Artificial Intelligence is becoming more sophisticated by the day. One of the most transformative developments is the emergence of AI agents.
Unlike generative AI, which primarily responds to prompts, AI agents can reason, plan and take action across multiple steps to complete tasks with minimal human input.
For businesses, this creates significant opportunities to improve efficiency, productivity and decision-making – but in parallel it raises new questions around governance, accountability and risk, which need to be navigated.
Here our dedicated AI lawyer Toby Irenshtain explains everything you need to know.
Generative AI vs agentic AI
GenAI holds tremendous power to create content, such as text, videos, and images based on human prompting. But it cannot act autonomously on behalf of humans. Rather than simply analysing prompts and generating responses, AI agents perform tasks on behalf of humans.
While humans provide AI agents with a goal and operating space, they do not necessarily provide AI agents with a roadmap for how to achieve the goal. Instead, AI agents autonomously develop a strategy to execute the task with the resources at their disposal and continuously refine the approach based on feedback of the effectiveness of their strategies.
Agentic AI works by connecting with GenAI: at a technical level, an AI agent is an autonomous software or bot connected to a large language model (LLM). This connection enables the agent to create various outputs. GenAI LLMs are the “brains” behind AI agents, whose operating space can be defined across multiple domains, including Web2 and Web3. Companies like Anthropic, Google, and OpenAI have developed their own AI agents, dubbed "computer use”, "Project Mariner", and "Operator". These agents interact with their respective LLMs to plan and execute tasks efficiently.
AI agents as impact multipliers
It's important to think about AI agents as impact multipliers, due to their ability to execute tasks at speeds unmatched by humans. When it comes to efficiency, AI agents significantly build upon advances offered by GenAI to streamline processes. While human actors may pose less risk relative to AI, they also are likely to be less efficient. LLMs build upon this efficiency, but only AI agents, and not LLMs, can act autonomously to complete assigned activities on behalf of a person, company or enterprise, significantly increasing efficiency by reducing human input.
Risks brought by AI agents
While AI agents offer the promise of numerous benefits, they also pose significant risks that need to be acknowledged and mitigated. These risks have been grouped by the International Scientific Report on the Safety of Advanced AI into three main categories:
- Malicious use risks: AI agents can be used for cyber offenses, disinformation, and manipulation of public opinion. Mitigating these risks can be challenging, especially for open-source products.
- Malfunction risks: AI agents may not always function as intended. As one example, they could face issues with control over simultaneous failures across different servers. Further, LLMs often have inherent biases that may impact agents’ function and result in disparate performance.
- Systemic risks: Systemic risks associated with the introduction of AI agents, particularly at the enterprise level, range from environmental and intellectual property risks to considerations for privacy and labour market impacts.
Governance approaches for AI agents
As with all types of AI, effective governance is essential for managing AI agents’ associated risks. Across the AI lifecycle, various levers can be manoeuvred with a goal to build a bespoke and effective governance strategy:
- Planning and design: Early-stage decisions about data sources and usage, instilling privacy and security by design, and considering de-biasing occur at this stage.
- Model development: At the model development stage, a variety of steps can be taken to protect against harmful risks. Illustrative examples include embedding awareness of uncertainty into the agent to guard against malfunctions based on misinterpretations or uncertainty or developing model cards to provide transparency about the system and its developers.
- Testing: Red teaming to assess potential negative outcomes, and harms and quality assurance to test the system's resilience, can secure the agent against future cybersecurity risks.
- Auditing: Both internal and external audits and activity logs can help with ongoing monitoring by keeping a continuous log of Agentic activity.
AI law and liability
Governments and regulatory bodies are recognising the importance of AI regulation.
As under usual agency law principles, enterprises remain responsible for what their AI agent does. Monitoring agentic tools’ behaviour is therefore a critical guardrail to effective implementation.
Some of the most important global developments in agentic AI law include the following:
- The EU AI Act was the world's first comprehensive AI law and remains the most significant regulatory framework. It does not specifically regulate “agentic AI” as a separate category but, instead, AI agents will be assessed under the Act’s existing framework according to their functionality, underlying model, intended purpose and deployment context. Depending on the use case, an AI agent may trigger rules for general-purpose AI, transparency obligations, prohibited practices or high-risk AI systems, with corresponding governance, oversight, documentation and risk-management requirements.
- China’s approach is more fragmented and interventionist, with AI agents likely to be assessed by reference to their functionality, outputs, user base and deployment context rather than under a single AI Act-style framework. Relevant obligations may arise under rules on generative AI, algorithm recommendation, deep synthesis, cybersecurity, personal information protection, content governance and cross-border data transfers, with corresponding expectations around testing, logging, auditability, oversight and risk management.
- In the US, the position is shaped by a growing patchwork of state-level AI laws and existing sectoral, privacy, consumer protection, biometric and employment rules rather than a single federal AI regime. Colorado's original risk-based AI law has been narrowed into a disclosure-based framework taking effect in January 2027. The state is one of many attempting to regulate AI in the absence of comprehensive federal legislation, while facing legal and political challenges amid the federal administration's lighter-touch approach to AI regulation. Depending on the use case, agentic AI may still trigger obligations under automated decision-making and profiling rules, California privacy and AI transparency requirements, biometric laws and workplace AI rules, with the overall focus on transparency, bias, discrimination, opt-out rights, risk management and accountability.
- Given that the UK is taking a principles-based, sector-led approach to regulating AI, agentic AI will generally be considered through existing laws and regulators’ expectations. Relevant regimes may include data protection, consumer protection, financial services, competition, employment, cybersecurity and online safety, with regulators such as the ICO, FCA, CMA and Ofcom focused on safety, security and operational resilience, transparency and explainability, fairness and redress, accountability and governance, and data protection.
Looking ahead
AI agents continue to offer new possibilities and efficiencies. At the same time, they introduce novel risks. As always, the balancing of innovation, risk, and legal liability will rely on effective and iterative governance approaches, robust implementation metrics, and comprehensive oversight strategies.
For advice on how your company can reap the rewards of agentic AI while mitigating the risks, reach out to any of our AI experts listed on our Artificial Intelligence service page.
This article was originally published in March 2025.