
Introduction
As AI agents move beyond generating responses to autonomously accessing databases, interacting with third-party tools and making decisions involving personal data, determining who bears legal responsibility for their actions becomes increasingly complex. Under India’s Digital Personal Data Protection Act, 2023 (“DPDP Act”), accountability rests on the role of the Data Fiduciary, the purpose and means of processing, and the obligations imposed on those who process data on its behalf.
However, when an AI agent operates through a cloud-hosted model, an external API or an organisation’s own infrastructure, the extent of control over the system can significantly influence the allocation of risk, the effectiveness of safeguards and the ability to establish accountability following a breach.
This raises a critical question: when an AI agent acts autonomously, who is legally responsible for the personal data it processes, and how does the architecture on which it operates shape that liability? The challenge, therefore, is not merely to identify who deploys an AI agent, but to determine how control over the model, its operating environment and its access to personal data shapes the resulting liability.
From Purpose and Means to the Fiduciary Test
For decades, enterprise technology liability rested on two elements: the purpose set by the organisation and the means supplied by technology providers. The foundation models add a third element: a general and probabilistic capability that the organisation did not author, and that an agent now chooses intermediate steps, selects tools, and improvises a path to a high-level goal.
The first two elements are close to the statutory test in the DPDP Act. A Data Fiduciary is a person who alone or jointly with others determines the purpose and means of processing personal data, and a Data Processor is a person who processes data on its behalf. The third element does not change the answer. An agent is not a person, and processing by a wholly or partly automated system is still processing by whoever put it to work. The organisation that sets the objective and gives the agent its credentials and tools has determined the purpose and the means, even if it did not foresee the particular step.
Section 8(1) of the DPDP Act then makes the Data Fiduciary responsible for compliance irrespective of any contract to the contrary and irrespective of processing carried out by a Data Processor on its behalf. In practice, this is a statutory non-delegable duty, and the decision to use a system at all stays with the deployer.
The Value Chain Through the DPDP Lens: Fiduciary, Processor or Both
The positions attach to roles and not to companies, so that one organisation can hold several positions at once. The same discipline is needed under the DPDP Act. An end-user organisation deploying an agent on its customer or employee data will ordinarily be the Data Fiduciary. An integrator that configures the system and a cloud provider that hosts it will usually be Data Processors, provided that they act on the organisation’s instructions under a valid contract, which Section 8(2) requires.
A product company that embeds an agent in its software is a Data Processor for its customers’ data but may be a Data Fiduciary for data it uses for its own purposes, for example to improve its product. The model provider’s position depends on what it does with the prompts. Where it processes them only to return a response, and in particular under zero-data-retention terms, it is closer to a Data Processor or to a party that holds no data at all.
Where it retains prompts or outputs to train its own models, it is pursuing its own purpose and may be a Data Fiduciary in its own right, with the user organisation also answerable for having sent the data. These are our views, since the Act and the Rules do not address AI value chains, and the contractual terms will in many cases decide the question.
Why Where the Model Runs Decides Liability
Where a closed model is consumed through a public cloud API, liability is shared across the value chain, the end user is the controller of the data, and the evidence needed to defend a claim sits partly with the providers. The silent model updates are possible, which matters because a change in behaviour can alter how personal data is processed without any change on the deployer’s side.
Country-specific or sovereign hosting keeps the sharing but largely removes residency exposure. A model deployed in the organisation’s own virtual private cloud moves most data protection liability to the organisation as the controller of the environment, and a private cloud or on-premises deployment of an open model places it almost wholly on the organisation, which becomes in effect the model provider.
Several DPDP provisions respond to those differences. The security safeguards duty in Section 8(5), with the access control, logging and monitoring measures in Rule 6, falls increasingly on the organisation as it takes over the runtime. Cross-border questions are governed by Section 16, under which transfers are permitted except to countries the Central Government restricts, so a public API hosted abroad is not unlawful in itself. Section 16(2) preserves stricter sectoral requirements, such as localisation directions for regulated financial data, and a Significant Data Fiduciary may be required under Rule 13 to keep specified data in India. A foreign provider serving Indian users also falls within the Act’s reach under Section 3(b) when it offers services to people in India.
Additionally, an open-source model lets the organisation audit the training record for personal data, and failing to do so is a failure of its own care. An open-weight model leaves the organisation blind to training origins and unable to check the weights for memorised data, and this is an asymmetric risk, because the organisation absorbs full hosting and modification liability while having little recourse upstream. The central paradox applies equally to data protection: the more of the stack an organisation controls, the more of the failure it owns.
Agent-Specific Exposures Under the Act
The traditional software analysis did not face, and several of them translate into DPDP Act exposure. Data that becomes instruction, such as an email or web page that redirects the agent, can lead to personal data being disclosed or moved to an unintended destination. That is a compromise of confidentiality and so, a personal data breach within Section 2(u), engaging Section 8(5) and the intimation duties in Section 8(6) and Rule 7.
Delegation across agents and tools, where permissions granted once travel further than intended, raises the question whether safeguards are reasonable when authority is not narrowed as tasks pass along a chain. Autonomous action with probabilistic judgement engages the accuracy duty in Section 8(3), which applies where personal data is used to make a decision that affects the Data Principal. Non-reproducible behaviour makes an incident difficult to replay, so that the record of what the agent saw and did becomes the main evidence.
Accountability, Evidence and the Gap Between Intent and Control
The distinction between an intent, the decision space the architecture leaves open, and the instruments that turn a decision into a hard constraint. A policy that an agent must not access certain customer fields is only intent until the platform allows a permission scope to be set and someone configures it. The investigators ask whether the organisation could have made the outcome happen, and whether it did, fits the way a regulator would approach Section 8(5) and Rule 6. The Data Protection Board, in setting a penalty under Section 33(2), weighs the nature and gravity of a breach, the type of data and the mitigation taken.
A control that was written down but never built or switched on shows that the risk was known and left open. The value-chain players who withhold the instruments needed to enforce policy expose themselves to responsibility as well. The hierarchy of controls, which prefers designing risk out, constraining permissions and confirming material actions before resorting to a kill-switch, is a useful template for what reasonable safeguards look like in an agentic system, and halting an agent mid-transaction can itself cause harm.
Allocating Risk by Contract
Because the statutory duty stays with the Data Fiduciary, contracts allocate cost and not responsibility, and recovery from upstream parties is limited by contractual caps and, for open models, is often unavailable. Contracts should therefore be drafted to give the Data Fiduciary what it needs to comply.
Useful terms include a prohibition on using prompts, outputs and tool data for training, zero-data-retention or defined retention periods, erasure on termination consistent with Section 8(7)(b), notice of model changes and the right to pin versions, notice of sub-processors, breach notification to the Data Fiduciary within a period that allows Rule 7 to be met, audit and cooperation rights in an investigation, data residency commitments, and the reasonable security safeguards that Rule 6 expects to be written into processor contracts.
AMLEGALS Remarks
The central message is that liability follows control, and the DPDP Act points the same way while placing the first claim on the Data Fiduciary. Indian organisations adopting agents would be well advised to inventory every route by which AI reaches their personal data, to record which position they and each vendor hold for each use, to document the hosting and model-type choice and the reasons for it, and to translate data policies into enforced permissions and thresholds in place of written rules alone. Contracts with model providers, cloud platforms and integrators should be reviewed against Section 8(2), Rule 6 and the breach intimation timetable in Rule 7, and incident plans should assume that the organisation will have to explain what the agent saw, decided and did.
For any queries or feedback, feel free to connect with mridusha.guha@amlegals.com or Khilansha.mukhija@amlegals.com
