AI & personal dataUpdated 22 November 20269 min read

AI Vendor Due Diligence Checklist Under the DPDP Act

A practical due diligence checklist for LLM APIs, vector databases, observability tools and AI SaaS vendors, mapped to DPDP section 8 and Rules 6, 7 and 8(3).

By Suhas Bhairav · Reviewed against the DPDP Rules, 2025 (G.S.R. 846(E), 13 Nov 2025)

Accountability
Fiduciary responsible for processor actions (s. 8(1))
Contract
Required for any processor engagement (s. 8(2))
Security clause
Rule 6(1)(f)
Board detailed breach report
Within 72 hours (Rule 7(2)(b))
Processor log retention
Minimum one year (Rule 8(3), Case 2)
Rules apply from
13 May 2027

Quick answer

Under section 8(1) and 8(2) of the DPDP Act, a Data Fiduciary stays responsible for processing done by its AI vendors and may engage them only under a valid contract. Due diligence should confirm prompt retention and training use, Rule 6 security safeguards, breach notice fast enough for your Rule 7 72-hour report, one-year log retention under Rule 8(3), erasure support and transfer locations.

  • You remain accountable for AI vendors' processing under section 8(1); the vendor contract is mandatory under section 8(2).
  • Rule 6(1)(f) expects processor contracts to provide for reasonable security safeguards.
  • Rule 8(3)'s cloud illustration makes you ensure processors retain data and logs for at least one year before erasure.
  • Vendor breach notice must be fast enough for you to send the Board a detailed report within 72 hours.
  • Ask explicitly whether prompts, files and outputs are used to train the vendor's models, and get the answer in the contract.

Who this is for: Procurement and vendor risk teams, CTOs and CISOs, Data Protection Officers, Legal teams negotiating AI contracts, Startups reselling LLM capabilities

Why AI vendor due diligence is a DPDP obligation

Section 8(1) of the DPDP Act makes the Data Fiduciary responsible for complying with the Act "in respect of any processing undertaken by it or on its behalf by a Data Processor". Section 8(2) allows a fiduciary to engage a processor only under a valid contract.

Most AI products depend on a chain of processors: a foundation-model API, an embedding API, a managed vector database, a tracing or evaluation tool, a speech-to-text service, a cloud host, and sometimes a human labelling vendor. Each one touches personal data. If any of them leaks, retains or reuses it, the penalty exposure lands on you — up to ₹250 crore for failing to take reasonable security safeguards.

The Rules make three vendor-facing obligations explicit, all from 13 May 2027:

  • Rule 6(1)(f): appropriate provision in the processor contract for reasonable security safeguards.
  • Rule 7: you must notify the Board and affected users of every breach, including one at a vendor, with a detailed report within 72 hours.
  • Rule 8(3), Case 2: a company using a cloud provider as processor must ensure the provider also retains data and logs for at least one year before erasure.

Due diligence is how you show those obligations were considered before data started flowing.

First, map who is fiduciary and who is processor

Before sending a questionnaire, decide the role of each vendor for each data flow.

Vendor typeTypical DPDP roleWatch for
Foundation-model APIProcessorTerms allowing training on inputs, abuse-monitoring retention
Managed vector DBProcessorSnapshot retention, region, soft deletes
Observability / tracing SaaSProcessorFull prompt capture by default
AI SaaS with its own consumer productPossibly a separate fiduciaryUsing your data to improve its own service
Labelling or annotation vendorProcessorHuman access, offshore teams, copies on laptops
Cloud hostProcessorLog retention, backup regions

A vendor that decides its own purposes for your users' data — for example, training a general model it sells to others — is acting as a fiduciary for that purpose and needs its own legal basis. Most enterprise AI contracts aim to keep the vendor strictly as a processor; check the clauses actually say so.

The distinction is covered in more depth in Data Fiduciary vs Data Processor.

The questions that matter for AI vendors

Generic security questionnaires miss the AI-specific risks. Add these:

Data use

  • Are prompts, files, embeddings or outputs used to train or improve any model? Is the default opt-in or opt-out, and can it be disabled contractually?
  • Is any data reviewed by humans for abuse monitoring or quality? Where are reviewers located?

Retention and erasure

  • How long are prompts, outputs, files and logs retained by default? Is zero or reduced retention available?
  • Is there a deletion API, and does it cover backups and derived stores (indexes, caches)?
  • Can the vendor retain processing logs for one year to meet Rule 8(3), and erase after?

Security

  • Encryption at rest and in transit; customer-managed keys; tenant isolation.
  • Access logging and its availability to you for investigation (Rule 6(1)(c)).

Incidents

  • Contractual breach notice period — hours, not "without undue delay".
  • What information the vendor will give you, mapped to the Rule 7(2)(b) list.

Location and sub-processors

  • Processing and storage regions; list of sub-processors and change notice.
  • Ability to meet sectoral localisation where it applies to you.

Contract clauses to insist on

Turn the answers into binding terms. A DPDP-aligned AI processor agreement should include:

  1. Purpose limitation. Process only on documented instructions, solely to provide the service; no training of general models on your data without written consent.
  2. Security safeguards. Encryption, access control, logging, backups and organisational measures consistent with Rule 6(1)(a)–(g).
  3. Breach notice. Notice to you within a fixed number of hours, with the facts you need for the Board's detailed report under Rule 7(2)(b) — circumstances, mitigation, findings about the cause, remediation.
  4. Retention. Retain personal data, traffic data and logs for at least one year from processing where Rule 8(3) applies, then erase; no longer unless law requires.
  5. Erasure on instruction. Delete on request and at termination, with written confirmation, so you can meet section 8(7).
  6. Sub-processors. Prior notice, flow-down of the same terms, and your right to object.
  7. Cooperation. Assistance with rights requests, Board inquiries and Rule 23 government information requests to the extent lawful.
  8. Transfers. Named regions; compliance with any Rule 15 requirements and section 16 restrictions.
  9. Audit. Reports, certifications or audit rights proportionate to the risk.

A ready-to-adapt clause set is in the DPA clauses template.

Sector overlays: RBI, SEBI and CERT-In

Regulated entities have extra layers on top of DPDP.

  • SEBI. In early 2025 SEBI amended its Intermediaries Regulations (regulation 16C) so a regulated person using AI tools is solely responsible for data privacy and security, outputs and compliance, regardless of the vendor. Vendor due diligence does not shift that responsibility. See SEBI AI/ML rules.
  • RBI. The FREE-AI committee report of August 2025 recommends board-approved AI policies, AI inventories, audits, model validation and AI incident reporting. These become binding only when RBI issues directions or circulars; check the latest position. Payment data storage rules from April 2018 continue to apply to vendors in the payment flow.
  • CERT-In. The April 2022 Directions require reporting specified cyber incidents within 6 hours of noticing and keeping ICT logs for a rolling 180 days within India. Your vendors' logs must let you meet both. See CERT-In rules for AI systems.

Where sectoral rules are stricter than DPDP, they continue to apply. Section 16(2) of the Act preserves stricter transfer restrictions under other laws.

Due diligence is not a one-time event

AI vendors change faster than most suppliers: new model versions, new regions, new default retention, new sub-processors. Build monitoring into the relationship.

  • Re-review triggers: a new model or feature that changes data use; terms-of-service updates; a vendor incident; a new region or sub-processor; your own use expanding to new data categories such as children's or health data.
  • Annual review: refresh the questionnaire, collect current certifications and test the deletion API with a synthetic user.
  • Technical verification: check that zero-retention or no-training flags are actually set on your account and API keys, not just available.
  • Inventory: keep a register of every AI vendor, data categories sent, region, retention setting, contract date and owner. Significant Data Fiduciaries, once notified, will also need this for the annual DPIA and audit under Rule 13(1).
  • Exit plan: know how to export and erase your data, and how long the vendor keeps logs after termination.

Before 13 May 2027, run a vendor sweep: list every AI processor in production, classify risk, and close contract gaps on the highest-risk flows first.

Further reading on suhasbhairav.com

Before 13 May 2027

Compliance checklist

  1. Classify each AI vendor's DPDP role

    Record whether the vendor is a processor or a separate fiduciary for each data flow.

    DPDP Act s. 8(1)

  2. Sign a written processor contract

    Do not send personal data to any AI vendor without a valid contract in place.

    DPDP Act s. 8(2)

  3. Prohibit training on your data

    Contractually bar use of prompts, files and outputs for training general models unless you consent in writing.

    DPDP Act s. 8(1)

  4. Fix a breach notice period in hours

    Require prompt vendor notice with the details needed for your 72-hour Board report.

    Rule 7(2)(b)

  5. Align vendor log retention with Rule 8(3)

    Ensure processors keep data and logs for at least one year, then erase.

    Rule 8(3)

  6. Test the deletion API

    Erase a synthetic user and confirm removal from indexes, caches and logs.

    DPDP Act s. 8(7)

  7. Record processing regions and sub-processors

    Keep a current list and require prior notice of changes.

    Rule 15; s. 16

  8. Include security safeguard clauses

    Spell out encryption, access control, logging and backup obligations.

    Rule 6(1)(f)

  9. Verify retention and training flags are enabled

    Check account and API-key settings, not just marketing claims.

    Rule 6(1)(g)

  10. Maintain an AI vendor register

    Track vendor, data categories, region, retention, contract date and owner.

    DPDP Act s. 8(4)

Key dates

  1. DPDP Rules notified

  2. Rules 6, 7 and 8 apply to fiduciaries and their processors

Answers

Frequently asked questions

Is an LLM API provider a Data Processor under the DPDP Act?

Usually yes, when it processes personal data only on your instructions to deliver the service. If its terms let it use your data for its own purposes, such as training models it sells to others, it may act as a fiduciary for that purpose. Read the data-use clauses and negotiate them so the vendor stays a processor.

Who is liable if my AI vendor leaks customer data?

You are, as the Data Fiduciary. Section 8(1) makes the fiduciary responsible for processing done on its behalf by a Data Processor. You must notify the Board and affected users under Rule 7. Your contract can give you indemnity and cooperation rights against the vendor, but it cannot transfer your statutory liability.

What breach notice period should I require from an AI vendor?

Short enough to let you send the Board a detailed report within 72 hours of becoming aware, as Rule 7(2)(b) requires, and to meet CERT-In's 6-hour reporting for specified incidents. Many teams set vendor notice in hours rather than days, with an obligation to share facts, cause findings and mitigation as they emerge.

Does Rule 8(3) apply to my vendors?

Yes. Rule 8(3) covers processing undertaken by you or on your behalf by a Data Processor, and its Case 2 illustration says a company using a cloud provider must ensure the provider also retains data and logs for at least one year before erasure, unless another law requires longer.

Do I need a vendor's servers to be in India under DPDP?

Not as a general rule. DPDP permits cross-border transfers except to countries the Central Government restricts under section 16, subject to Rule 15 requirements. Localisation applies through sectoral rules such as RBI payment data storage, and to Significant Data Fiduciaries for data the Government specifies under Rule 13(4).

When should I involve legal counsel in AI vendor reviews?

Involve counsel when a vendor refuses to remove training-use rights, when data includes children's, health or financial information, when the vendor may be acting as a fiduciary, or when sectoral rules from RBI, SEBI or IRDAI apply. Those are the points where contract wording changes your liability materially.

Legal references and sources

  • Digital Personal Data Protection Act, 2023 — Sections 8(1), 8(2), 8(5), 8(7) (official text)
  • Digital Personal Data Protection Rules, 2025 — Rule 6(1)(f) (official text)
  • Digital Personal Data Protection Rules, 2025 — Rule 7(2) (official text)
  • Digital Personal Data Protection Rules, 2025 — Rule 8(3), Illustration Case 2 (official text)
  • Digital Personal Data Protection Rules, 2025 — Rule 15 (official text)

This guide explains the law for engineering and business teams. It is not legal advice; confirm how it applies to your organisation with qualified counsel, and check the Gazette for later notifications or amendments.

About the author

Suhas Bhairav

Suhas Bhairav

Systems architect and applied AI practitioner who designs production AI systems for Indian and global companies, with a focus on privacy-by-design under the DPDP Act and the EU AI Act.