The Rise of AI Agents
One of the more exciting developments in the rapidly evolving AI space has been the emergence of the AI agent. Very briefly, this is not just an application that responds to specific questions or performs specific tasks on your instruction, but an application that is capable of making decisions and taking actions independently of the programmer, or user as the case may be. This is exciting for some people, but terrifying for others.
An example of the power of AI agents is OpenAI’s recent solution of a mathematical problem arising from the Navier–Stokes equations, which are used to describe the motion of fluids. This was a problem that had remained unsolved for decades but was solved by OpenAI using some 10 000 AI agents, each working independently on different aspects of the problem, checking each other’s outputs, managing teams of other agents and so forth.
The Legal Problem
Now if I were any good at maths, I would not have become a lawyer at all but would instead have sought a more meaningful vocation; be that as it may, AI agents have every day commercial uses. For example, AI agents can buy airline tickets for you, book your hotel accommodation at the same time, and make payment. You can set up an AI agent to do all this for you autonomously. What if the agent misunderstands you and books you first-class tickets and three nights at a five-star hotel when you wanted a budget option?
If an AI agent enters into a contract on your behalf, is that a binding contract? If your AI agent books your tickets, are you contractually bound, and if you interact with the airline’s AI agent, can you enforce the resulting agreement with the airline? What if the agent makes an error or misunderstands what you want it to do?
The Law as it Stands
Parliament anticipated this issue some time ago in the Electronic Communications and Transactions Act (ECTA), which was passed in 2002 (antediluvian legislation on the AI timescale).
The ECTA introduced the concept of an “electronic agent” into South African law. This is a computer program or other automated means that works independently to action or respond to messages in an “automated transaction”. An “automated transaction” is a transaction conducted electronically where one or more of the parties does not review the agreement – like the airline website when you book that ticket. The airline does not have a person reviewing the transaction.
Section 20 of ECTA then says that an agreement can be formed when any or all of the parties to the agreement use an electronic agent, and that an agent can perform any of the acts needed for agreement formation (for example, the offer or acceptance of the terms of the agreement).
If anyone uses an electronic agent to form an agreement, that person is presumed to be bound by the terms of the agreement whether or not that person checked whether the electronic agent acted correctly or reviewed the terms of the agreement itself.
However, a party who interacts with an electronic agent is only bound if the system used allows a natural person to review the terms of the agreement before it is finalised. The electronic agent must also give the other party the opportunity to notify it of any errors made in the contracting process and to correct those errors.
Section 20 was presumably intended to cover more traditional programmatic interfaces like online shopping portals, online banking sites, etc., but fortunately it is broad enough to apply to AI agents too.
To see how this all works, let us look at a few examples.
The AI agent is the Retailer
You operate an e-commerce website which you use to sell motor spares. The site uses an AI agent to trade on your behalf – it not only handles simple orders, but it can also negotiate bulk discounts depending on the purchaser’s purchase history. Moreover, it can decide whether to allow sales on credit, and it does credit checks on potential purchasers to decide if it will do so. It handles payments and also occasionally threatens to beat bad payers with a tyre-iron.
An agreement made by the AI agent on your behalf is presumed to be binding, but remember that other applicable law will always apply, for example the Consumer Protection Act and those aspects of the ECTA dealing with electronic transactions generally. For example, certain information must be made available to consumers, and a cooling-off period applies to electronic transactions.
You are the Consumer
Let’s flip it around: you are the consumer and, having forgotten to pay your AI subscription, purchase an oil filter on the site by interacting with the retailer’s AI agent. The agreement is binding, as long as the AI agent allows you to review the terms of the agreement before you accept.
Now imagine that you order the wrong part. If the retailer’s AI agent allowed you the opportunity to prevent or correct the error, and you did not do so, then you are bound. However, if no such opportunity is given, then if you notify the retailer (or indeed the retailer’s AI agent) of the error as soon as practicable after you discover it, and follow the other required steps, then you are NOT bound. Remember to keep a record of your communications in this case, because you just can’t trust AI.
Bot-on-bot
Finally, consider two AI agents interacting. You instruct your AI agent to find the best deal for oil filters, consider each brand’s reputation, select the best price for a reliable oil filter, and make payment using your credit card. The retailer also has an AI agent running its site.
The ECTA is clear that an agreement may be formed between two parties that use an electronic agent, and as an AI agent qualifies as an “electronic agent”, you would be presumed to be bound, provided that the terms were capable of being reviewed by you before the agreement was formed. Note that if your own AI agent does not give you the opportunity to review the terms, that is not the retailer’s concern, and you are still bound.
What if your AI agent makes a mistake in its order? The exception for mistakes that I described above applies only to natural persons, so if your AI agent makes a “material error”, then you are still bound.
Remember, though, that all of this is still subject to the general contract law relating to formation of contracts. So, if your AI agent made a material error because the retailer’s AI agent posted misleading information on the retailer’s portal, then the agreement would be voidable just like any other contract.
