Agentic AI Creates Legal No-Man's Land: Courts Haven't Decided Who's Liable When Agents Go Wrong

AI agents are signing contracts, executing financial transactions, writing and deploying code, and browsing the web on behalf of users — and when they cause harm, no court has yet produced consistent answers about who's responsible. That's the legal reality in October 2026, as Lewis Silkin's analysis of AI litigation trends documents an emerging body of agentic AI cases without settled doctrine to resolve them.

The stakes are rising fast. An AI agent that purchases the wrong product, transfers funds based on a misread instruction, triggers an automated business process that causes downstream damage, or deploys code with a security vulnerability isn't a thought experiment anymore. These are live scenarios reaching courts and regulators in multiple jurisdictions simultaneously — without any of those jurisdictions having established the case law needed to guide decisions.

What "agentic AI" means legally

An AI agent, in the legal context Lewis Silkin is examining, is software that takes autonomous actions — without requiring human approval at each step — to achieve a goal. The defining feature isn't intelligence or sophistication; it's autonomy. A chatbot that answers questions isn't agentic. An AI system that reads your email, identifies a bill you need to pay, logs into your bank account, and pays it without asking you first — that's agentic.

The practical spectrum in 2026 runs from research agents that browse the web and compile summaries, to code-writing agents that commit to production repositories, to financial agents that execute trades or payments, to contract agents that draft, negotiate, and in some implementations countersign agreements.

Each of these creates liability exposure that existing legal frameworks weren't designed to handle.

The three-way liability problem

When an AI agent causes harm, courts are grappling with three potential liability assignments — and no jurisdiction has yet established a clear hierarchy:

Developer liability. Did the AI developer create a product that was defective, inadequately tested, or deployed without sufficient safeguards? Product liability frameworks exist for this analysis, but they assume the defective product behaved in a consistent, testable way. AI systems behave probabilistically — the same input can produce different outputs — making "defect" harder to define and prove in court.

Deployer liability. Did the company that deployed the AI agent for a business use case configure it negligently, fail to implement adequate guardrails, or deploy it in a context the developer's terms of service didn't authorize? This is the most active liability area in 2026, particularly in enterprise settings where companies deployed off-the-shelf AI agents and configured their permissions broadly.

User liability. Did the end user grant the AI agent permissions, resources, or access that enabled the harm — and can that authorization transfer responsibility to the user even if they didn't anticipate the specific harmful action?

The EU AI Act's high-risk and general-purpose AI provisions, which entered enforcement in August 2026, added a compliance layer on top of this three-way uncertainty. The Act requires risk assessments and oversight measures for certain AI deployments, but it doesn't resolve the underlying liability question — it adds regulatory exposure on top of unresolved civil liability.

Where courts are landing

Lewis Silkin's analysis identifies several early patterns in agentic AI litigation, though none constitute settled doctrine:

• Financial transaction errors are being treated most like existing product liability cases, with deployers bearing significant exposure for inadequate guardrails — particularly where agents had access to payment systems with insufficient human-in-the-loop checkpoints before execution
• Contract execution cases are in early stages, with courts examining whether AI-executed agreements bind the authorizing party under standard agency law, or whether the agent's deviation from intended instructions voids the contract
• Code deployment incidents are being analyzed under professional negligence frameworks in some jurisdictions and product liability in others, with no consensus on which standard applies when AI-generated code causes production damage

The common thread in early adverse rulings: deployers who gave agents broad permissions without proportionate oversight are bearing the heaviest liability. Courts appear to be applying a "you authorized it, you own the outcome" logic that tracks more closely with employer liability doctrine — where employers are responsible for actions taken by employees within the scope of their authority — than with product liability.

What the EU AI Act enforcement adds

August 2026 marked the enforcement of the EU AI Act's provisions covering general-purpose AI models — the category that covers the foundation models underlying most agentic systems. Deployers using these models for high-risk applications are now required to maintain:

• Risk assessments documenting potential harms and mitigations
• Human oversight mechanisms proportionate to the agent's autonomous action scope
• Incident logging and reporting for harmful outcomes

Meeting these requirements doesn't shield a deployer from civil liability, but failing to meet them creates a second exposure — regulatory enforcement by EU member state AI authorities — that compounds the civil risk. Expect the first significant enforcement actions against agentic AI deployments in Q1 2027.

What companies need to do now

Businesses that have deployed AI agents — for customer service, procurement, financial operations, IT management, or any other function where the agent takes autonomous actions — face a concrete risk management problem. The absence of settled law doesn't mean absent liability; it means unpredictable liability, which is often worse because it makes insurance, legal reserves, and contractual risk allocation harder to calculate.

Practical steps that matter in the current environment:

• Audit agent permissions. What can your deployed agents actually do without human approval? Every autonomous capability — send email, make purchases, execute payments, deploy code — is a potential liability vector that needs proportionate oversight.
• Implement checkpoints for high-stakes actions. For financial transactions, external communications sent on your behalf, and code pushed to production, require human confirmation before execution. The litigation record shows courts hold deployers responsible for outcomes when they gave agents unrestricted access.
• Document your oversight. When incidents occur, courts and regulators will ask what oversight measures were in place. Documented risk assessments and guardrail implementation matter — both for defensibility and for EU AI Act compliance.
• Review vendor agreements carefully. Most enterprise AI agreements shift significant liability to the deployer through broad indemnification clauses. Understand precisely where the developer's liability ends and yours begins before expanding agent permissions.

What to watch

The next six to twelve months will produce more precedent. Watch for the first significant EU AI Act enforcement action against an agentic deployment, and for U.S. appeals court decisions in financial transaction cases to establish clearer doctrine on deployer versus developer liability. No jurisdiction is close to a comprehensive agentic AI liability framework — but the legal no-man's land is actively shrinking as cases accumulate and early patterns harden into doctrine.

By Hector Herrera