AI Sales Agents in the US: TCPA, CAN-SPAM and AI Rules 2026

Short answer: what the rules require in 2026

If you automate outbound sales in the United States with AI, three bodies of law do most of the work: the Telephone Consumer Protection Act (TCPA) governs calls and texts, the CAN-SPAM Act governs commercial email, and a growing set of state AI and privacy laws governs what you must disclose and how automated scoring is used. An AI voice agent that dials prospects is treated as an "artificial" voice under the TCPA, so telemarketing calls to mobile numbers need prior express written consent and revocations must be honored within 10 business days. Most AI-written email and most B2B chat sequences are lawful today, provided headers are honest, opt-outs work, and nothing about the system claims to be a human when a prospect clearly asks.

Stratgik is a technology company, not a law firm. Nothing here is legal advice. What follows is the engineering and operating view: which rules change your architecture, which change your copy, and which only change your record-keeping. Confirm your specific obligations with US counsel before you launch.

Channel by channel: what actually applies

Teams buy "AI SDR" as one product, then discover each channel carries a different legal regime. Email is the most permissive, outbound voice the least. Mapping this before you build is what keeps a pipeline experiment from becoming a liability.

AI voice agents: the TCPA is the binding constraint

In February 2024 the Federal Communications Commission adopted a Declaratory Ruling confirming that calls using AI-generated voices are "artificial" within the meaning of the TCPA (FCC, "FCC Makes AI-Generated Voices in Robocalls Illegal"). The practical consequence is simple and often missed: your AI caller does not get the latitude of a human SDR. It sits in the same bucket as a prerecorded robocall, which means marketing calls to wireless numbers require prior express written consent, and the TCPA's private right of action makes mistakes expensive at volume.

Two 2025 and 2026 developments matter to the design of your system.

First, consent is slightly broader than it briefly looked. The FCC's "one-to-one consent" rule, which would have limited a single consent to one identified seller and required calls to be topically related to the interaction that produced the consent, was vacated by the Eleventh Circuit in Insurance Marketing Coalition v. FCC days before it was due to take effect in January 2025, on the basis that the Commission had exceeded its statutory authority (Morrison Foerster analysis). The older definition of prior express written consent governs. That is not permission to dial a bought list; it means a properly drafted consent can cover more than one seller.

Second, revocation is now a real engineering requirement. Rules effective April 11, 2025 require you to accept revocation by any reasonable method, treat standard keywords such as stop, quit, revoke, opt out, cancel, unsubscribe and end as valid, and process the request within 10 business days rather than the previous 30. The caller bears the burden of showing a revocation attempt was unreasonable. The related "revoke-all" provision — under which opting out of one message type stops everything from that sender — has been delayed twice and now carries an effective date of January 31, 2027 while the FCC considers changes (Burr & Forman). Build for revoke-all now; retrofitting a suppression list across five tools later is the expensive path.

In our experience the consent record, not the model, is where these programs fail. An AI agent that produces a transcript but cannot link a call to a timestamped, signed consent artifact is unauditable — and unauditable is indefensible.

Email: AI changes the volume, not the rules

CAN-SPAM is older and considerably less demanding than most founders expect. The Federal Trade Commission's compliance guide sets out the core duties: accurate header information, a subject line that matches the message, clear identification that the message is an advertisement, a valid physical postal address, a working opt-out mechanism that stays live for at least 30 days, opt-outs honored within 10 business days, and no fee or extra personal information required to unsubscribe. Responsibility is not transferable to an agency or a sending vendor, and the law makes no exception for business-to-business email (FTC CAN-SPAM compliance guide). Penalties are assessed per email, which is what makes an automated sender dangerous rather than merely untidy.

AI makes two of these harder to hold. Generated subject lines drift from message content, which is a deceptive-subject-line risk rather than a copywriting quibble. And multi-inbox sending infrastructure, now standard in AI outbound tooling, fragments suppression: if an opt-out registered in one sending domain does not propagate to the other eleven, you are out of compliance in nine business days regardless of intent.

Disclosure: when your agent must admit it is not human

Utah's Artificial Intelligence Policy Act is the most directly relevant to commercial interactions. As amended by SB 226 in 2025, a supplier must disclose that a consumer is interacting with generative AI rather than a human if the consumer makes a clear and unambiguous request, and must disclose proactively in high-risk interactions — those involving sensitive personal data or personalized financial, legal, medical, or mental health advice. The amendments added a safe harbor: a provider is shielded from disclosure-violation enforcement if the AI clearly and conspicuously identifies itself as non-human at the start of, and throughout, the interaction (Perkins Coie).

There is no single federal rule requiring an AI sales agent to identify itself, but state law is thickening. A set of chatbot statutes — California's SB 243, New York's AI companion law, and 2027 laws in Oregon, Washington, Nebraska and Idaho — impose non-human disclosure duties. These target companion chatbots rather than B2B sales agents, so most sales deployments fall outside them, but they set the baseline regulators now expect. Oregon's carries statutory damages of $1,000 per violation.

The engineering conclusion is the same under all of these laws and costs nothing: identify the agent as AI at the opening of the interaction, make "are you a bot?" a hard-coded honest answer rather than a model output, and never let the agent give regulated advice. A persona that denies being AI is also straightforwardly deceptive under Section 5 of the FTC Act, and the FTC has brought AI deception cases steadily since its 2024 "Operation AI Comply" sweep.

Lead scoring, profiling and the ADMT rules

The piece most teams miss is the privacy layer underneath the outreach. California's Privacy Protection Agency finalized regulations covering automated decisionmaking technology, risk assessments and cybersecurity audits, approved in September 2025 and effective January 1, 2026, with ADMT obligations for significant decisions applying from January 1, 2027 (CPPA announcement).

Read the scope before you panic or relax. "Significant decision" is defined narrowly: the provision or denial of financial or lending services, housing, education, employment or independent contracting, or healthcare services. Ordinary B2B lead scoring and sequence routing generally are not significant decisions. But if your AI decides who gets offered credit, a mortgage product, a job or care, the ADMT rules bite — pre-use notice, an opt-out with at least two request methods, an access right, and a human appeal path.

Colorado is the cautionary tale about building to a moving target. The Colorado AI Act (SB 24-205) was repealed and replaced by SB 26-189, signed on May 14, 2026, which swapped the algorithmic-discrimination framework for narrower transparency duties: pre-use notice before deploying covered automated decision technology that materially influences a consequential decision, an explanation within 30 days after an adverse outcome, developer documentation to deployers, correction and human-review rights, and three years of compliance records, enforceable only by the Attorney General and effective January 1, 2027 (Davis Wright Tremaine). Teams that spent 2025 building to the repealed version wasted the effort. The durable investment is the capability — notice, logs, explanations, human override — not any one state's checklist.

What this costs to build properly

These are the cost drivers we see in US deployments, not a quote. Costs vary widely with volume, integration depth and existing data quality.

Two patterns are worth stating plainly. Inbound-only AI agents are far cheaper and lower-risk than outbound dialing, because the prospect initiated contact and the TCPA consent machinery largely falls away — if you are starting, start there. And the compliance layer is reusable while the model layer is not: models and vendors will turn over, but your consent ledger and suppression service will outlive all of them, which argues for owning that piece rather than renting it inside a point tool. Our AI automation opportunity scanner helps sort which functions to automate first, and build vs buy which layers to own.

Pre-launch checklist

• Classify every channel you intend to automate and record which law governs it. Do not treat "outreach" as one system.
• Prove consent per number. For SMS and outbound voice, store a timestamped, signed consent artifact linkable to each contact attempt. If you cannot produce it on demand, do not dial.
• Centralize suppression. One suppression service, consulted by every sender and dialer, processing revocation within 10 business days and honoring all standard keywords.
• Build for revoke-all now, ahead of the January 31, 2027 effective date, so an opt-out in any channel stops everything from your brand.
• Hard-code AI disclosure. Identify the agent as AI at the start of every interaction and answer "are you human?" honestly outside the model. Take Utah's safe harbor.
• Bound the agent's topics. No financial, legal, medical or mental-health advice. Escalate to a human instead.
• Check CAN-SPAM mechanics on every sending domain: honest headers, matching subject lines, ad identification, valid postal address, working unsubscribe.
• Check state call-recording law before you record or train on call audio.
• Log everything. Transcripts, model version, prompt version, decision inputs, human overrides. Keep at least three years for ADMT-adjacent uses.
• Assess ADMT exposure. If any automated output influences credit, housing, education, employment or healthcare outcomes, treat January 1, 2027 as a hard deadline.
• Get US counsel to sign off on consent language, disclosures and the vertical you are selling into.

Frequently asked questions

Can I use an AI voice agent for cold outbound calls in the US?

Not to mobile numbers for marketing purposes without prior express written consent. The FCC has confirmed AI-generated voices are "artificial" under the TCPA, so an AI caller is treated like a prerecorded robocall rather than like a human SDR. Cold-calling a purchased list with an AI voice agent is the single highest-risk configuration, because the TCPA carries a private right of action and liability accrues per call.

Does my AI chatbot have to tell prospects it is not a human?

Under Utah's AI Policy Act you must disclose if the consumer clearly and unambiguously asks, and proactively in high-risk interactions involving sensitive data or personalized financial, legal, medical or mental health advice. Other state chatbot laws target companion chatbots rather than B2B sales. Disclosing upfront in all cases is both cheap and the route to Utah's safe harbor, so there is no commercial reason not to.

Is AI-written cold email legal in the United States?

Generally yes. CAN-SPAM does not require prior consent and covers B2B email, so the obligations are about honesty and mechanics: accurate headers, a non-deceptive subject line, identification as an advertisement, a valid postal address, and an opt-out honored within 10 business days. The risk AI introduces is volume and fragmentation across many sending domains, not the drafting itself.

Do the California ADMT rules apply to AI lead scoring?

Usually not. The ADMT obligations attach to "significant decisions," defined as the provision or denial of financial or lending services, housing, education, employment or independent contracting, or healthcare services. Ordinary B2B lead scoring and sequence routing fall outside that. If your model influences credit, insurance, hiring or care decisions, the January 1, 2027 compliance date applies to you.

What changed with the Colorado AI Act?

It was repealed. SB 26-189, signed on May 14, 2026, replaced the original algorithmic-discrimination framework with narrower transparency duties — pre-use notice, adverse-outcome explanations within 30 days, developer documentation, correction and human-review rights, and three-year record retention — effective January 1, 2027 and enforceable only by the Attorney General. The lesson for build planning is to invest in the underlying capabilities rather than any single state's rulebook.

What is the lowest-risk way to start with AI in US sales?

Inbound. An AI agent that answers calls, chats and form submissions from prospects who contacted you sidesteps most TCPA consent machinery. It also tends to show results faster, since response speed on inbound leads is usually the clearest bottleneck in a funnel.

Automating US sales outreach is less an AI problem than a consent-and-records problem wearing an AI costume. The teams that scale this successfully build the suppression ledger and the audit trail first, then add agents on top. If you are planning a US rollout, our United States practice and US privacy compliance pages set out how we approach it, and a technology decision sprint is the usual way we pressure-test an outreach architecture before anyone writes code. Stratgik is a technology firm, not a law firm; pair any build with US counsel.