Indian enterprise AI RFPs in 2026 treat DPDP as a product question, not a legal footnote. Winning answers include a data map, lawful grounds for training and inference, retention owners, subprocessors, logging for high-impact decisions, and a human override path. A privacy-policy hyperlink alone fails security review.
If you sell chatbots, RAG copilots, or document AI into Indian BFSI, manufacturing, or GCCs, procurement already copies language from MeitY’s AI governance draft and DPDP training-data rules. This explainer is the questionnaire playbook — not a statute reprint.
What Changed
- Security questionnaires now bundle DPDP + AI oversight in one section instead of a one-line “GDPR-like” checkbox.
- Buyers ask for artefacts (tables, owners, log fields, deletion SLAs) more than philosophy slides.
- Cross-border model APIs trigger transfer and subprocessor follow-ups even when the UI is India-hosted.
- Regulated sectors add India-region and kill-switch language on top of DPDP — see RBI model-risk checklist.
- Related: AI regulation business guide, 30-day MSME sprint, AI Compliance Kit.
The Details
The eight questions that decide the RFP
Treat these as a fixed set. If your deck cannot answer them in writing, do not submit “we take privacy seriously.”
- What personal data enters training, fine-tuning, or RAG indexes?
- What is the lawful ground / purpose for each field?
- Where do prompts, embeddings, and logs live (region + vendor)?
- Who can access production logs, and for how long?
- How do you delete or anonymise on customer request?
- Which subprocessors touch personal data (model APIs, observability, support)?
- For high-impact decisions, where is human review and override?
- What is your incident path if a model leaks personal data?
Sample answers that pass (and ones that fail)
| Question | Fails | Passes |
|---|---|---|
| Training data | “We use anonymised data” | “Tickets used for fine-tunes are redacted of email/phone; purpose = retrieval quality; ground = contractual notice §4.2; owner = ML lead” |
| Model API region | “OpenAI / Azure” | “Inference via Azure OpenAI swedencentral with DPA on file; India option available within 30 days for BFSI pilots” |
| Logging | “We log for security” | “Fields: model_version, prompt_hash, output_hash, reviewer_id, ts_utc; retention 30 days hot / 90 days cold; access: SRE + security only” |
| HITL | “Humans can intervene” | “Credit and HR flows require reviewer approval before customer-visible action; kill switch owner = on-call eng; last drill: date” |
Copy the pass column into your annex. Vague verbs (“ensure”, “strive”, “best efforts”) score as Weak.
Copy-ready artefact: one-page data map
| Pipeline stage | Personal fields | Purpose | Ground | Retention owner | Region |
|---|---|---|---|---|---|
| Support RAG index | Ticket text, email | Answer employee queries | Contract / notice | Eng lead | India VPC |
| Fine-tune snapshot | Redacted tickets | Improve retrieval | Explicit consent addendum | ML lead | India + named cloud |
| Prompt logs | Free text | Debug + abuse | Legitimate ops + notice | SRE | 30 days, then delete |
| Eval set | Synthetic + redacted samples | Regression tests | Internal R&D; no live PII | ML lead | India staging |
| Support export | Name, ticket ID | Customer success | Contract | CS lead | India CRM |
Fill real rows. Empty cells are audit findings. Re-export this table per customer — do not send the same screenshot to a bank and a SaaS SMB.
Sector overlays (what gets added after DPDP hygiene)
| Buyer | Extra asks beyond the eight questions |
|---|---|
| BFSI / NBFC | Model inventory language, kill switch, India-region preference — map to RBI MRM checklist |
| Health / hospitals | Purpose limitation on clinical text; no silent fine-tunes on discharge summaries |
| GCC / IT services | Customer-code redaction; no paste into public model endpoints; subprocessor list for client security |
| Government / PSU | Disclosure of AI use to end users; grievance officer contact; India hosting preference |
Scoring rubric buyers already use (desk synthesis)
| Score | Pattern | Typical outcome |
|---|---|---|
| Fail | Privacy URL only; “we use OpenAI” with no subprocessor table | Out before technical round |
| Weak | Policies exist; no owners, no retention, no log fields | Clarification loops; lose to faster vendor |
| Pass | Data map + subprocessors + HITL for high-tier flows | Shortlist |
| Strong | Pass + tested deletion SLA + India-region option + sample log schema | Preferred for regulated pilots |
IndiaAIBrief desk estimate from founder conversations in 2026: Weak annexes cost 2–6 weeks of back-and-forth; Strong annexes cut security review to one or two calls.
But here’s what others won’t tell you: procurement grades speed and specificity. A four-page annex delivered in 48 hours beats a polished 40-page PDF arriving after the shortlist closes.
Red flags that kill deals
- Training on customer data “to improve the model” without refreshed notices
- Sharing production prompts with public model endpoints for “debugging”
- No distinction between training and inference processing
- Claiming anonymisation when free-text still names employees or customers
- No link between AI logs and your DPDP grievance / incident process
- Subprocessor list last updated “at launch” with no review date
48-hour annex pack (minimum viable)
- Cover note (half page): product, data categories, India/non-India regions
- Data map table (one page)
- Subprocessor table with DPAs / regions
- Log field list + retention
- HITL / kill-switch for high-tier flows
- Deletion + incident contacts
If you lack time, start the 30-day compliance sprint and ship this pack at day 22 — not day 90.
For a deeper training-data treatment, read How India’s DPDP Act Affects AI Training Data. For sovereign and MeitY context, see India’s AI strategy.
What This Means for Indian MSMEs and CTOs
- Build the eight-answer pack once; reuse across RFPs with customer-specific data-map rows.
- Name owners — anonymous “security team” fails when the buyer asks for an email.
- Budget 5–10 days for the first questionnaire; less thereafter.
- Use the AI Compliance Starter Kit for checklist + workspace templates before counsel hours.
- Book an AI Readiness Audit if you need a scored gap list before a BFSI pilot.
- Track policy changes on the Policy Tracker so your annex does not cite last year’s advisories.
Frequently Asked Questions
What DPDP artefacts do Indian enterprises ask for in AI RFPs?
Expect a personal-data map for training and inference, lawful ground or consent language, retention/deletion owners, subprocessors, and transfer clauses — plus logging fields for high-impact outputs.
Is a privacy policy link enough for DPDP in an AI RFP?
No. Buyers want purpose-specific answers for model training, RAG indexes, and prompt logs. A generic privacy URL without a data map usually scores as non-compliant in security review.
How long does a DPDP-ready AI questionnaire take to prepare?
IndiaAIBrief desk estimate for a first-time MSME vendor: 5–10 working days if you already know your pipelines; 3–4 weeks if you must inventory personal data and vendors from scratch.
Should we hire a law firm before every AI RFP?
Use counsel for regulated BFSI/health/government deals and novel training on personal data. For hygiene questionnaires, ship operational artefacts first — then escalate material legal risk.
What is a good deletion SLA to put in an AI RFP annex?
State a named owner, channel (email/ticket), acknowledgment within 2 business days, and completion within 7–30 days depending on whether data sits in hot logs, RAG indexes, or cold backups — then prove it once in a dry run.
