Mayur PatilContact

Last checked 2026-11-01

AI agents in India

India has no AI-specific statute yet. The government's guidelines are voluntary, and the binding duties for an agent come mainly from the data protection law and from the rules of sector regulators. For a global capability centre in India, the practical questions are usually about personal data, and about how the centre shows control over what it builds for its parent. For a financial firm, the questions are about governance, customer rights and explainability.

What buyers here ask me first

  • Headquarters has asked the centre to do agentic AI with the same team. How does forward-deployed engineering work inside a capability centre without creating a second delivery function?
  • Readiness for the data protection duties before they apply. What must be in place first?
  • Moving from offshore delivery to owned outcomes, where the centre runs what it builds.
  • Which operating model suits a single centre that serves several business units.

The data protection law

The Digital Personal Data Protection Act, 2023 and its Rules, notified in November 2025, apply to the processing of digital personal data in India. In some cases they also apply to processing outside India in connection with offering goods or services to people in India. Most of the obligations apply after a transition period of eighteen months, which places them around May 2027. The exact date should be taken from the official notification.

For an agent, the duties attach to what it reads, writes and logs. The processing needs a lawful basis, which is usually consent or another basis the Act recognises. People must be given notice of what is collected and why. People must be able to exercise their rights, including correction and erasure, through a route that works. The data must be protected with security safeguards, and a breach must be handled according to the Act.

The governance guidelines

MeitY's India AI Governance Guidelines, released on 5 November 2025, set seven principles, called sutras, and six pillars. They are voluntary. They say that a separate AI law is not needed at this stage and that existing laws should be applied consistently. The guidelines shape what procurement and risk teams ask for, and a vendor that can answer those questions in writing has an advantage.

The guidelines name the same kinds of concern that a data protection review raises: accountability for outcomes, transparency about the system's behaviour, and safety. An agent built to answer those concerns in its design is easier to defend than one that answers them after an incident.

Financial firms

The Reserve Bank of India's FREE-AI Committee report, published on 13 August 2025, sets out 26 recommendations for AI in the financial sector. The report covers commercial banks, cooperative banks, non-banking financial companies and payment system operators. Commentary on the report describes several recommendations that matter to an agent. A regulated firm should adopt an AI governance policy approved by its board. Customers should be told when they are dealing with AI, and they should have a way to challenge an AI-based decision. Models used for credit, underwriting and customer service should be explainable, auditable and free of discriminatory bias. Incidents should be reported in standard formats, and the report proposes a standing committee on AI within the regulator.

The report's recommendations are not all binding. Commentary describes some as changes to existing directions and others as proposals for new guidance. The report itself should be read for the exact wording and the status of each recommendation before a firm relies on it.

How an agent is built to meet these expectations

The data protection duties and the governance guidelines ask for the same things. An agent built for them is easier to defend, and it is easier to show a regulator or a client that the controls exist.

Context and data. Keep a record of the personal data each agent uses and the purpose for each use. Limit the agent to the data the task needs. Where data is retrieved for an answer, the retrieval should respect the same access rules as the source system.

Rights and notice. Build a route for rights requests into the workflow, not as a separate manual process. Where a customer is talking to an agent, the agent should say so, and the customer should be able to reach a person.

Tools and approvals. Each tool has a written contract with its permissions. An action that changes a customer or employee record requires an approver, and the approval is logged. For a financial firm, a decision that affects credit or service eligibility needs a reason the customer can understand, and the reason must be stored with the decision.

Evaluations and incidents. The evaluation set covers the cases that matter to the business owner, including cases where the outcome must not differ across groups. An incident process records what happened, what was done and what changed, in a form that can be reported.

The harness engineering term describes the parts around a model that carry most of this work.

GCCs and the operating model

A capability centre that builds agents for its parent needs to show who owns each agent after handover. The operating model answers that. In practice there are three common patterns. A central pool assigns engineers to projects. Embedded teams stay inside one business unit. A hub and spoke design keeps standards in a small central group and places teams in the business units. Each pattern has a different owner for each agent after launch.

The FDE operating model term describes these options, and the article on operating models for Indian GCCs sets out how I see them in practice. The handover pack term describes what the owner receives when the work passes to the centre.

Where the picture is still moving

The governance guidelines are voluntary and may be revised. The data protection rules come into full effect around May 2027, and their detail should be read against the notification. The RBI recommendations are at the level of a report and may be turned into formal directions, and that step should be watched for each regulated client.

What a first project looks like

The first project is one workflow, one owner and one set of personal data, chosen so that the controls can be tested in a few weeks. I start with a readiness review of two weeks. The first week maps the workflow, the personal data it uses and the rights a person could exercise over that data. The second week tests the riskiest assumption, which is usually the access model or the quality of the records.

If the project proceeds, the first working slice runs outside production on real data. For a financial client, the evaluation set includes cases where the explanation given to a customer must be understandable and accurate, and the test results are kept with the version of the system. The handover passes the agent to a named owner inside the centre, together with the runbook, the rights-request route, the incident process and the record of decisions. The owner then runs the agent for a period with the vendor on call but not operating it.

Where the sources disagree

The start date for the data protection duties is reported as around May 2027 in several sources, and the notified Rules are the reference for the exact date. The status of the RBI recommendations varies by commentary, from proposals to planned directions. The report itself should be read for the wording of each recommendation before a firm relies on it. The guidelines' seven sutras are described consistently across the official release and commentary, but the official release is the reference for their wording.

A composite example

[Composite] A capability centre serving a European parent wanted an agent to answer employee questions about benefits. The readiness review found that the questions included health information, which is sensitive personal data under the Act's framework. The centre limited the agent to the policy documents, blocked health details from the logs, and gave employees a route to a person. The parent's data protection team reviewed the design before the pilot began.

Common questions

Are the MeitY guidelines binding?

No. They are voluntary guidance. They still shape what procurement and risk teams ask for, and a design that answers their questions is easier to approve.

What should a GCC have in place before the data protection obligations apply?

A record of the personal data each agent uses, the purpose for each use, the basis for the processing, a working route for rights requests, and security safeguards. Confirm the full list against the notified Rules.

Does the data protection law apply to processing done outside India?

In some cases, yes. The Act applies to processing outside India when it is connected with offering goods or services to people in India. A centre that serves only foreign clients should still check whether any Indian personal data is involved.

Does the RBI report apply to my company?

It applies to the regulated entities the report covers, such as banks and non-banking financial companies. If your client is a regulated firm, expect it to ask for the controls the report describes, even if the report does not bind you directly.

Can an FDE work inside a capability centre?

Yes. It is one of the most common places for the role in India. The operating model decides who reports to whom and who owns the system after handover.

What is the first step for a centre that is starting?

Pick one workflow, name its owner and map the personal data it touches. The readiness review then decides whether the workflow should proceed, be reframed or be paused.

How do I know which version of a guideline to follow?

Use the official notification or report and record its version and date in the design. Commentary changes quickly, and the design should be checked against the official text at each review.

Resources

Work with me

If you are deciding whether an agent is worth building for an Indian centre or enterprise, start with the AI Delivery Readiness Review. It is a two-week check on one workflow that ends in a written decision to proceed, reframe or pause. For teams that want to build the capability themselves, the AIXL Academy forward-deployed AI engineering cohort is the route, with its own schedule and terms on that page. I work with teams in India on engagements, and I travel for engagements. The work location is stated in each engagement. You can also send an enquiry.

Work with me in India

This page summarises public sources as of the date above. It is not legal advice. Where sources disagree, the page says so. Sources marked as commentary are secondary and should be checked against the official text before a decision.

Explainers on India

Explainers on the rules and the technology for this country will be listed here as they are published.

Sources

This page summarises public sources as of the date above. It is not legal advice. Where sources disagree, the page says so.

Enquiry

Tell me about the work

The use case, where it is blocked and who is involved. I reply with a suggested starting point. Or email hello@mayurpatil.ai.

Goes to my CRM and my inbox. No newsletter, no sharing. See Privacy (opens in a new tab).

Close