AI Proof of Concept Development, security and the DPDP Act: a compliance checklist
Is AI proof of concept development compliant with the DPDP Act?
AI proof of concept development is compliant with the DPDP Act only when the pilot treats real personal data as production would: a lawful basis, a written purpose, minimised fields, restricted access, a deletion date and an audit log. The word pilot creates no exemption.
AI proof of concept development is compliant with the DPDP Act only when the proof of concept treats real personal data with the same controls as production: a lawful basis, a written purpose, a named data fiduciary, minimised fields, restricted access, a scheduled deletion date and an audit log. Nothing about the word pilot creates an exemption.
This article maps each DPDP obligation to the concrete engineering control that satisfies it during a short pilot, shows where the data physically sits when model providers are involved, gives the real cost of running a compliance-first proof of concept, and sets out the checklist to complete before any extract leaves your systems.
What the DPDP Act actually asks of a short pilot
The Digital Personal Data Protection Act, 2023 is India's general data protection law. It applies to digital personal data processed in India, and to processing outside India that relates to offering goods or services to people in India. It names the organisation deciding the purpose as the data fiduciary, the person the data describes as the data principal, and anyone processing on the fiduciary's behalf as a data processor. Your AI vendor is a processor. You remain the fiduciary.
There is no proof-of-concept exemption in the statute. A four-week pilot that loads six months of loan applications into a vector store is processing personal data for a new purpose, and every obligation that applies to your production platform applies to that pilot. The practical difference is scale, not duty: a smaller sample, a shorter retention window and a narrower set of people with access.
That is the good news. A proof of concept is the cheapest place in the whole programme to get the data handling right, because the sample is small enough to inspect by hand and the architecture is not yet load-bearing. Our AI POC Sprint treats the data agreement as a deliverable rather than paperwork that trails the build. The broader statutory picture is covered in DPDP Act 2023 and AI: what Indian companies must do and defined in the DPDP Act glossary entry.
DPDP obligations mapped to proof-of-concept controls
Compliance arguments fail when they stay abstract. The table below is the mapping we work through in the first two days of every pilot that touches personal data, and it is the document we hand a client's security reviewer.
| DPDP obligation | What it means during a pilot | The control we build |
|---|---|---|
| Lawful basis and notice | Every record in the sample needs consent or a listed legitimate use | Consent status travels with the extract; records without it are dropped at ingest |
| Purpose limitation | Pilot data may be used only for the stated evaluation | Written purpose in the data note, enforced by a single-purpose bucket and no reuse |
| Data minimisation | Only the fields the model reads, not the whole table | Column allow-list in the extract script, signed off before the first run |
| Storage limitation | The sample has an end date, not an indefinite life | Deletion job scheduled for the pilot end date, with a written destruction certificate |
| Reasonable security safeguards | Encryption, access control and logging at production standard | Private network deployment, role-based access, encrypted volumes, request-level audit log |
| Breach notification | You must be able to detect and report an incident | Alerting on unusual access patterns and a named incident owner on both sides |
| Processor obligations | Every vendor touching the data is under contract | NDA before the first working session, plus a processing agreement and a sub-processor list |
| Data principal rights | Erasure and correction requests must reach the pilot copy | Stable record identifiers kept so a deletion request can be honoured in the sample too |
Does an AI proof of concept need real personal data at all?
Often it does not, and that is the single largest compliance saving available to you. A proof of concept exists to retire technical risk, and technical risk usually lives in document structure, language mix, retrieval quality or latency rather than in the identity of the people described.
We run the question in this order. Can the hypothesis be tested on synthetic records that match the real schema and messiness? If not, can it be tested on production records with identifiers redacted at extract time using PII redaction? If not, can it be tested on a consented subset of a few hundred records? Only when all three fail does a pilot need a broad extract of live personal data, and by then everyone understands exactly why.
Redaction is not a checkbox. Names inside free-text remarks, account numbers in scanned pages and phone numbers in call notes all survive naive masking, so the redaction pass is itself something to evaluate before it is trusted.
Where the data physically sits
Residency
The DPDP Act permits cross-border transfer except to countries the central government restricts, but sector rules bite harder: RBI-regulated entities have their own storage expectations, and many enterprise contracts simply require India-only processing. Decide residency before the architecture, because retrofitting it means changing model provider, vector store and logging stack at once. Data residency is a design constraint, not a deployment setting.
Model providers and training
If the pilot calls a hosted model, personal data leaves your perimeter. Read the provider's enterprise terms for whether inputs are used to train, what the retention window for abuse monitoring is, and which region the endpoint serves. Where the answer is unacceptable, an open-weight model on your own infrastructure is the alternative, which is the ground our self-hosted agentic AI work covers, from $31,500 or ₹20,80,000 plus infrastructure.
Logs, traces and the quiet leak
Observability is where most pilots break their own rules. Prompt traces contain the retrieved documents, evaluation dashboards cache sample answers, and a debugging notebook on an engineer's laptop is a copy nobody listed. Traces need the same retention clock and the same access control as the source data, and laptops are not a storage tier.
What a compliant proof of concept costs
A compliance-first pilot does not cost more than a careless one; it costs the same money spent in a different order. Our AI POC Sprint is a fixed three-week engagement from $6,250 or ₹4,00,000, rising to $10,500 or ₹6,80,000 where the data environment is complex. A ten-day Sprint Zero at $3,250 or ₹2,00,000, credited to the next build, is often the better first step when the data-sharing route is not yet agreed, because legal review, not engineering, is what delays pilots. Starting prices for every programme sit on the pricing page, and you own the code, prompts, infrastructure and documentation at the end of it.
Budget separately for the things compliance adds: a redaction pass, a private deployment if residency demands it, and the security questionnaire your own risk team will send. A security questionnaire for AI vendors lists the questions worth asking before you sign.
When this approach is the wrong choice
A full compliance wrapper around a pilot is wrong when no personal data is involved. If you are testing retrieval quality over public product manuals or internal engineering documentation with no individuals named, the DPDP Act is not engaged and building consent plumbing for it wastes a week of a three-week sprint.
It is also the wrong shape when the honest blocker is organisational. If your data protection officer has not agreed that the use case is lawful, no architecture rescues it, and the correct next step is a written purpose note and a decision, not a pilot. Finally, if the compliance work would consume more than about a third of the engagement, the pilot is too ambitious: narrow the hypothesis until it fits.
What this looked like on a real engagement
An NBFC came to us wanting document intelligence for KYC and loan onboarding, which is about as sensitive as Indian personal data gets: identity documents, addresses, income proofs. The pilot ran on a consented subset inside their own environment, with redaction at extract, record-level identifiers preserved for erasure requests and a deletion date fixed at the start. The full engagement is described in the KYC document intelligence case study, and the same pattern suits most regulated lending and insurance work.
The lesson was not that compliance slowed the pilot. It was that the compliance conversation forced a smaller, sharper sample, and the smaller sample produced clearer evaluation results than a full data dump would have.
The checklist before any extract leaves your systems
- Name the fiduciary and the owner. One person on your side signs off the purpose and the deletion date.
- Write the purpose in one sentence. If it needs a paragraph, the pilot scope is not yet locked.
- Test the synthetic route first. Prove you cannot answer the question without real personal data.
- Allow-list the columns. Start from nothing and add fields the model demonstrably needs.
- Fix residency before architecture. Region, model provider and vector store are one decision, not three.
- Schedule deletion on day one. A dated job plus a destruction certificate, not an intention.
- Cover logs and traces. Same retention clock, same access control, no laptop copies.
- Sign the NDA and processing agreement first. We sign NDAs before the first working session as standard.
Related reading
From POC to production: the checklist covers what has to be true before a pilot becomes a system, AI governance for mid-size companies without a compliance team is the practical route for organisations with no dedicated privacy function, and permission-aware retrieval explains how to stop a retrieval system answering outside a user's access rights. The Ministry of Electronics and Information Technology publishes the Act and its data protection framework material as the primary source for the obligations described here.
Treat the pilot as production with a smaller sample and a shorter clock, and the compliance question stops being the thing that kills your proof of concept.
Frequently asked questions
Does the DPDP Act apply to an AI pilot or only to production systems?
▾
It applies to both. The Act governs the processing of digital personal data and contains no exemption for pilots, proofs of concept or internal experiments. A four-week evaluation using real customer records carries the same obligations on lawful basis, minimisation, security, retention and data principal rights as a live platform does.
Can an AI proof of concept use data hosted outside India?
▾
Usually yes, since the DPDP Act permits cross-border transfer except to countries the central government restricts. Sector rules are stricter: RBI-regulated entities and many enterprise contracts require India-only processing. Decide residency before you pick a model provider, because changing it later means rebuilding the retrieval and logging stack too.
What happens to pilot data when the proof of concept ends?
▾
It should be deleted on a date agreed before the pilot starts, through a scheduled job rather than a manual promise, with a written destruction certificate. Deletion must cover derived copies as well: vector indexes, prompt traces, evaluation caches and any extract on an engineer's machine, all of which count as retained personal data.