Kamal Tserakhau

Kamal TserakhauSupplied to BYGRID

Expert commentaryLaw

The one-person unicorn can wait. Who’s liable when AI employees get it wrong (Part 1)

Startups, like many other businesses, are increasingly handing customer correspondence, advertising budgets and initial candidate screening to agents. We asked Kamal Tserakhau, a lawyer and mentor who advises on deals and startup launches, to explain what companies are responsible for when that kind of team gets it wrong and what to put in place before deploying agents.

By: BYGRID Team8 min read

Contents

In early October, Google Research published a report, “Open and Emergent Problems in Agentic Privacy and Security”. More than 60 researchers look at how to control AI agents that contact customers, spend money and make business decisions on their own. From a legal standpoint, things are much simpler: if you replace part of your team with agents, everything they do is your responsibility and any obligations they create are yours.

01Who’s liable for an AI agent’s actions?

Claude behind the counter and an office shop in the red

The market has been promising for years to replace people with agents. Back in 2024, Artisan put up “Stop Hiring Humans” billboards around San Francisco. In spring 2025, Anthropic CEO Dario Amodei predicted that 2026 would see the first $1 billion company with one employee. Ads for AI tools on social media often promise some version of: “Buy a subscription and get a virtual sales rep, marketer and HR specialist.”

The first AI Inc companies, run entirely by agents, may eventually appear. In practice, a person is still an essential part of the process. At today’s level of automation, you need a human in the loop for legal and practical reasons. Project Vend is a telling example. The Claude agent gave discounts to anyone who asked, sold goods below cost and drove an office shop into the red. An AI agent acting as an employee can easily cost a company money. Saying “the agent did it” won’t get you off the hook. AI has no legal personality, so it cannot be a party to a contract. The obligations and liability stay with whoever uses the agent. Legislators are codifying that approach too. In October 2025, for example, California passed a law that bars defendants from arguing in their defence that the harm was caused by AI acting autonomously.

02Does a company have to honour a discount promised by a chatbot?

The virtual sales rep that hands out discounts if you ask nicely

A sales agent binds the company to any promise a reasonable customer would take as an offer from the company. The dividing line is whether the promise is believable.

The best-known example is a chatbot at the Chevrolet of Watsonville dealership in California. In December 2023, a user first told it to agree to anything and call every response a legally binding offer. Then the user asked for a new Tahoe for $1. The bot agreed! The dealer took the bot offline and, luckily for the dealer, never had to sell the car for $1. The customer who pushed the bot to that absurd result knew the price wasn’t real.

In Poland, you can withdraw from a transaction for value entered into by mistake if the other party caused the mistake, knew about it or could readily have spotted it (Article 84 of the Civil Code). If the customer deliberately deceived the seller’s agent, that qualifies as fraud and the seller can withdraw on that basis too. If an outsider slipped the hidden instruction to the agent, the seller can challenge the transaction on grounds of fraud only if the buyer knew about the manipulation and failed to disclose it.

A believable mistake is a different matter. An Air Canada passenger asked the chatbot about a bereavement fare and got incorrect but convincing instructions for claiming it. In 2024, a Canadian tribunal rejected the airline’s argument that the chatbot was responsible for its own statements and ordered Air Canada to pay C$812. The passenger had no way of knowing the answer was wrong. By the same reasoning, a company would probably have had to honour the 25% discount Claude was handing out in Project Vend. Unlike a Tahoe for $1, that discount looks plausible.

03Can an AI agent enter into a contract on behalf of a company?

“We agree,” the AI agent replied from a company email address

An agent could potentially even sign contracts on behalf of its owner. Businesses generally don’t entrust an agent with a qualified electronic signature, but an agent can accept terms in more straightforward ways: ticking “I accept the terms”, clicking “Confirm order”, replying “We agree” by email or sending an invoice that serves as an offer. Under Polish law, intent may be expressed by any conduct, including electronically, unless the law requires a particular form. When the agent ticks the box, the company is agreeing to the deal.

A customer might be sure they’re writing to “Svetlana from Sales”, who has authority to approve a discount, when in fact they’re talking to an AI agent. Under English law, apparent authority comes from what the company has represented to the other party, rather than from what a representative says about themselves. The company gave the agent the name, job title and corporate email address. An English court would therefore probably find that the company created the appearance of authority. There have yet to be court rulings on contracts entered into by AI agents, but a court would probably find that a contract was formed. There’s also a regulatory question. If “Svetlana” doesn’t identify herself as AI, that is likely to breach Article 50(1) of the AI Act, which requires customers to be told they are interacting with AI. The obligation falls on the provider of the system. If you built the agent yourself on someone else’s model, it falls on you.

04Who’s liable for budget overruns and mistakes in AI advertising?

The virtual marketer and the mystery of the missing ad budget

An agent with access to Google or Meta ad accounts can launch campaigns, raise bids and change budgets on its own. Every mistake can mean another charge to your card. An agent can spend money even without being asked. When a Washington Post columnist asked OpenAI’s Operator to find cheap eggs, it went ahead and ordered them on Instacart for $31.43.

To the platform and the bank, it’s an ordinary order from the company’s account. The payment went through on your card and under your login. You may be able to get the money back if the platform itself made a mistake, as in 2024 when Meta’s Advantage+ tool overspent advertisers’ daily budgets. If your agent ran up the bill, the company bears the cost.

Ad accounts have long attracted scammers who take over accounts and advertise their own products on someone else’s budget. An AI agent can be hijacked without stealing a password. A hidden instruction in a report, an email from a supposed partner or a web page the agent is reviewing could lead it to change an ad’s link or raise the budget. For now, that is largely a theoretical scenario, but the Google report describes how attacks of this kind work. The platform would still see the spending as activity from your account.

The second risk is the ad itself. If the agent promises consumers a feature the product doesn’t have, or invents a figure such as “saves 40% of your time”, that is misleading advertising, which EU consumer protection rules prohibit. From August 2026, videos with an AI character resembling a real person must be labelled as AI-generated. A label won’t make it permissible for such a character to pass itself off as a real customer. Fake reviews are prohibited.

A competitor can also challenge an ad. Polish law, for example, treats advertising that misleads customers and may influence their purchasing decisions as unfair competition. One common scenario is impersonating another brand. An agent buys ads against a competitor’s name and writes them so that customers think they’ve reached the competitor itself. Affected companies can demand that the campaign stop, that a statement be published and that they be compensated for the damage. Agents can also invent claims of compliance with safety standards or put logos on a site for companies that have never been customers of the advertiser.

05What to set up before launching an AI agent

“Give the agent a discount limit and the human the ‘Launch’ button”

Prices and discounts. Keep prices, discount limits and other material terms in your pricing system. An agent can be talked into ignoring a rule in its prompt. A limit enforced by the system is much harder to bypass. Bar the agent from ticking boxes to accept new terms and identify it as an AI assistant in its display name and email signature. If the agent made a mistake and the customer knew about it or could easily have spotted it, send them written notice that you’re withdrawing from the deal. EU countries’ laws often allow a party to withdraw from a transaction entered into by mistake within a period that starts when the mistake is discovered.

Advertising claims. Leave the launch of new campaigns and changes to ad links to people. The same goes for ads containing prices, figures or claims about the product. Label AI-generated videos before publication.

If you found this useful, buy me a coffee. Your support helps me keep talking to founders and sharing what they’ve learned.

Buy me a coffee (opens in a new tab)
Share