Artificial intelligence is rapidly moving beyond systems that merely generate responses to user prompts. The emergence of autonomous AI and agentic AI systems represents a significant technological shift: AI systems can increasingly plan tasks, use external tools, access databases, interact with software, make decisions and execute actions with limited human intervention.

This evolution creates a corresponding legal problem: who should be liable when an autonomous or agentic AI system causes harm?

Traditional liability frameworks generally assume that a human being or legally recognised entity makes a decision, performs an act or controls a product. Agentic AI complicates that assumption because the causal chain may involve a foundation-model developer, application provider, system integrator, deployer, user, third-party tool provider and the AI system itself.

The central legal challenge is therefore not simply whether AI should be “liable”. AI systems are generally not recognised as independent legal persons capable of bearing legal responsibility in their own right. The more important question is how existing doctrines—and potentially new AI-specific rules—should allocate responsibility among the human and corporate actors surrounding the system.

The European Union’s Artificial Intelligence Act (“EU AI Act”), the EU’s revised Product Liability Directive and existing national liability laws illustrate an emerging approach based on risk management, traceability, human oversight, product safety and responsibility across the AI value chain. In India, the legal position remains more fragmented, with potential liability arising under areas including consumer protection, product liability, data protection, contract, tort and information technology law.

This article examines the principal liability frameworks applicable to autonomous and agentic AI systems and considers how legal systems can preserve innovation while ensuring meaningful accountability.

1. Understanding Autonomous AI and Agentic AI

Before analysing liability, it is necessary to distinguish between conventional AI, autonomous AI and agentic AI.

1.1 Conventional AI

A conventional AI system generally performs a defined function based on inputs provided by a human or another system. Examples include:

image classification;
recommendation systems;
fraud detection;
predictive analytics;
automated translation; and
generative AI chatbots.

Although these systems can produce unexpected results, the user’s interaction with the system is often relatively direct.

1.2 Autonomous AI

An autonomous AI system is capable of operating with varying degrees of independence from direct human intervention.

The EU AI Act itself defines an AI system by reference to a machine-based system designed to operate with varying levels of autonomy and potentially exhibit adaptiveness after deployment.

Autonomy, however, should not be confused with legal personhood. A system may act autonomously in a technical sense while responsibility remains legally attributable to the people and organisations that designed, supplied, deployed or controlled it.

1.3 Agentic AI

Agentic AI goes further. An AI agent may be capable of:

interpreting an objective;
developing a plan;
selecting among possible actions;
calling external tools or APIs;
retrieving information;
executing transactions or operations;
evaluating intermediate results; and
modifying its subsequent actions.

The legal significance of agentic AI lies in the increased distance between human instruction and eventual action.

A user may give an agent a broad objective without specifying every individual action the system ultimately takes.

That creates an important liability question:

If the user authorised the objective but did not specifically authorise the harmful act, who bears responsibility for the resulting damage?

2. Why Existing Liability Models Are Under Pressure

Traditional liability law often relies upon concepts such as fault, causation, control, foreseeability, defective products and reasonable care.

Agentic AI challenges each of these concepts.

2.1 The causation problem

Suppose an AI agent causes financial loss after independently selecting an incorrect course of action.

The causal chain could look like:

Developer → Foundation Model → AI Provider → Application → Tool/API → Deployer → Agentic Decision → Harm

The claimant may therefore face difficulty identifying which actor’s conduct legally caused the damage.

2.2 The foreseeability problem

A developer may argue that it could not reasonably predict the exact sequence of actions undertaken by an autonomous system.

However, the fact that a particular output was unpredictable does not necessarily mean that the resulting risk was unforeseeable.

Courts may increasingly distinguish between:

unpredictability of the exact outcome, and
foreseeability of the category of risk.

This distinction could become central to future AI litigation.

2.3 The control problem

Another difficult question is determining who had sufficient control over the AI system at the relevant time.

Potential actors include:

model developers;
AI application providers;
system integrators;
employers;
businesses deploying AI;
users;
tool providers; and
platform operators.

The greater the autonomy of the system, the less useful it may become to ask simply, “Who clicked the button?”

Instead, liability analysis may need to ask:

Who designed the system, who controlled the relevant risk, who had the ability to prevent the harm, and who benefited from deployment?

3. The Principal Actors in the AI Liability Chain

A comprehensive liability framework should recognise that AI systems operate within an ecosystem rather than in isolation.

3.1 Developers and model providers

Developers may be responsible where harm results from:

defective model architecture;
inadequate safety testing;
inadequate cybersecurity;
insufficient documentation;
known failure modes;
inadequate safeguards; or
negligent design choices.

The EU Product Liability Directive expressly recognises software developers and AI system providers within the product-liability framework, treating a developer or producer of software as a manufacturer for relevant purposes.

This is particularly significant because traditional product liability was historically associated with physical products.

3.2 AI application providers

An application provider may combine a third-party foundation model with:

proprietary instructions;
databases;
retrieval systems;
plugins;
APIs;
workflow automation; and
autonomous execution capabilities.

If the harmful behaviour results from this integration rather than the underlying model, liability may shift toward the application provider.

3.3 Deployers

A deployer is the entity that actually uses an AI system under its authority.

The EU AI Act expressly places obligations on deployers of high-risk AI systems, alongside obligations imposed on providers.

A business that deploys an AI agent therefore cannot necessarily avoid responsibility merely by arguing:

“The AI made the decision.”

If the organisation selected the system, configured it, determined its operational environment and failed to implement appropriate safeguards, ordinary principles of organisational responsibility may still apply.

3.4 Users

Individual users may bear responsibility where they:

misuse an AI system;
deliberately override safeguards;
provide unlawful instructions;
use an AI system outside its intended purpose; or
ignore warnings concerning known risks.

However, user responsibility should not become a mechanism for automatically transferring every AI-related loss to the end user.

3.5 Tool and infrastructure providers

Agentic AI frequently depends upon external tools.

For example, an agent may have access to:

payment systems;
cloud infrastructure;
email;
databases;
scheduling platforms;
enterprise software; or
external APIs.

This creates the possibility of distributed causation, where several actors contribute to the final harmful event.

4. A Multi-Layered Liability Framework

A future-ready framework for autonomous and agentic AI should not depend upon one universal liability rule.

Instead, liability should operate across multiple layers.

Layer 1: Contractual Liability

Contracts should allocate responsibility between:

model providers;
application developers;
enterprise customers;
cloud providers;
integrators; and
end users.

Important contractual provisions may address:

permitted use;
allocation of risk;
warranties;
indemnification;
limitation of liability;
audit rights;
incident reporting;
data responsibilities;
security obligations; and
termination rights.

However, contractual allocation cannot necessarily eliminate statutory or tortious liability toward third parties.

Layer 2: Fault-Based Liability

Fault-based liability remains appropriate where a human or organisation:

failed to exercise reasonable care;
ignored known risks;
failed to conduct adequate testing;
failed to implement safeguards;
deployed an unsuitable system; or
failed to supervise a high-risk system.

The advantage of fault-based liability is that it can distinguish responsible conduct from genuinely unforeseeable technological failure.

Its disadvantage is evidentiary difficulty.

A claimant may not know:

what the model was trained on;
which safety tests were performed;
what logs existed;
what system instructions were used;
what warnings were provided; or
which configuration caused the failure.

This makes access to evidence and disclosure obligations especially important.

5. Strict and Product Liability

Product liability may provide a more claimant-friendly mechanism in appropriate circumstances.

The EU Product Liability Directive represents an important development because its modernised framework expressly addresses software, AI systems and product-related digital services.

This matters because AI can be embedded within physical products, software products and digital services.

A defective autonomous system could therefore potentially be analysed similarly to other defective products, depending upon the applicable jurisdiction and circumstances.

6. The European Union Approach

The EU provides one of the most developed regulatory approaches to AI.

The EU AI Act establishes a risk-based regulatory framework covering prohibited AI practices, high-risk AI systems, transparency obligations, general-purpose AI models and governance mechanisms.

Importantly, the AI Act is primarily a regulatory compliance framework, rather than a comprehensive civil damages regime.

This distinction is crucial.

A violation of an AI regulatory obligation does not automatically answer every question concerning private compensation. Civil liability may still depend upon applicable national or EU liability rules.

6.1 High-risk AI and risk management

The EU AI Act imposes extensive obligations concerning high-risk AI systems.

These include requirements relating to:

risk management;
data governance;
technical documentation;
record-keeping;
transparency;
human oversight;
accuracy;
robustness; and
cybersecurity.

Providers of high-risk AI systems must also comply with specified conformity-assessment and quality-management requirements.

For agentic AI, these requirements provide an important regulatory philosophy:

The law should regulate not only the harmful outcome but also the risk-management processes that preceded it.

7. Product Liability and AI

The revised EU Product Liability Directive is particularly important for AI-related liability.

It modernises product liability for the digital age and expressly brings software and AI systems within its scope in relevant circumstances.

This can reduce one traditional obstacle to AI litigation: the argument that software is not a “product”.

The Directive also reflects the increasingly complex nature of digital supply chains.

An AI-enabled product may involve:

hardware + software + AI model + digital service + continuous updates.

The legal framework must therefore account for defects that emerge not merely at the moment of sale but through subsequent software updates and digital functionality.

8. The Withdrawn EU AI Liability Directive: An Important Lesson

The EU had proposed an AI Liability Directive designed to adapt non-contractual civil liability rules to artificial intelligence.

However, the legislative proposal was withdrawn in October 2025.

This development is legally significant.

It demonstrates that creating a dedicated AI civil-liability regime is considerably more difficult than establishing AI regulatory standards.

It also suggests that future AI liability may continue to rely heavily upon:

existing tort principles;
product liability;
contractual allocation;
sector-specific rules; and
evidentiary mechanisms,

rather than one comprehensive “AI liability law”.

9. The Indian Legal Position

India currently does not have a single comprehensive statute establishing a standalone civil liability regime specifically for autonomous or agentic AI.

Instead, liability may arise through several existing legal frameworks.

These include:

contract law;
tort law;
consumer protection law;
product liability;
information technology law;
data protection law;
intellectual property law; and
sector-specific regulation.

This fragmented structure means that the legal classification of the AI system and the nature of the harm become particularly important.

10. Consumer Protection and AI Liability in India

The Consumer Protection Act, 2019 contains a dedicated product-liability framework.

Section 82 provides for claims concerning harm caused by defective products, while Section 83 permits product liability actions against manufacturers, product service providers and product sellers.

Section 84 identifies circumstances in which a product manufacturer may be liable, including manufacturing defects, design defects, deviation from manufacturing specifications, failure to conform to express warranties and inadequate instructions or warnings.

This framework could become relevant to AI-enabled products where the AI functionality contributes to a legally recognised defect or harm.

The statute also imposes circumstances of liability upon product sellers and service providers. Section 86, for example, addresses situations involving substantial control over design or testing, alteration or modification, warranties and failure to exercise reasonable care.

Thus, Indian consumer law already contains concepts capable of being adapted to some AI-related disputes.

11. Information Technology Law and AI

The Information Technology Act, 2000 may also become relevant depending upon the nature of the AI system and the conduct involved.

For example, Section 43A historically provided for compensation where a body corporate negligently failed to implement and maintain reasonable security practices and thereby caused wrongful loss or wrongful gain in relation to sensitive personal data or information.

Section 79 provides conditional intermediary protection for certain third-party information, subject to statutory requirements.

However, an important legal distinction must be maintained:

An AI developer or agentic AI provider should not automatically be treated as an intermediary merely because its system processes third-party information.

The applicability of intermediary protection depends upon the statutory requirements and the actual functions performed.

12. Data Protection as a Separate Liability Layer

Agentic AI systems can process substantial amounts of personal information.

An autonomous agent may:

retrieve personal information;
combine information from multiple sources;
send information to third parties;
make decisions based on personal data; or
initiate actions affecting individuals.

Consequently, privacy and data-protection liability may arise independently from liability for the ultimate physical or economic harm.

This produces an important principle:

One AI incident may generate multiple independent causes of action.

For example, an AI agent’s unlawful processing of personal data could create a data-protection issue even if no physical injury occurs.

Conversely, physical or financial harm may occur without a data-protection violation.

13. The “Black Box” Problem and Burden of Proof

One of the most significant legal challenges concerns the opacity of AI systems.

A claimant may know:

“The AI caused the harmful result.”

But proving:

“The defendant’s legally actionable conduct caused the AI to produce that result”

may be considerably harder.

Agentic systems make this problem more severe because decision-making may occur through multiple steps.

For example:

Prompt → Planning → Tool selection → Data retrieval → Intermediate reasoning → Action → Feedback → Further action

A meaningful liability framework should therefore address evidentiary asymmetry.

Potential mechanisms include:

mandatory logging;
audit trails;
preservation of system records;
disclosure obligations;
technical documentation;
incident reporting;
model and system documentation; and
presumptions concerning causation in narrowly defined circumstances.

The EU AI Act’s requirements concerning documentation and logging for high-risk AI illustrate the importance of traceability.

14. Human Oversight and the “Human-in-the-Loop” Defence

A common regulatory response to AI risk is human oversight.

However, simply placing a human somewhere in the workflow should not automatically eliminate liability.

There is a significant difference between:

Meaningful human oversight

A human:

understands the system’s limitations;
receives relevant information;
has sufficient time to intervene;
possesses authority to override the system; and
can realistically prevent harm.

Formal human oversight

A human merely:

receives an automated recommendation;
clicks “approve”;
lacks technical understanding; or
has no practical ability to challenge the AI.

The second model risks becoming a legal fiction.

Therefore, courts and regulators should evaluate actual control rather than nominal human involvement.

15. Should AI Agents Be Given Legal Personhood?

One proposed solution to the accountability problem is to recognise autonomous AI systems as separate legal persons.

At present, this approach raises substantial legal difficulties.

Legal personality could potentially:

complicate attribution;
create uncertainty concerning ownership and control;
facilitate artificial shielding of developers;
create enforcement problems where the AI has no independent assets; and
weaken incentives for human actors to manage technological risks.

A stronger near-term approach is therefore responsibility without personhood.

The law can recognise technological autonomy while continuing to assign legal responsibility to the natural or legal persons who create, deploy, control or commercially benefit from the system.

16. A Proposed Liability Matrix for Agentic AI

A practical liability framework could operate as follows:

Actor	Potential Liability Trigger	Relevant Legal Principle
Model developer	Defective model, inadequate safeguards, known failure modes	Negligence / product liability
AI provider	Unsafe system design or deployment	Regulatory / contractual / tortious liability
Integrator	Faulty integration or configuration	Negligence / contract
Deployer	Improper deployment or inadequate supervision	Organisational negligence
User	Misuse or unlawful instructions	Fault-based liability
Tool provider	Defective external functionality	Contract / product / negligence
Product manufacturer	AI-enabled defective product	Product liability
Seller/service provider	Defective service or inadequate warnings	Consumer / product liability
Employer	Employee deployment and supervision failures	Vicarious / organisational liability

This model recognises that AI liability should follow the distribution of risk and control rather than the mere location of the final automated decision.

17. A Risk-Based Liability Model for Autonomous AI

A particularly promising framework would classify AI systems according to the seriousness of potential harm.

Low-risk systems

Examples may include:

writing assistance;
low-impact productivity tools;
basic recommendation systems.

Liability could primarily remain under ordinary contract and negligence rules.

Medium-risk systems

Examples could include:

financial workflow automation;
enterprise decision-support tools;
automated customer-management systems.

These systems may require stronger logging, contractual allocation and auditing.

High-risk autonomous systems

Examples may include systems affecting:

healthcare;
employment;
financial access;
critical infrastructure;
safety-sensitive operations; or
significant legal rights.

Such systems should attract stronger requirements concerning:

testing;
monitoring;
human oversight;
documentation;
incident reporting;
cybersecurity; and
insurance or financial responsibility.

The EU AI Act’s risk-based architecture provides an important comparative model for this approach.

18. Cybersecurity and Agentic AI Liability

Agentic AI introduces a further complication: an AI system can itself become an attack surface.

An attacker might manipulate:

prompts;
retrieved information;
external tools;
credentials;
APIs;
databases; or
connected applications.

This raises the question of whether a developer or deployer should be liable where an agent is compromised.

The answer should depend upon factors such as:

whether the vulnerability was reasonably foreseeable;
whether reasonable security measures were implemented;
whether the vulnerability was known;
whether security updates were provided;
whether the deployer followed security instructions; and
whether a third party’s conduct was an intervening cause.

The EU’s broader digital regulatory framework increasingly connects AI safety with cybersecurity. The EU Cyber Resilience Act, for example, requires manufacturers of products with digital elements to address cybersecurity risks across design, development, production and maintenance.

19. Insurance and Financial Responsibility

As autonomous systems become more capable, insurance may become an important complement to liability rules.

Potential products include:

AI professional liability insurance;
technology errors-and-omissions insurance;
cyber insurance;
product liability insurance; and
specialised autonomous-system coverage.

For high-risk AI, regulators could eventually require evidence of financial capacity or insurance.

Such a mechanism would not determine legal fault. Instead, it would ensure that victims have a realistic source of compensation when several actors contribute to an AI-related loss.

20. Challenges in Cross-Border AI Liability

Agentic AI systems are often inherently international.

A single system may involve:

Developer in Country A → Cloud provider in Country B → User in Country C → Harm in Country D

This creates questions concerning:

jurisdiction;
applicable law;
enforcement of judgments;
contractual choice of law;
cross-border evidence;
data-transfer restrictions; and
regulatory cooperation.

AI liability therefore cannot be considered exclusively as a domestic legal issue.

International interoperability will become increasingly important.

21. Key Legal Principles for Future AI Liability

A mature liability framework for autonomous and agentic AI should be based upon several principles.

21.1 Accountability should follow control

The person or entity capable of controlling a relevant risk should generally bear an appropriate share of responsibility.

21.2 Autonomy should not equal immunity

The fact that an AI system acted autonomously should not by itself absolve its developers or deployers.

21.3 Liability should be proportionate to risk

Low-risk applications should not face the same compliance burden as systems capable of causing serious physical, financial or fundamental-rights harms.

21.4 Evidence must be accessible

AI providers should maintain sufficient records to reconstruct significant automated decisions and incidents.

21.5 Contract should supplement—not replace—public law

Private agreements can allocate risk between commercial actors but should not automatically eliminate statutory protections available to consumers or third parties.

21.6 Human oversight must be substantive

A nominal human approval mechanism should not become a shield against responsibility.

22. The Way Forward for India

India’s rapidly expanding AI ecosystem makes the development of a coherent liability framework increasingly important.

Rather than immediately adopting an entirely separate AI liability statute, India could consider a layered framework built upon existing law.

Such a framework could include:

clarification of how AI-enabled software interacts with product liability;
sector-specific liability rules for high-risk autonomous systems;
mandatory incident and audit records for specified high-risk applications;
stronger contractual standards for AI procurement;
clear allocation of responsibilities among developers, deployers and integrators;
mechanisms for preserving AI-generated evidence;
regulatory guidance concerning human oversight;
cybersecurity obligations for autonomous systems; and
appropriate compensation mechanisms for affected consumers.

India’s existing Consumer Protection Act already provides a useful starting point because it recognises product liability and identifies responsibilities across manufacturers, sellers and service providers.

The objective should not be to eliminate technological risk entirely—a practically impossible task—but to ensure that the party best positioned to prevent or mitigate a particular risk bears an appropriate legal responsibility.

23. Conclusion

The rise of autonomous and agentic AI challenges one of the foundational assumptions of traditional liability law: that a human actor can be readily identified as the immediate decision-maker.

Agentic AI can transform a simple instruction into a complex chain of autonomous decisions and actions. As a result, legal responsibility cannot depend solely on identifying the last human who interacted with the system.

The emerging approach should instead focus on risk, control, foreseeability, causation, traceability and the distribution of responsibility across the AI value chain.

The EU’s AI Act demonstrates the growing importance of risk-based governance, while the revised EU Product Liability Directive illustrates how traditional product-liability principles are being adapted to software and AI. The withdrawal of the proposed EU AI Liability Directive also demonstrates the difficulty of constructing an entirely new civil-liability regime specifically for AI.

For India, the immediate challenge is not necessarily the creation of a completely separate body of AI law. Existing principles under consumer protection, product liability, contract, tort, information technology and data protection can provide much of the foundation. What is required is greater clarity concerning how these principles apply when an AI system is capable of acting autonomously.

Ultimately, the guiding principle should be simple:

AI may act autonomously, but accountability should remain humanly attributable.

A legally sustainable AI ecosystem must therefore ensure that technological autonomy does not create an accountability vacuum. The objective of AI liability law should be neither to punish innovation nor to impose absolute liability for every AI failure, but to establish a predictable and proportionate system in which those who design, control, deploy and profit from autonomous systems also bear appropriate responsibility for the risks they create.

Frequently Asked Questions

What is AI liability?

AI liability refers to the legal responsibility of individuals or organisations for harm caused by the development, deployment, operation or misuse of an artificial intelligence system.

Who is liable when an AI agent causes harm?

Liability depends upon the circumstances. Potentially responsible parties include the AI developer, model provider, application provider, integrator, deployer, product manufacturer or user. The appropriate defendant depends on factors such as control, fault, causation, contractual obligations and applicable product-liability rules.

Are autonomous AI systems legally responsible for their own actions?

Generally, technological autonomy does not itself create legal personality. Current liability systems generally attribute responsibility to legally recognised persons or entities associated with the AI system.

Does the EU AI Act create AI civil liability?

The EU AI Act primarily establishes regulatory obligations and risk-based requirements for AI systems. Civil compensation questions may also depend on applicable product-liability and national civil-liability rules.

Does India have a specific AI liability law?

India does not currently have a single comprehensive statute dedicated exclusively to civil liability for autonomous or agentic AI. Existing legal frameworks, including consumer protection and product liability, may apply depending on the facts.

Why is agentic AI more difficult to regulate than ordinary generative AI?

Agentic AI can independently plan and execute multiple actions using external tools. This increases the complexity of determining causation, foreseeability, control and responsibility.

Keywords

Primary:
Liability Frameworks for Autonomous & Agentic AI Systems

Secondary:
AI liability, agentic AI liability, autonomous AI liability, AI legal framework, AI accountability, AI regulation, AI product liability, AI developer liability, AI deployer liability, AI governance, autonomous systems law, agentic AI regulation, AI liability India, EU AI Act, AI law India.

Long-tail keywords:

Who is liable when an AI agent causes harm?
legal liability for autonomous AI systems
liability framework for agentic AI
AI product liability in India
autonomous AI legal responsibility
developer liability for AI systems
deployer liability under AI law
EU AI Act and AI liability
AI accountability framework in India
legal challenges of agentic AI

Table of Contents

Introduction