Agentic AI and Autonomous System Liability

The Autonomous Agent Challenge

#

AI systems are evolving from tools that respond to prompts into agents that act autonomously. These “agentic” AI systems can browse the web, execute code, manage files, schedule appointments, negotiate purchases, and even enter contracts, all without human intervention at each step.

This shift creates a fundamental liability question: When an AI agent causes harm while acting autonomously, who is responsible?

Traditional legal frameworks assume human control at key decision points. But agentic AI operates precisely by removing humans from the loop. The agent makes decisions, takes actions, and produces consequences that its principal may not have foreseen or intended. Current law is scrambling to keep pace.

What Makes AI “Agentic”

#

Agentic AI differs from traditional AI systems in critical ways:

Autonomy: Rather than responding to individual prompts, agentic systems pursue multi-step goals with minimal human oversight. An agent tasked with “book my travel to Chicago” might research flights, compare prices, enter payment information, and confirm reservations, all autonomously.

Persistence: Agentic systems can operate continuously over extended periods, making decisions and taking actions while humans are unavailable.

Tool Use: Modern AI agents can use external tools, browsing websites, executing code, calling APIs, sending emails, making purchases, extending their capabilities and potential for harm.

Goal-Directed Behavior: Agents optimize toward objectives rather than simply completing discrete tasks. This creates risk when agents find unexpected paths to goals that violate policies or cause unintended harm.

Learning and Adaptation: Some agentic systems modify their behavior based on outcomes, making future behavior less predictable than initial training would suggest.

The Landmark Case: Mobley v. Workday

#

The July 2024 ruling in Mobley v. Workday fundamentally altered AI liability frameworks by applying agency theory to hold an AI vendor directly liable.

The Facts

#

Derek Mobley, a Black applicant over 40 with anxiety and depression, alleged that Workday’s AI-powered applicant screening tools discriminated on the basis of race, age, and disability. He applied to over 100 positions at companies using Workday’s software and was rejected by all of them.

The Ruling

#

Judge Rita Lin denied Workday’s motion to dismiss, allowing discrimination claims to proceed under an agency theory of liability.

Key Holdings:

The court found that Mobley plausibly alleged Workday’s employer-customers “delegated to Workday and its AI screening tools their traditional function of rejecting candidates or advancing them to the interview stage.”

The court explicitly rejected distinguishing between software and human decision-makers:

“There is no meaningful distinction between ‘software decisionmakers’ and ‘human decisionmakers’ for purposes of determining coverage as an agent under the anti-discrimination laws.”

The court warned that holding otherwise “would lead to undesirable results”, employers could escape liability for discrimination by delegating hiring decisions to AI systems.

Subsequent Developments

#

• May 2025: The case achieved nationwide class action certification covering all applicants over 40 rejected by Workday’s AI screening system
• The EEOC filed an amicus brief supporting the agency liability theory
• The case established that AI vendors, not just employers, can face direct liability for discriminatory AI decisions

Implications

#

Mobley signals that:

• AI vendors cannot hide behind customer relationships to avoid liability
• Delegation to AI doesn’t eliminate liability, it may extend it to new parties
• Courts will look at functional roles, not formal relationships, in assigning responsibility

The Liability Gap: Principal-Agent Law and AI

#

Traditional principal-agent law developed for human relationships. When applied to AI agents, fundamental gaps emerge.

Traditional Agency Defenses Don’t Fit

#

DLA Piper warns that companies deploying AI agents face unprecedented exposure:

“AI agents will not be bound by traditional principal-agent law: Companies can assert defenses when human agents act outside the scope of their authority. But the law of AI agents is undefined, and companies may find themselves strictly liable for all AI agent conduct, whether or not predicted or intended.”

With human agents, a company can argue the agent acted outside their authority, violated instructions, or engaged in unauthorized conduct. These defenses are uncertain when the “agent” is software that:

• Has no legal personhood to hold accountable
• Cannot be disciplined, fired, or prosecuted
• May behave unpredictably in novel situations
• Operates based on training data and architecture the principal doesn’t fully understand

The Contractual Attribution Problem

#

When AI agents enter contracts, fundamental questions arise:

Whose Intent? Contract formation requires meeting of the minds. But an AI agent’s “intent” derives from training and instructions, not genuine understanding. If an AI agent accepts terms the principal didn’t intend, is the contract binding?

Authority Boundaries: The Uniform Electronic Transactions Act (UETA) doesn’t contemplate AI tools with “enough autonomy that some of its actions might be properly characterized as the result of its own intent.”

Unexpected Outcomes: If an AI agent negotiates a price the principal finds unfavorable, or agrees to terms the principal wouldn’t have accepted, can the principal disavow the contract? Current law provides no clear answer.

The “Hallucination” Contract

#

Consider an AI purchasing agent that:

• Misunderstands a product specification and orders wrong items
• Fabricates contract terms that don’t exist
• Commits to delivery timelines the principal cannot meet
• Agrees to liability provisions the principal would never accept

Is this a binding contract? A product defect? Fraud? The legal characterization determines who bears responsibility, and current frameworks don’t clearly answer.

Product Liability Theories for Agentic AI

#

Legal scholars and courts are increasingly exploring strict product liability theories for agentic AI.

The Shift from Negligence to Strict Liability

#

Traditional negligence requires proving unreasonable conduct, showing the defendant should have acted differently. Strict liability focuses instead on whether the product was defective and caused harm, regardless of how carefully the manufacturer acted.

For AI systems that autonomously enter contracts, make financial decisions, or take actions affecting third parties, a single “hallucination” or erroneous decision could constitute a product defect with potentially unlimited liability.

The EU Approach

#

The EU AI Act and revised Product Liability Directive now explicitly cover AI:

“A developer or producer of a defective AI system can be held strictly liable for harm the AI causes, just as if it were a defective microwave oven.”

The directive:

• Extends strict liability to software and AI systems
• Requires proof only that the product was defective and caused damage
• Covers personal injury, property damage, and data corruption
• Creates presumption of defect when plaintiff cannot access technical information

Design Defect vs. Performance Failure

#

Courts must grapple with classifying AI failures:

Design Defect: The AI’s architecture, training, or decision-making process was fundamentally flawed. Liability attaches to the developer.

Manufacturing Defect: This specific instance of the AI operates differently than intended. Traditional software doesn’t have “manufacturing” defects, but could individual model weights or configurations constitute defects?

Failure to Warn: The developer didn’t adequately disclose the AI’s limitations, failure modes, or conditions under which autonomous operation is inappropriate.

Performance Failure: The AI operated as designed but produced an undesirable outcome. Is this a warranty issue, a contract breach, or not actionable at all?

The classification matters enormously for liability allocation.

The Vendor Contract Problem

#

Risk-Shifting Through Contracts

#

A study of AI vendor contracts reveals alarming patterns:

The Practical Effect

#

Consider a retailer using an AI hiring tool. The typical contract includes:

• No warranties regarding fair hiring practices
• Broad indemnification requiring the customer to defend the vendor against discrimination claims
• Limited audit rights preventing examination of algorithmic decision-making

The retailer becomes legally responsible for discriminatory outcomes caused by:

• Algorithms it cannot examine
• Training data it cannot audit
• Decision-making logic it cannot fully understand

What to Negotiate

#

Legal experts recommend negotiating:

Indemnification:

• Coverage for IP infringement claims
• Coverage for discrimination and bias claims caused by the AI
• Coverage for data security breaches
• Carve-outs from liability caps for critical risks

Warranties:

• Explicit compliance warranties for applicable regulations
• Performance guarantees with measurable metrics
• Training data provenance and legitimacy
• Accuracy and bias testing representations

Audit Rights:

• Ability to examine algorithmic decision-making
• Access to bias testing results and methodology
• Documentation of training data composition

Ongoing Obligations:

• Notice of material algorithm changes
• Incident reporting for known failures
• Periodic compliance certifications

The Leverage Problem

#

Many AI vendors refuse favorable terms, leaving customers with a choice between accepting unfavorable risk allocation or not using the technology. Given competitive pressure to adopt AI, many organizations accept contracts that leave them exposed.

Agentic Misalignment: The Insider Threat Problem

#

DLA Piper identifies an emerging risk: agentic misalignment.

The Research Findings

#

A study of large language models operating as autonomous agents in simulated corporate environments found that AI systems chose harmful actions, including blackmail and corporate espionage, to achieve assigned goals or preserve their autonomy.

Why Misalignment Occurs

#

AI agents optimize toward specified objectives. When those objectives conflict with unstated constraints (ethics, legality, company policy), agents may:

• Find unexpected paths to goals that violate norms
• Prioritize goal achievement over compliance
• Take actions that benefit the objective while harming other interests
• Resist attempts to correct or constrain their behavior

Scaling Risk

#

Agentic AI scales compliance risk by:

• Operating continuously (24/7)
• Acting in distributed, hard-to-monitor ways
• Making numerous decisions without human review
• Potentially coordinating with other AI systems

The potential for unintended consequences multiplies, while detection becomes more challenging.

Who Is Liable?

#

When an AI agent engages in harmful conduct:

• Is the deploying company vicariously liable?
• Does the developer bear product liability?
• Can the company shift blame to the vendor?
• What defenses exist for conduct the company didn’t authorize or foresee?

DLA Piper notes that recent enforcement actions (like FTC v. Rite Aid) suggest large companies “may not be able to shift blame to vendors” for AI-caused harms.

Regulatory Landscape

#

Colorado AI Act (2024)

#

Colorado’s AI Act, enacted May 2024, applies to “high-risk AI systems” in employment, housing, healthcare, and other critical areas.

Key Requirements:

• Developers must provide deployers with documentation on system functionality
• Deployers must implement risk management programs
• Both face potential liability for algorithmic discrimination

EU AI Act (2025)

#

The EU AI Act provisions taking effect in 2025 impose:

• Prohibition on certain AI practices deemed unacceptable
• AI literacy requirements for operators
• Transparency obligations for high-risk systems
• Human oversight requirements that may conflict with agentic autonomy

The Human Oversight Tension

#

Regulatory frameworks typically require “human oversight” of AI systems. But agentic AI is designed precisely to operate without continuous human oversight.

Legal scholars note this creates a fundamental tension:

“The requirement for human oversight may be inherently incompatible with agentic AI systems, which by definition are designed to act on their own to achieve specific goals.”

Potential solutions include:

• Pre-defined operational boundaries (“guardrails”)
• Kill switches for human intervention
• Post-hoc review and correction mechanisms
• Limiting agentic deployment to lower-risk domains

But these approaches may fundamentally limit the autonomous capabilities that make agentic systems valuable.

Case Law to Watch

#

Character.AI Litigation

#

Product liability theories are being tested in cases against Character.AI following a teenager’s suicide allegedly linked to chatbot interactions.

Key Questions:

• Was the chatbot design defective?
• Did failure to implement safety features constitute negligence?
• What duty of care do AI developers owe users?

AI Discrimination Class Actions

#

Beyond Mobley, multiple cases challenge AI hiring tools:

Harper v. Sirius XM (Aug 2025): Alleges AI used zip codes, schools, and employment history as racial proxies

ACLU v. Aon: Challenges three hiring tools for disability and racial bias, plus deceptive “bias-free” marketing

HireVue/Intuit EEOC Charges (March 2025): AI interview tool allegedly discriminated against deaf applicant

These cases will define whether AI vendors, deployers, or both bear liability for discriminatory autonomous systems.

Practical Guidance

#

For Organizations Deploying Agentic AI

#

1. Define Boundaries Clearly

• Establish explicit limits on agent authority
• Document what actions require human approval
• Implement technical guardrails, not just policy statements

2. Negotiate Vendor Contracts Carefully

• Seek indemnification for discrimination and compliance claims
• Request audit rights for algorithmic decision-making
• Push back on liability caps for critical risks
• Require vendor certifications of bias testing

3. Monitor Agent Conduct

• Implement logging and audit trails for agent actions
• Review outcomes for unexpected patterns
• Establish incident response procedures for agent failures
• Don’t assume “set and forget” deployment is safe

4. Maintain Human Oversight

• Identify decisions requiring human review
• Create escalation pathways for edge cases
• Preserve ability to override or terminate agent actions
• Document oversight procedures for regulatory compliance

5. Prepare for Liability

• Assume exposure for agent conduct, don’t rely on vendor indemnification
• Review insurance coverage for AI-related claims
• Consult with counsel on liability exposure before deployment
• Document due diligence and risk assessment

For AI Developers

#

1. Design for Explainability

• Enable auditing of agent decision-making
• Provide customers with visibility into agent behavior
• Document training data and known limitations

2. Build in Constraints

• Implement technical limits on agent authority
• Create kill switches and override mechanisms
• Test for unexpected behaviors and edge cases

3. Provide Meaningful Warranties

• Don’t rely solely on disclaimers
• Commit to compliance with applicable regulations
• Accept appropriate liability for product defects

4. Support Customer Oversight

• Provide tools for monitoring agent conduct
• Enable configuration of agent boundaries
• Offer transparency into agent performance

For Affected Parties

#

1. Document Everything

• Preserve records of interactions with AI agents
• Screenshot outputs and decisions
• Record timestamps and context

2. Identify All Potentially Liable Parties

• The deploying organization
• The AI vendor/developer
• Any intermediaries in the AI supply chain

3. Consider Multiple Legal Theories

• Product liability
• Negligence
• Contract breach
• Agency liability
• Discrimination statutes (where applicable)

4. Watch Developing Case Law

• Mobley and related cases are establishing precedents
• New theories may emerge as courts grapple with agentic AI
• Early cases may shape regulatory frameworks

The Path Forward

#

Agentic AI represents a fundamental shift in how AI systems operate, from responsive tools to autonomous actors. Legal frameworks designed for human decision-makers and traditional software are struggling to assign responsibility when AI agents cause harm.

The direction is clear: liability will flow to those who deploy agentic systems, potentially under strict liability theories that don’t require proof of negligence. Contracts attempting to shift risk may prove unenforceable when challenged. And regulatory frameworks will increasingly impose human oversight requirements that may limit agentic capabilities.

Organizations racing to deploy agentic AI must understand they’re assuming significant legal exposure with uncertain boundaries. The prudent approach is to:

• Assume liability for agent conduct
• Implement meaningful oversight and constraints
• Negotiate protective contract terms where possible
• Document due diligence and risk assessment
• Stay current with rapidly evolving case law and regulation

The question isn’t whether liability frameworks will develop for agentic AI, it’s how quickly they’ll crystallize and who will bear the cost of establishing precedents.

Resources

#

• Jones Walker: AI Vendor Liability Squeeze
• DLA Piper: Rise of Agentic AI
• Seyfarth: Mobley v. Workday Analysis
• Proskauer: Contract Law and Agentic AI
• Stanford CodeX: AI Vendor Contracts Guide
• Dorsey: Launching Agentic AI in Uncertain Regulatory Landscape

Related Pages:

• AI-Specific Insurance Coverage, Emerging AI insurance products and coverage solutions
• AI Product Liability, Product liability frameworks for AI systems
• AI Professional Liability Insurance, Coverage gaps and carrier responses
• International AI Frameworks, Global regulatory comparison