What is an agent under English law?

Agency arises where one party (the "Agent") is appointed by and authorised to act on behalf of another (the "Principal"). This is underpinned by long-established and well-understood legal principles but this is not (yet) the case in the world of AI, not least because AIs are not humans or corporate entities in the traditional sense.

Granting the agent Authority

The Principal can grant the Agent authority to act on its behalf in a number of ways, including both express authority or implied authority as well as apparent authority.

Acting on behalf of the Principal

Under English law, an agent has the authority to create (or terminate, depending on the context) legal relations between the Principal and third parties.

Case study

Let's look at a specific example to demonstrate the above points: Company Y enters into a service agreement with the provider of an AI Agent. Within that agreement, it is stated that the AI Agent's instructions are to: operate on Company Y's website, handle queries from customers and encourage sales.

• Express authority: given the clear instructions in the service agreement, the AI Agent responds to website user queries on Company Y's website, advises on the different products available and how to purchase.

• Implied authority: the AI Agent, with the belief that it is acting within the scope of its instruction to encourage sales, ends up offering a website user a special discount that Company Y would not typically offer.

• Apparent authority: the website user interacting with the AI Agent on Company Y's website is likely to come to the conclusion that the AI Agent is representing Company Y and therefore the special offer being made is legitimate.

This example is a good demonstration of where unintended consequences of using AI Agents may arise in light of the different types of authority that may be granted to them.

Why does all of this matter?

An agency is a genuine relationship with real, legal consequences and describing some AI systems as "agents" may arguably lead to the unintentional granting of power for these AI tools to meddle with and create legal relationships.

The English courts are likely to consider the following when assessing whether an agency arises at law:

• Whether the Agent was acting under authority of a Principal
• Whether a third party reasonably believed the AI agent had authority to act, or
• Whether the actions of the Agent resulted in a legally binding contract or agreement

If an agency relationship is found based on any of the above, the organisation that has adopted the AI tool may have to respond to arguments that they are legally responsible for detrimental actions or agreements made by the agent, potentially with adverse or harmful consequences (whether those be financial or reputational).

In addition to this, and as demonstrated in the example above, the ability to grant implied authority based on surrounding circumstances may result in AI tools acting beyond their intended scope set by both developers and adopters. Even though the organisation didn't explicitly provide authority to make such an offer to a customer, the AI Agent may have "believed" that this was an essential part of acting out its instruction of encouraging sales.

As AI becomes even more sophisticated it will become more autonomous in various ways. Until "AI agency" is properly tested by the courts, the following should be considered:

Practical tips for adopters of AI agents

• Ensure contractual clarity by clearly defining and documenting what AI agents are authorised to do in supplier agreements, customer agreements and T&Cs
• Be careful about using the term "agent" or "agency" when adopting AI tools
• Regularly monitor and audit your AI agents
• Upskill the organisation on the legal concept of agency
• Ensure that the limits of AI agents are made clear

Practical tips for developers of AI agents

• Try to avoid using the word "agent" in a legal sense when marketing your tool
• Add disclaimers into marketing materials
• Give your AI agents a tight scope and tune them cautiously – mitigate the risk that their actions or words could be seen as an intention to create a legally binding relationship
• Consider using role-based permissions and structure dialogue to explicitly indicate boundaries in ability
• Ensure all conversations and actions by AI agents are fully logged