execution engine, governed tool and data connectors, and a compliance runtime: DPDP consent records, Cedar policy checks, PII redaction, and a hash-chained audit trail are enforced on every model call, not documented after the fact.

Is there a DPDP Act-compliant AI agent platform for Indian companies?

Yes. EAIP was built India-first for the DPDP Act 2023: every tenant gets consent records with a legal basis at provisioning, cross-border model routing is consent-gated per data principal, erasure requests cascade with signed receipts, and personal data is redacted before any prompt leaves the tenant boundary. Inference is pinned to Indian regions (AWS ap-south-1) by default.

How do RBI-regulated lenders audit AI agent decisions?

Every agent turn in EAIP writes an append-only audit event — actor, tenant, tool calls, model, region, token counts — into a SHA-256 hash-chained store with 7-year retention aligned with PMLA record-keeping expectations. The chain is verifiable end-to-end, so an auditor can prove no event was inserted, altered, or removed.

Can EAIP run inside our own AWS account?

Yes. The reference deployment is BYOC: the full platform runs as containers in your AWS account in ap-south-1 (Mumbai), provisioned by Terraform you can read. Your data and model traffic never leave your cloud. The stack can be paused to near-zero cost when idle and resumed in minutes.

Which LLMs does EAIP support?

EAIP is model-neutral. Out of the box it routes to Claude models on AWS Bedrock in ap-south-1, with bring-your-own keys for other providers. Cross-border models can be enabled per tenant behind an explicit consent gate; personal-data workloads stay on India-resident endpoints regardless of configuration.

How does EAIP handle prompt injection and untrusted tool output?

All connector and MCP tool output passes through a fail-closed scanning gate before it reaches the model: if the injection scanner is unavailable, the call is blocked rather than passed through. Agent capabilities are allow-listed per tenant with Cedar policies, so a compromised tool cannot reach tools it was never granted.

What does "compliance as a runtime" mean?

It means regulatory controls execute in the request path instead of living in a policy document. In EAIP, a model call that would violate residency is rejected at routing time; a tool call without a capability grant never dispatches; PII is redacted before inference; and the audit event is written before the response returns. The evidence pack falls out of normal operation.

How is data residency actually enforced?

A tenant's region is stored server-side and read from the tenant record on every call — never trusted from a request body. For Indian tenants the model router only selects India-resident endpoints, and financial-grade identifiers are hard-floored: they never leave India in cleartext even if someone opts into cross-border models.

What is stored in the audit trail, and for how long?

Each event stores the actor, tenant, action, resource, timestamps from a CERT-In-recognised NTP source, and a SHA-256 link to the previous event. Retention is 7 years, aligned with PMLA record-keeping, on append-only storage with a deletion registry for DPDP erasure rather than physical row deletion.

How do I get started with EAIP?

Sign up on the platform, verify your email, and a free starter tenant is provisioned automatically in ap-south-1 with default guard policies, a consent ledger, and the audit chain already active. From there you can create your first agent, connect data sources, and invite teammates.