Cerata / Blog

Who is liable when an AI agent breaks the law?

The short answer is you. Here is how AI agent liability works in plain language: the old rule behind it, the cases and regulators that have tested it, why scale changes the math, and a checklist for staying on the right side of the law.

When an AI agent breaks the law, the person or company it acts for is almost always the one who answers for it. That holds whether the agent is a sales bot texting leads, a voice agent making calls, a recruiting tool screening applicants, or a personal assistant posting a listing for you.

Where courts and regulators have faced "the AI did it" as a defense, they haven't accepted it. The law already has a framework for someone acting on your behalf. It's called agency, and it's centuries old. Software doesn't escape it. It fits inside it.

This guide walks through why, with the cases and rules behind each point. It ends with what you can do about it.

A well-behaved agent can still break the law

Most of the work on agent risk today is about safety: blocking prompt injection, keeping agents inside their permissions, catching an agent that goes rogue. That work matters. It isn't the same as legal compliance.

Picture an agent that does exactly what it was told. It follows up with every lead on the list, politely and on time. Nothing about it is rogue. But some of those leads never agreed to marketing texts. A few live in states with stricter calling hours. One follow-up goes out as a call in a synthetic voice to a number that never gave consent. Each of those can be a violation.

Guardrails stop an agent from doing what you didn't ask. They can't tell you that what you asked is against the law.

The law doesn't care whether the agent was obedient. It cares what happened: who was contacted, when, with what consent, saying what. Answering that means knowing which laws apply to a specific action, for a specific person, in a specific place. That's a legal question, not a security one.

The old rule: whoever the agent works for answers for it

Agency law covers what happens when one party acts on behalf of another. The core idea is simple. If you authorize someone to act for you and they act within that authority, their acts bind you, and you can be liable for the harm they cause. Employers answer for employees. Companies answer for the contractors they hire to do their marketing.

The Federal Communications Commission applied this to telemarketing in a 2013 ruling involving DISH Network. It said a seller "may be held vicariously liable under federal common law principles of agency" for violations of the Telephone Consumer Protection Act (TCPA) committed by the third-party telemarketers it uses.2 The Supreme Court acknowledged that ruling in Campbell-Ewald v. Gomez.3

An AI agent isn't a legal person, so it can't be an agent in the legal sense on its own. That doesn't help you. If anything, it makes things simpler: what your software does is treated as what you did. And if a vendor runs the agent for you, the same agency rules that reach companies through their outsourced telemarketers can reach you through your vendor.

The chatbot that invented a refund policy

The clearest test so far came from Canada. In November 2022, after his grandmother died, Jake Moffatt used the chatbot on Air Canada's website while booking a flight. The chatbot told him he could apply for a bereavement fare retroactively. Air Canada's policy didn't allow that, and the airline refused the refund.

Before British Columbia's Civil Resolution Tribunal, Air Canada argued it couldn't be held liable for information from its agents and representatives, including a chatbot. The tribunal's February 2024 decision was blunt:1

In effect, Air Canada suggests the chatbot is a separate legal entity that is responsible for its own actions. This is a remarkable submission. While a chatbot has an interactive component, it is still just a part of Air Canada's website. It should be obvious to Air Canada that it is responsible for all the information on its website.

Moffatt v. Air Canada, 2024 BCCRT 149, para. 27

The tribunal found negligent misrepresentation and ordered Air Canada to pay $812.02, including $650.88 in damages. It also rejected the idea that customers should have to double-check the chatbot against another page of the same website.

The amount is small, and the tribunal is a Canadian small-claims body, so the decision doesn't bind any US court. But the reasoning is the kind any court would reach for. Your agent's words are your words.

Calls and texts: AI voices count as "artificial"

The TCPA is one of the most litigated consumer laws in the US, and it's the one AI sales and voice agents run into first. It bars calls made with "an artificial or prerecorded voice" to cell phones and residential lines without the called party's prior express consent, with narrow exceptions.4

In February 2024 the FCC removed any doubt about AI. Its declaratory ruling confirmed that the TCPA's restrictions on "artificial or prerecorded voice" cover "current AI technologies that generate human voices," so those calls "require the prior express consent of the called party."5 Voice clones, synthetic voices and AI agents that sound human are all covered.

Texts can be covered too. The Supreme Court called it "undisputed" that a text message to a cell phone counts as a "call" under the TCPA's autodialer provision.3 But whether a particular text falls under the TCPA depends on how it was sent and which provision applies. The Court later read "autodialer" narrowly in Facebook v. Duguid,13 and some courts have questioned whether texts count under the Do Not Call provisions. This is heavily litigated, so check before you assume either way.

The penalties are what make this serious. Anyone who receives an unlawful call can sue for their actual loss or $500 per violation, whichever is greater. A court can triple that, to as much as $1,500, if the violation was willful or knowing.4 State attorneys general can bring their own cases as well.

If you run AI sales agents or AI voice agents, consent is the first thing to get right.

Regulators: "There is no AI exemption"

In September 2024 the Federal Trade Commission announced Operation AI Comply, a set of enforcement actions against companies using AI to deceive consumers. Then-Chair Lina Khan put the principle plainly:6

The FTC's enforcement actions make clear that there is no AI exemption from the laws on the books.

FTC press release, September 25, 2024

That principle outlasted the change in leadership. In August 2025 the FTC approved a final order against Workado, which had advertised its AI content detector as 98 percent accurate. According to the FTC, independent testing put its accuracy on general-purpose content at 53 percent.7

The FTC Act prohibits unfair and deceptive practices, and it applies the same way whether a person or a model wrote the claim. If your agent tells a customer something false about a price, a refund, a product or a deadline, that's your company making the claim.

Hiring: automation doesn't move the responsibility

Employment law has been dealing with automated screening for years. In 2023 the Equal Employment Opportunity Commission settled its suit against iTutorGroup for $365,000. The EEOC said the company's application software was set to automatically reject women aged 55 or older and men aged 60 or older. It said more than 200 qualified US applicants were turned away, in violation of the Age Discrimination in Employment Act. The company settled without admitting wrongdoing.8

That wasn't sophisticated AI. It was a rule in software, which is the point. Once a hiring decision is automated, the employer still owns it.

Cities are adding rules aimed at automated tools. New York City's Local Law 144 bars employers and employment agencies from using an automated employment decision tool unless it has had a bias audit within the past year, a summary of the results is public, and candidates have received the required notices.9

Federal guidance in this area changes with administrations. The statutes don't. Title VII, the ADEA and the ADA still apply to every hiring decision, whoever or whatever makes it. If you're building AI recruiting agents, plan for the federal laws and the local ones.

Housing: the words in the listing are what count

The Fair Housing Act makes it unlawful to make, print or publish any notice, statement or advertisement for the sale or rental of a dwelling that indicates a preference, limitation or discrimination based on race, color, religion, sex, disability, familial status or national origin.10

That rule turns on what the ad says. An agent that writes "adults only" into a rental listing has published a preference based on familial status, unless the property legally qualifies as housing for older persons. It doesn't matter how helpful the agent was trying to be.

Algorithms have been tested here too. In 2022 the Justice Department settled its case against Meta over housing ads. It had alleged that Meta's ad tools, including a "Special Ad Audience" feature, relied on characteristics protected by the Fair Housing Act to decide who saw ads. Under the settlement, Meta agreed to stop using that tool, change its ad delivery system under outside review, and pay a civil penalty of $115,054.11

Federal guidance has moved here as well. HUD withdrew its 2024 guidance on digital housing ads, effective September 2025.12 The statute's ban on discriminatory ad statements wasn't withdrawn. It applies when a personal AI agent posts a sublet for you just as much as when a property manager's agent posts a hundred listings.

Scale turns a small mistake into a large one

This is the part that's new. A sales rep who misreads the rules might send a few bad messages before someone notices. An agent works at machine speed, and it makes the same mistake every time until someone changes it.

Many consumer protection laws count each call, text, email or listing as a separate violation. So the math multiplies.

1 callin an AI voice, without consent $500–$1,500statutory damages
10,000 callsthe same mistake, repeated $5M–$15Mstatutory damages
Calls using an artificial voice without prior express consent, under 47 U.S.C. § 227(b)(1) and the FCC's 2024 ruling. Statutory damages under § 227(b)(3): $500 per violation, up to three times that if willful or knowing. Before defense costs, and before any class is certified.

A mistake a person would make once, an agent makes ten thousand times, in your name. That's why the timing of the check matters more than its sophistication.

Can you push the liability onto someone else?

Sometimes others share it. Vendors and platforms can be liable alongside you, as the Meta case shows, and a contract can require a vendor to pay you back. But sharing isn't escaping. Consumers and regulators tend to start with the business whose name is on the message, and an indemnity is only worth what the vendor can pay.

Plan as if the liability stays with you. In most cases, at least part of it will.

What to do about it: a checklist

None of this means you shouldn't deploy agents. It means you should treat each action an agent takes as something with legal weight, and build for that.

  1. Know which laws apply to each action

    "We follow the TCPA" isn't enough. The same text can fall under different rules depending on where the recipient lives and what time it is there. Map each kind of action your agent takes, such as calls, texts, emails, posts, listings and payments, to the federal, state and local laws that cover it.

  2. Check before acting, not after

    Logs and audits find problems once the damage is done. The check has to happen between the agent deciding to act and the action going out. Checks are cheap. Misses are counted one violation at a time.

  3. Keep a record of what was checked

    When a complaint or demand letter arrives, you'll need to show what your agent knew, which rules it checked and why it went ahead: consent records, timestamps, the exact message and the reason. Records made at the time are worth far more than ones rebuilt later.

  4. Put a human on the edge cases

    Not every action is clear-cut. When the law is unsettled or the stakes are high, the agent should stop and wait for a person to decide, not guess. Build that path in from the start, with a named owner and a short response time.

  5. Keep up with changes

    States pass new calling and consent laws, cities add hiring rules, and agencies withdraw and reissue guidance. Whatever does your checking has to change with them.

Where Cerata fits

We built Cerata because that checklist is hard to run by hand at agent speed. Cerata is legal compliance for AI agents. Before an agent calls, texts, emails or posts, it sends Cerata the action. Cerata works out which laws apply and answers with one of four decisions, a reason and a fix, and keeps a signed receipt of every check.

Checked today: US federal rules for sales calls and texts, plus Florida, marketing email and rental listings. New York City job posts are being verified. Other domains are coming soon.

Clear Flag Hold Block

To see it on your own agent, follow the five-minute quickstart or contact us.

General information, not legal advice. Laws change, and how they apply depends on your facts. Talk to a lawyer about your situation. Cerata is not a law firm.

Sources

  1. Moffatt v. Air Canada, 2024 BCCRT 149 (B.C. Civil Resolution Tribunal, Feb. 14, 2024). decisions.civilresolutionbc.ca↩
  2. FCC, Declaratory Ruling on the DISH Network petitions, FCC 13-54 (May 9, 2013), para. 1. docs.fcc.gov↩
  3. Campbell-Ewald Co. v. Gomez, 577 U.S. 153 (2016). law.cornell.edu↩
  4. Telephone Consumer Protection Act, 47 U.S.C. § 227(b)(1) and (b)(3). law.cornell.edu↩
  5. FCC, Declaratory Ruling, Implications of Artificial Intelligence Technologies on Protecting Consumers from Unwanted Robocalls and Robotexts, CG Docket No. 23-362, FCC 24-17 (adopted Feb. 2, released Feb. 8, 2024), para. 2. docs.fcc.gov↩
  6. FTC, "FTC Announces Crackdown on Deceptive AI Claims and Schemes" (Sept. 25, 2024). ftc.gov↩
  7. FTC, "FTC Order Requires Workado to Back Up Artificial Intelligence Detection Claims" (Apr. 28, 2025) and final order announcement (Aug. 28, 2025). ftc.gov (April), ftc.gov (August)↩
  8. EEOC, "iTutorGroup to Pay $365,000 to Settle EEOC Discriminatory Hiring Suit" (Sept. 11, 2023). eeoc.gov↩
  9. NYC Department of Consumer and Worker Protection, "Automated Employment Decision Tools (AEDT)," Local Law 144 of 2021. nyc.gov↩
  10. Fair Housing Act, 42 U.S.C. § 3604(c). law.cornell.edu↩
  11. Settlement Agreement, United States v. Meta Platforms, Inc., No. 22-cv-5187 (S.D.N.Y., filed June 21, 2022), para. 15. justice.gov↩
  12. HUD Office of Fair Housing and Equal Opportunity, "Notification of Withdrawal of Fair Housing and Equal Opportunity Guidance Documents," Docket No. FR-6571-N-01 (withdrawal effective Sept. 17, 2025). federalregister.gov↩
  13. Facebook, Inc. v. Duguid, 592 U.S. 395 (2021). law.cornell.edu↩

More from the Cerata blog Subscribe by RSS

Let your agents ask first.

Give every action a legal check before it goes out, with a signed receipt you can show anyone.