The AI agent compliance glossary

The laws an AI agent runs into when it calls, texts, emails, hires, rents, collects or handles health data, in plain language. Each entry says what the rule is, why it matters for agents and where to read the law itself.

Most tools for AI agents stop rogue behavior. This page is about the other risk: a well-behaved agent that breaks the law because of when it sent a message, who got it or what consent was on file.

45 terms

Terms starting with A

Adverse action notice

MoneyComing soon

A notice a creditor must give when it turns down a credit application or changes terms against the applicant. Under Regulation B it generally goes out within 30 days of a completed application and must give the specific principal reasons for the decision, or say how to ask for them. A vague reason like “you didn’t meet our internal standards” isn’t enough. A similar notice is required under the Fair Credit Reporting Act when a decision relies on a consumer report.

Why it matters for AI agentsIf an agent or a model makes the credit decision, it still has to produce specific, accurate reasons a person can understand.

Agency liability

Agents

The principle that a business is responsible for what its agents do on its behalf. In 2013 the FCC ruled that sellers can be held liable under federal agency law for TCPA violations by the telemarketers they hire. In Mobley v. Workday, a federal court in 2024 let job-discrimination claims go forward against an AI screening vendor on the theory that it acted as the employers’ agent. As of early 2026 the case was still pending.

Why it matters for AI agentsWhen an AI agent breaks the law, the business it acts for usually pays, whoever built the agent.

AI agent

Agents

Software that uses an AI model to work toward a goal by taking actions, such as sending messages, calling APIs, filling in forms or making payments, rather than only answering questions. An agent chooses its own next steps and uses tools to act. The law generally treats those actions as the actions of whoever put the agent to work.

Why it matters for AI agentsOnce software acts instead of advising, the laws that govern the action apply to it.

AI disclosure

Agents

Rules that require telling people when they’re dealing with AI. California’s bot law makes it unlawful to use a bot to knowingly mislead a person in California about its artificial identity to drive a sale or influence a vote, unless it discloses that it’s a bot. Utah requires businesses to disclose AI use when a consumer clearly asks, and licensed professionals to disclose it up front in high-risk interactions. On calls, the FCC treats AI-generated voices as artificial voices under the TCPA.

Why it matters for AI agentsAn agent that talks to people may need to say it’s an AI, and what triggers that duty differs by state and channel.

Artificial or prerecorded voice

Calls and texts

The TCPA restricts calls that use an “artificial or prerecorded voice”. In February 2024 the FCC ruled that AI-generated voices, including voice clones, are artificial voices under the TCPA. Calls that use them need the same consent as prerecorded calls, and the message has to identify who is calling.

Why it matters for AI agentsCalls placed by a voice agent need consent for an artificial voice, however human the voice sounds.

Autodialer

Automatic telephone dialing system, or ATDS

Calls and texts

Equipment covered by the TCPA’s limits on automated calls and texts. In Facebook v. Duguid (2021), the Supreme Court held that a device counts as an autodialer only if it can store or produce numbers using a random or sequential number generator. Some state laws, such as Florida’s, use their own definitions.

Why it matters for AI agentsThe federal autodialer rules may not reach an agent’s dialer, but artificial-voice and state rules can still apply to the same call.

Automated employment decision tool

AEDT, under New York City Local Law 144

HiringComing soon

New York City’s term for software that uses machine learning, statistics or AI to substantially help decide who gets hired or promoted. Employers and employment agencies in the city may use one only if it had an independent bias audit within the past year, a summary of the results is public, and candidates got notice at least 10 business days before it was used.

Why it matters for AI agentsAn agent that screens or ranks candidates may itself be an AEDT, with audit and notice duties before it runs.

Terms starting with B

Block

Cerata term, not law

A Cerata decision meaning the action breaks a rule and shouldn’t happen. Cerata gives the reason and the fix. For example, a sales text that would arrive after quiet hours where the recipient lives is blocked, and the fix is to send it in the morning.

Why it matters for AI agentsThe agent learns why it can’t act and what to do instead, so it can recover on its own.

Terms starting with C

CAN-SPAM Act

Controlling the Assault of Non-Solicited Pornography and Marketing Act of 2003

EmailChecked today

The federal law on commercial email, enforced by the FTC. It doesn’t require consent before sending, but every commercial email needs honest header information and subject lines, a valid physical postal address, a clear way to opt out and, unless the recipient agreed to receive it, a notice that it’s an ad. Opt-outs must be honored within 10 business days. Each email that breaks the law can bring a civil penalty of up to $53,088.

Why it matters for AI agentsEvery marketing email an agent sends needs the address and a working unsubscribe, and must skip everyone who already opted out.

Clear

Cerata term, not law

A Cerata decision meaning the action passed every rule Cerata checked for it, and the agent can go ahead. The receipt lists which rules were checked and as of which date. When Cerata is unsure, it doesn’t clear.

Why it matters for AI agentsThe agent gets a plain yes it can act on, with proof attached.

Coverage

Cerata term, not law

The actions, laws and places Cerata checks. Today that’s US federal rules for sales calls and texts (quiet hours, consent and Do Not Call) plus Florida’s, marketing email (CAN-SPAM) and rental listings (fair housing ad rules). New York City’s pay transparency rule for job posts is being verified. Other actions and domains are coming soon, and every receipt names the rules it checked.

Why it matters for AI agentsAn agent should know which of its actions are checked, so nothing it does falls into a gap unnoticed.

Terms starting with D

Disparate impact

Housing

When a policy that looks neutral falls harder on a protected group without enough justification. The Supreme Court recognized it for employment in Griggs v. Duke Power Co. (1971), Congress wrote it into Title VII in 1991, and the Court held that it applies under the Fair Housing Act in Texas Department of Housing v. Inclusive Communities Project (2015). Federal agencies have moved to pull back from it since 2025, but those court decisions and the statute still stand, and many state laws apply it too.

Why it matters for AI agentsAn agent that screens applicants or targets ads can cause disparate impact without any intent to discriminate.

Do Not Call Registry

National Do Not Call Registry

Calls and textsChecked today

The national list, run by the FTC, of phone numbers whose owners don’t want telemarketing calls. Sellers and telemarketers must check it and can’t make sales calls to listed numbers unless an exception applies, such as an established business relationship or the person’s written permission. It covers sales calls, not political, charity or purely informational calls. Companies must also keep their own internal do-not-call list of people who asked them to stop.

Why it matters for AI agentsBefore every sales call or text, an agent should check the national registry, any state list and the company’s own list.

Terms starting with E

ECOA

Equal Credit Opportunity Act

MoneyComing soon

The federal law that bans discrimination in any part of a credit transaction based on race, color, religion, national origin, sex, marital status or age, because income comes from public assistance, or because the applicant exercised consumer credit rights. Regulation B, at 12 C.F.R. part 1002, puts it into practice.

Why it matters for AI agentsAn agent that markets, prequalifies or decides credit must not use a protected trait, or a stand-in for one.

Terms starting with F

Fair Housing Act

HousingChecked today

The federal law, passed in 1968, that bans discrimination in housing based on race, color, religion, sex, national origin, familial status and disability. It also makes it unlawful to publish a housing ad that states a preference, limitation or discrimination based on those traits, so wording like “adults only” can break it. Many states and cities protect more groups.

Why it matters for AI agentsIn a rental listing the wording itself can be the violation, so every ad needs a check before it’s published.

FDCPA

Fair Debt Collection Practices Act

MoneyComing soon

The federal law on how third-party debt collectors deal with people. It bans harassment, false or misleading statements and unfair practices, and by default bars contact before 8 a.m. or after 9 p.m. local time. Collectors must stop most contact when a person asks in writing. People can sue for actual damages plus up to $1,000 in statutory damages.

Why it matters for AI agentsA collections agent’s timing, tone and every claim it makes about a debt are regulated.

Flag

Cerata term, not law

A Cerata decision meaning the action is risky as written. Cerata returns a safer version the agent can use instead, such as the same text with the opt-out language added. When Cerata is unsure, it flags rather than clears.

Why it matters for AI agentsThe agent gets a fix it can apply itself, not just a warning.

Terms starting with G

Guardrails

Agents

Controls that keep an AI system’s behavior within set limits. Safety guardrails stop harmful, off-topic or rogue behavior, such as leaked secrets, prompt injection or toxic text. They don’t tell an agent whether a well-formed action is legal: a polite, on-brand text can still break the TCPA because of when it’s sent, who gets it or what consent is on file. Legal compliance is a separate check, based on the law that applies to that specific action.

Why it matters for AI agentsGuardrails stop rogue agents. A legal check keeps well-behaved agents legal.

Terms starting with H

HIPAA

Health Insurance Portability and Accountability Act

Health and privacyComing soon

Federal law whose Privacy and Security Rules govern how health plans, health care clearinghouses and most health care providers, plus the business associates working for them, use and protect protected health information. It sets when health information can be used or shared without the patient’s permission and requires safeguards for electronic records.

Why it matters for AI agentsAn agent that books appointments or answers patients handles protected health information, and the company running it usually needs a business associate agreement.

Hold

Cerata term, not law

A Cerata decision meaning a person on your team should decide before the action goes ahead. Holds are for rare, high-stakes cases where the answer depends on facts Cerata can’t confirm, such as whether consent given two years ago still holds. The receipt records who decided.

Why it matters for AI agentsOnly the uncertain, high-stakes cases wait for a person, so review stays fast.

Human in the loop

Agents

A design where a person reviews or approves an AI system’s action before it happens, or can step in while it’s happening. It’s most useful where the right answer depends on facts or judgment the system can’t verify on its own.

Why it matters for AI agentsSending only the risky, uncertain cases to a person keeps review practical without letting those actions through unchecked.

Terms starting with M

MCP

Model Context Protocol

Agents

An open standard, introduced by Anthropic in 2024, for connecting AI models and agents to tools and data. An MCP server offers tools, and an MCP client inside the agent finds and calls them. One integration then works across many agents and apps.

Why it matters for AI agentsA legal check offered as an MCP tool can be added to an agent without rewriting it.

Mini-TCPA

Calls and texts

A state law that adds its own rules for telemarketing calls and texts on top of the federal TCPA. Florida’s Telephone Solicitation Act, Oklahoma’s Telephone Solicitation Act and Maryland’s Stop the Spam Calls Act are well-known examples. These laws can set tighter calling hours, define consent or autodialers differently, and let people sue.

Why it matters for AI agentsThe rules depend on where the recipient is, so an agent needs the state law as well as the federal one.

Minimum necessary standard

Health and privacyComing soon

A HIPAA rule that covered entities and business associates must make reasonable efforts to use, share or request only the minimum protected health information needed for the purpose. It doesn’t apply to sharing for treatment or to disclosures to the patient.

Why it matters for AI agentsAn agent should pass along only the health data a task needs, in prompts and tool calls as well as messages.

Terms starting with N

Negative option

MoneyComing soon

An offer where silence or failing to cancel counts as agreeing to pay, such as a free trial that turns into a paid subscription. For online offers, the federal Restore Online Shoppers’ Confidence Act requires clear terms, express informed consent before billing and a simple way to cancel, and many states have automatic renewal laws. The FTC’s 2024 click-to-cancel rule was struck down by a federal appeals court in July 2025, but those laws still apply.

Why it matters for AI agentsAn agent that signs someone up for a subscription has to disclose the terms and get clear consent before anyone is billed.

Terms starting with O

Opt-out

Calls and texts

A person’s request to stop getting messages. Under FCC rules, people can revoke consent to robocalls and robotexts in any reasonable way, such as replying STOP (from November 25, 2026, a caller can instead name one opt-out method it accepts, if every call or text states it), and callers must honor it within a reasonable time, no later than 10 business days. Under CAN-SPAM, commercial email must offer a working unsubscribe that’s honored within 10 business days.

Why it matters for AI agentsBefore it sends, an agent has to check for opt-outs on every channel, including a reply that arrived a minute ago.

Terms starting with P

Pay transparency

HiringBeing verified

Laws that require employers to include a pay range in job postings. There’s no federal rule, but states including Colorado, California, Washington and New York have one, as does New York City. The details differ, such as which employers are covered, whether benefits must be listed and how a good-faith range is defined, and remote roles can bring in several states’ rules.

Why it matters for AI agentsAn agent that drafts or posts a job ad has to include a range that meets the rules for every place the job could be filled.

Personal AI agent

Agents

An AI agent that acts for one person rather than a business, for example booking appointments, sending email, negotiating or buying things in that person’s name. Some laws reach only commercial activity, but others, such as call recording consent rules, apply to anyone.

Why it matters for AI agentsBefore a personal agent records a call or sends a pitch, it needs the same check a business would make.

Preflight check

Cerata term, not law

The call an agent makes to Cerata before it acts. The agent sends the planned action, such as a text with its recipient and context, and gets back clear, flag, hold or block, with a fix when one is needed and a signed receipt. In the HTTP API it’s a POST to /v1/preflight; in the TypeScript SDK it’s cerata.preflight().

Why it matters for AI agentsOne call before every action is the whole integration, whatever the agent is built on.

Terms starting with Q

Quiet hours

Calls and textsChecked today

The times of day when sales calls aren’t allowed. Under federal rules, telephone solicitations can’t be made before 8 a.m. or after 9 p.m. in the called person’s local time. Some states set tighter windows: Florida, for example, bars commercial solicitation calls before 8 a.m. or after 8 p.m. The clock that counts is the recipient’s, not the sender’s.

Why it matters for AI agentsAn agent working across time zones has to work out the recipient’s local time before it sends.

Terms starting with R

Reassigned Numbers Database

Calls and texts

A database, set up by the FCC, that shows whether a phone number has been permanently disconnected since a given date, which means it may now belong to someone else. Callers can check a number before calling. A caller that checks and reasonably relies on the answer gets a safe harbor from TCPA liability if the number turns out to have been reassigned.

Why it matters for AI agentsConsent belongs to a person, not a number, so an agent texting an old lead may be reaching a stranger who never agreed.

Receipt

Cerata term, not law

The signed record every Cerata check leaves: what the agent tried to do, which laws were checked as of which date, what was decided and who approved it. Anyone can verify a receipt without an account, and any change after signing shows. A receipt isn’t legal advice and doesn’t say an action is lawful everywhere, only what these rules, at these versions, decided.

Why it matters for AI agentsWhen a lawyer, partner or regulator asks what your agent checked, the receipt is the answer.

Regulation F

MoneyComing soon

The CFPB’s rule that puts the FDCPA into practice, at 12 C.F.R. part 1006. It covers calls, texts, email and social media. A collector is presumed to call too often if it calls about a debt more than seven times in seven days, or within seven days after a phone conversation about it. Emails and texts must include a clear way to opt out.

Why it matters for AI agentsAn agent that follows up on debts needs a running count of attempts for each debt and an opt-out in every text and email.

Primary source12 C.F.R. part 1006

Robocall

Calls and texts

The everyday name for a call made with an autodialer or with an artificial or prerecorded voice. The TCPA restricts these calls, and the FCC applies the same rules to texts sent with an autodialer. Robocalls to cell phones need the called person’s prior express consent, and marketing robocalls need prior express written consent.

Why it matters for AI agentsAn AI voice agent that places calls is making robocalls in the legal sense, with every consent rule that follows.

Terms starting with S

Safe harbor

Calls and texts

A provision that protects you from liability if you followed set steps, even when something goes wrong. Under the Do Not Call rules, a seller isn’t liable for a call made in error to a listed number if it has written procedures, trains its staff and uses a version of the registry no more than 31 days old. Checking the Reassigned Numbers Database correctly is another safe harbor.

Why it matters for AI agentsA safe harbor usually depends on proof that you checked, so an agent needs a record of every check it made.

Refusing a renter because of a lawful source of income, such as a housing voucher, Social Security or child support. Federal law doesn’t ban it, but many states and cities do, including California, New York and Washington. In those places a listing that says “no Section 8” can be illegal.

Why it matters for AI agentsA listing agent needs the local law, not just the federal one, to know which phrases it can’t use.

Statutory damages

Calls and texts

A fixed amount set by law that a person can recover without proving actual loss. The TCPA allows $500 per violation, up to $1,500 if the violation was willful or knowing. The FDCPA allows up to $1,000 per individual case plus actual damages. When the amount is per violation, it multiplies with volume.

Why it matters for AI agentsAgents act at scale, so one wrong rule repeated 10,000 times can be 10,000 violations.

Terms starting with T

TCPA

Telephone Consumer Protection Act

Calls and textsChecked today

The main federal law on telemarketing calls and texts, passed in 1991 and enforced by the FCC. It limits calls made with an autodialer or an artificial or prerecorded voice, sets the basis for calling hours and the Do Not Call Registry, and treats texts as calls. FCC rules under it are at 47 C.F.R. § 64.1200. People can sue directly, which is why TCPA class actions are common.

Why it matters for AI agentsAn agent that sends one message to thousands of people can repeat the same violation thousands of times, each with statutory damages.

Tool call

Agents

When an AI model asks to run a function, such as send_sms or create_listing, with specific inputs, and the software around it runs the function and returns the result. Tool calls are how an agent’s decisions become actions.

Why it matters for AI agentsThe tool call is the moment an agent acts, which makes it the natural place to check the law first.

Terms starting with U

UDAAP

Unfair, deceptive or abusive acts or practices

MoneyComing soon

The standard for consumer financial products and services under the Consumer Financial Protection Act of 2010, which adds “abusive” to the older UDAP standard. The CFPB and state attorneys general enforce it. Abusive conduct includes taking unreasonable advantage of a person’s lack of understanding of a product’s risks, costs or conditions.

Why it matters for AI agentsA finance agent that recommends a product or explains its terms is held to this standard.

Primary source12 U.S.C. § 5531

UDAP

Unfair or deceptive acts or practices

Money

Section 5 of the FTC Act bans unfair or deceptive acts or practices in commerce, and every state has its own UDAP law, enforced by the attorney general and sometimes through private lawsuits. A claim is deceptive if it’s likely to mislead a reasonable consumer about something that matters to their decision.

Why it matters for AI agentsAnything an agent tells a customer, including something it made up, can be a deceptive claim the business answers for.

General information, not legal advice. Cerata is not a law firm. Definitions summarize the law as of November 2026 and leave out exceptions and details; the linked primary source is the authority. Spotted something out of date? Write to hello@cerata.ai.

Check the law before your agent acts.

Your agent sends Cerata the action. Cerata answers clear, flag, hold or block, with the fix and a signed receipt.