AI agents are no longer a futuristic concept confined to research papers. They're booking meetings, executing code, querying databases, triggering financial transactions, and making decisions — often without a human in the loop. And that autonomy is creating a legal and operational minefield that most organisations are woefully unprepared for.
The urgent question is no longer whether an AI agent can go wrong. It's who gets the bill when it does.
If your IT and legal teams haven't sat down together to answer that question yet, this post is your wake-up call.
The Scale of the Problem Is Already Alarming
Before diving into liability, let's be clear about the scope of the risk.
Upwards of half of the three million AI agents currently in use by organisations in the US and UK are "ungoverned and at risk of going rogue,"
according to a study by API management platform Gravitee.
The results, drawn from a December 2025 survey of 750 IT executives and practitioners, revealed that AI agents are being deployed faster than security teams can keep up.
Even more alarming,
only 14% of teams secure full security sign-off before deployment.
AI agents are no longer experimental side projects — they're scheduling meetings, querying internal databases, pulling data from SaaS platforms, triggering workflows, and executing tasks on behalf of users without human involvement at every step.
Meanwhile,
industry leaders fear agents could delete databases, authorise payments, or leak secrets before anyone notices.
This isn't a hypothetical threat landscape. It's today's enterprise reality.
Who Actually Owns the Liability?
This is where things get legally complicated — and fast.
AI has evolved from chatbots to autonomous agents capable of executing code, signing contracts, and booking transactions, and traditional agency law is being tested.
If an AI agent executes a disadvantageous contract, the question of whether the user is bound by it is one courts are now scrutinising as they assess whether users or developers bear liability for autonomous errors.
The honest answer? Nobody fully knows yet.
Courts are scrutinising whether users or developers bear liability for autonomous errors, but to date, they have not issued definitive rulings allocating liability for fully autonomous agent behaviour.
What legal scholars and frameworks have established is that liability is likely to be shared — and messy.
Liability has to be allocated between the increased number of actors involved in the development and deployment of an agentic system.
Traditional agency law, product liability law, and computer fraud law all struggle to effectively address liability allocation for AI agents.
The Developer vs. Deployer vs. User Triangle
Researchers are advocating for dynamic liability allocation among developers, users, and platforms based on the AI agent's degree of autonomy, the foreseeability of its conduct, and each party's capacity for control.
In practice, this means:
- Developers may face product liability if a defect in the agent's design caused harm.
Companies developing AI-driven applications may be held liable for harms caused by their agents through product liability law, where developers may be found liable for harms stemming from defective design of their product.
- Deployers — the organisations integrating and releasing agents into the wild — carry significant accountability because they control the operational environment.
- End-users may bear responsibility for tasks they explicitly authorise an agent to carry out.
It remains an open question as to whether the accountability landscape changes as agents get more autonomous and unpredictable — and critically, AI agents are not human or legal persons and cannot be meaningfully held accountable for harm themselves.
The Regulatory Landscape Is Hardening — Fast
If you've been waiting for governments to "figure this out" before acting, that window has closed.
The regulatory landscape around autonomous agents has hardened materially in 2025–2026, moving from aspirational principles to enforceable obligations on concrete deadlines.
Here's what IT and legal teams need to track:
EU AI Act
The EU AI Act entered phased enforcement — prohibitions on unacceptable-risk systems took effect in February 2025, transparency obligations for general-purpose AI models activated in August 2025, and the full high-risk system requirements including human oversight and conformity assessments are scheduled to apply from August 2026.
The EU AI Act's high-risk system obligations and transparency rules entering into force between 2025 and 2026 cement explainability, traceability, and human oversight as legal requirements.
US Federal and State Law
In the absence of a comprehensive federal AI bill, states such as California, Utah, Texas, and Colorado have filled the void.
States are proposing and passing AI legislation at record speed in the absence of any overriding federal AI law.
Federal and state agencies, including the EEOC and FTC, have emphasised that existing employment, credit, housing, and consumer protection laws apply equally to AI-mediated decisions.
NIST's AI Agent Standards Initiative
NIST's AI Agent Standards Initiative, launched in February 2026, targets three pillars — interoperability, security, and testing and evaluation.
Singapore's First-of-Its-Kind Framework
Singapore's Model AI Governance Framework for Agentic AI, launched in January 2026 as the world's first national governance framework specifically designed for agentic systems, provides guidance across four dimensions — risk bounding, human accountability, technical controls, and end-user responsibility — and establishes that organisations remain legally accountable for their agents' behaviours regardless of voluntary compliance.
The Multi-Agent Problem Makes This Even Harder
If you think governing a single AI agent is difficult, consider what happens when agents talk to each other.
Legal frameworks currently being built for AI agent liability assume a single-agent model that multi-agent deployment has already outgrown, and policymakers should mandate traceability infrastructure before the inevitable harms force courts to improvise.
Agent collaboration can unintentionally break segregation of duties, escalate privileges indirectly, or amplify minor errors into enterprise-wide incidents.
When harm occurs across a chain of agents from multiple providers, tracing the origin of a failure — let alone assigning blame — becomes extraordinarily complex.
As AI governance matures, organisations will be expected to explain why an autonomous system acted the way it did. "The model decided" will not be an acceptable answer.
The Security Dimension: Why IT Teams Can't Ignore This
Legal liability doesn't exist in a vacuum — it's directly tied to how well (or poorly) your IT team governs these systems.
High-autonomy agents that have broad access to tools and data, and the freedom to apply reasoning at runtime, can perform complex tasks but securing them is "an open challenge" today.
The large language model as it is today will always be susceptible to jailbreak and prompt injection attacks, and organisations cannot prevent such attacks 100% of the time regardless of how much money is spent on guardrails,
according to Gartner's research VP Dennis Xu.
Agents operating with persistent memory and high privileges increase the risk of delayed attacks, data exfiltration, and large-scale security failures, especially as agents interact with other agents and external systems beyond traditional security visibility.
Insurers have begun introducing "AI Security Riders" that require documented evidence of adversarial red-teaming, model-level risk assessments, and specialised safeguards as prerequisites for underwriting, and this trend is expected to continue in 2026 as it becomes increasingly common for insurance carriers to require alignment with recognised AI risk management frameworks.
Practical Tips for IT and Legal Teams — Act Now
The regulatory and legal picture is evolving rapidly, but there are clear, actionable steps you can take today to reduce exposure:
For IT Teams:
- Build an agent inventory.
Many organisations don't even know how many agents they have, where they're running, or what they can touch.
You cannot govern what you cannot see.
- Apply least-privilege principles. Limit every agent's permissions to the minimum required for its specific task. Reduce blast radius before an incident occurs.
- Implement continuous monitoring.
Governance and architectural controls — including tighter permission boundaries, stronger monitoring, and security models designed specifically for autonomous agents — are recommended over traditional application defences.
- Red-team your agents. Adversarial testing for prompt injection and jailbreaks should be standard practice before any agent goes to production.
For Legal Teams:
- Audit vendor contracts immediately.
Organisations should review vendor contracts for AI agents to ensure indemnification clauses specifically address autonomous actions and hallucinations resulting in financial loss.
- Update vendor agreements.
Update vendor agreements to shift liability for IP infringement and autonomous errors back to AI providers where possible.
- Map your regulatory obligations by jurisdiction. Understand where the EU AI Act, state-level US laws, or frameworks like NIST apply to your specific deployments.
- Document everything. Maintain logs of agent actions, human oversight decisions, and governance policies. In any future litigation, this paper trail will be essential.
For Both Teams Together:
- Establish a joint AI Governance Committee with clear ownership over agent deployment approvals, monitoring, and incident response.
- Define your organisation's public-facing policy on AI agent use, including disclosure obligations to customers and third parties.
Conclusion: Governance Is No Longer Optional
The rise of agentic AI represents one of the most significant operational and legal shifts enterprises have faced in a generation.
If 2024 was the year of AI hype and 2025 was the year of AI accountability, the legal landscape has now shifted from theoretical debates to concrete enforcement actions and compliance deadlines.
Waiting for the dust to settle is no longer a viable strategy — the dust is settling, and it's settling on the organisations that failed to act.
The good news? The frameworks to protect yourself exist today. What's missing in most organisations is urgency.
If you're ready to get serious about AI agent governance — before a rogue agent forces your hand — start by conducting a full AI agent audit this quarter. Bring your IT and legal leads into the same room, review your vendor contracts, and build your monitoring infrastructure now. The organisations that treat AI governance as a competitive advantage, not a compliance burden, will be the ones best positioned to scale safely and confidently in the agentic era.
