AI strategy consulting, security and the DPDP Act: a compliance checklist
Is AI strategy consulting compliant with the DPDP Act?
It can be, and it is not automatic. The DPDP Act applies to the personal data a strategy engagement touches, not to the engagement itself. Compliance depends on four choices: what data leaves your systems, whose consent covers it, where it is processed, and how long anyone keeps it.
It can be, and it is not automatic. The Digital Personal Data Protection Act, 2023 applies to the personal data a strategy engagement touches, not to the engagement itself. Compliance depends on four choices: what data leaves your systems, whose consent covers it, where it is processed, and how long anyone keeps it.
What follows is a checklist rather than a legal opinion. It walks the obligations that actually bite during AI strategy consulting, the architecture that satisfies them without slowing the work down, and the two points where Indian companies most often get caught out.
What the DPDP Act requires of a consulting engagement
The Digital Personal Data Protection Act, 2023 is India's general law on digital personal data. It covers personal data processed in India, and processing outside India that relates to offering goods or services to people in India. During AI strategy consulting the Act reaches your engagement the moment a consultant sees a customer name, a phone number, a loan file or a support transcript.
Two roles matter. You, the organisation that decides why and how the data is processed, are the data fiduciary. A consultant processing that data on your instructions is a data processor. Accountability stays with the fiduciary, which is why the contract, rather than the consultant's good intentions, is the control that counts. The Ministry of Electronics and Information Technology publishes the Act and the implementing rules on its data protection framework page.
Three obligations shape the engagement. Purpose limitation: data collected for loan servicing cannot be repurposed for a machine learning experiment without a lawful basis. Storage limitation: data is erased once the purpose is served. Security safeguards: the fiduciary must take reasonable measures, and a personal data breach must be reported to the Data Protection Board and to the people affected.
None of that forbids AI strategy work. It shapes it. Sample instead of copying, redact instead of trusting, process inside your boundary instead of outside it, and delete on a schedule that somebody owns by name. The short definition worth circulating internally is the DPDP Act glossary entry.
Obligation by obligation: what each one means during the engagement
This table maps the obligations onto the artefacts a discovery or strategy engagement actually produces, and onto the evidence you should be able to show an auditor twelve months later.
| DPDP obligation | What it means during a strategy engagement | Evidence to keep |
|---|---|---|
| Lawful basis and notice | Confirm the original consent or legitimate use covers analysis and system design before any sample moves | Consent text, purpose mapping note, sign-off from your data owner |
| Purpose limitation | Data sampled for a use-case study is used for that study only, never retained as training material | Written scope of processing in the statement of work |
| Data minimisation | Sample a few hundred records rather than dumping a table, and drop fields the question does not need | Sampling script and field list, with exclusions named |
| Storage limitation | A deletion date fixed at kickoff, performed on schedule and confirmed in writing | Deletion certificate and log entry, countersigned |
| Security safeguards | Encryption in transit and at rest, named individuals with access, no personal laptops, no shared logins | Access list, key management note, device policy |
| Processor obligations | A written contract naming the processor, every sub-processor and the model providers involved | Signed data processing agreement and sub-processor register |
| Breach reporting | An agreed path to notify you within hours so you can meet your own reporting duty | Incident runbook with named contacts and stated timings |
Do you need to send production data to a consultant?
Usually not, and the default answer should be no. A strategy engagement needs to know whether the data can support the use case; it does not need to hold the data. Four handling tiers cover almost every question we are asked to answer, and most engagements stop at the second.
- Schema and statistics only. Column names, types, null rates, cardinality and distributions, with no rows at all. This is enough to answer most feasibility questions and to size a retrieval index.
- Redacted sample inside your boundary. A few hundred rows with direct identifiers removed, reviewed on your infrastructure over screen share and never copied out. PII redaction is applied before anyone reads a record.
- Pseudonymised export under a signed agreement. Identifiers replaced with stable tokens, a deletion date fixed at the start, and access limited to named engineers on managed devices.
- Processing inside your VPC. Models and code run where the data already lives, with nothing leaving the boundary. The pattern is described in zero data egress.
- Full production access. Rare, and justified only for regulated workloads where the strategy must prove behaviour on real records. It requires the strictest controls on this list and an explicit board-level decision.
Where the data is processed, and why residency is not the whole question
Residency answers where data sits at rest. It does not answer who can reach it. An Indian region on a hyperscaler with a support team that can open a ticket and read your bucket satisfies a residency box and very little else. Decide both questions: the region, and the access list.
Model calls are the leak most often missed. If a consultant pastes a sample of customer records into a hosted model to test an idea, that is a cross-border transfer and a purpose change in a single action. Agree the model policy before the first working session: which providers are approved, whether zero-retention terms are in place, and what may never be sent at all. Sector rules stack on top, and for regulated lenders the RBI outsourcing guidelines are stricter than DPDP on audit and location.
Where the work has to stay entirely inside your boundary, the answer is self-hosted. Agentic AI running on your own infrastructure starts at $31,500 or ₹20,80,000 plus infrastructure, and the strategy phase should say plainly whether that constraint applies, because it changes the build budget substantially.
What does compliant AI strategy consulting cost?
The compliance work sits inside the engagement price rather than arriving as a surcharge. AI product strategy and use-case discovery starts at $4,250 or ₹2,80,000 for a two to four week engagement, and a ten-day AI discovery sprint is $3,250 or ₹2,00,000, fixed and credited to your next build. Both include the data handling plan, the deletion schedule and the model policy. Starting prices for every service are published on the pricing page.
We sign NDAs before the first working session, and you own the code, prompts, infrastructure and documentation that come out of the work. Indian clients are invoiced in INR with GST. The controls behind all of this are set out on the security page.
Controls we run on every engagement
Access that expires
Named individuals, time-boxed credentials, and removal at handover rather than at some unspecified point afterwards. Shared logins are the fastest way to lose the ability to answer the question of who saw what.
An audit trail from the first day
Log who accessed which sample, when, and why. Regulators and boards ask the same question in different words, and the answer has to be a record rather than a recollection. AI audit trails covers what that record needs to contain.
Deletion you can prove
A deletion date set at kickoff, executed on schedule, and confirmed in writing with every storage location named, including backups and notebook caches. Unproven deletion is the most common gap we find when we take over work someone else started.
When this checklist is the wrong tool
If the engagement will never touch personal data, most of this is overhead. A strategy exercise on demand forecasting from aggregated sales, on document templates with no customer content, or on public policy corpora carries different risks, mainly confidentiality and intellectual property, and a tight NDA plus a short access list covers them.
It is also the wrong tool if compliance is being used as a reason not to start. Schema-level discovery, synthetic data and redacted samples let you answer feasibility questions without moving a single identifier. Organisations that wait for a perfect governance programme before any AI work usually wait a year and learn nothing in the meantime.
And a checklist is not a legal opinion. For regulated sectors, or where cross-border transfer is unavoidable, involve counsel early. We will tell you when the question in front of you is legal rather than architectural, because getting that wrong is expensive in a way that rework is not.
A worked example
An NBFC asked us to look at KYC and loan onboarding documents, which are about as sensitive as Indian data gets: identity documents, addresses and income proof. The discovery work ran on a redacted sample inside their own environment, and the production system was built so that those documents never left their infrastructure.
Exceptions went to a human queue with a full audit trail rather than being silently auto-approved, because the audit requirement was treated as a design input rather than a later addition. The engagement is described in the KYC document intelligence case study.
The checklist
- Confirm in writing that the lawful basis covers analysis and design, before any sample moves
- Sign the NDA and the data processing agreement before the first working session
- Name every sub-processor and every model provider the engagement will use
- Choose the lowest data tier that answers the question, and record why that tier was enough
- Fix the deletion date at kickoff and diarise the confirmation, including backups
- Agree which categories of data may never be sent to a hosted model, and write it down
- Log access by named individual from day one, not from go-live
- Set the breach notification path with names and hours, not a generic inbox
Related reading
DPDP Act 2023 and AI sets out what Indian companies must do across the whole lifecycle, a security questionnaire for AI vendors gives the questions to send before you sign anything, and sovereign AI in India explains when a fully self-hosted stack is genuinely required rather than merely reassuring.
Compliance in a strategy engagement is decided in the first week, by how little data you agree to move, not in the last week by a paragraph in the report.
Frequently asked questions
Does the DPDP Act apply to an AI strategy consulting engagement?
▾
It applies to the personal data the engagement touches, not to consulting as an activity. The moment a consultant sees customer names, phone numbers, transcripts or loan files, you are the data fiduciary and they are a processor acting on your instructions. Your obligations on purpose, retention, security and breach reporting follow the data into the engagement.
Can a consultant put our customer data into a hosted AI model?
▾
Only if your lawful basis covers it, the provider is named in your processing agreement and zero-retention terms are in place. Most feasibility questions need no personal data at all, so the safer default is schema and statistics first, then a redacted sample reviewed inside your own environment.
Do we have to keep AI strategy data inside India?
▾
The DPDP Act does not impose blanket localisation, though the government may restrict transfers to specified countries and sector regulators impose stricter rules. Reserve Bank of India expectations for regulated lenders are tighter than the Act. Decide residency and the access list together, because a local region with wide support access proves very little.