The Telemarketing Sales Rule and AI-Generated Sales Calls

Two federal regimes govern a sales call and they belong to different agencies. Getting an AI calling programme wrong usually starts with reading the TCPA case law and assuming it answers the TSR question, or the reverse. It does not: the exemptions are different and so is the enforcer.

The TSR is not the TCPA

The Telemarketing Sales Rule is a Federal Trade Commission rule at 16 C.F.R. Part 310, made under the Telemarketing and Consumer Fraud and Abuse Prevention Act. It regulates sellers and telemarketers engaged in telemarketing, and it is enforced by the FTC and by state attorneys general, with civil penalties per violation. The Telephone Consumer Protection Act, 47 U.S.C. § 227, is a statute administered by the Federal Communications Commission with a private right of action and statutory damages, which is why most litigation is TCPA litigation.

The scopes diverge in ways that matter. The TSR does not reach entities outside the FTC Act’s jurisdiction — common carriers acting as such, banks, federal credit unions, airlines and insurance in some respects — and it treats calls to induce a charitable contribution, business-to-business calls and certain others under separate provisions in § 310.6. The TCPA has no such carve-outs but hinges on the technology used to place the call. A programme can be compliant with one and not the other. The rule text is on the eCFR.

The prerecorded message rule

Section 310.4(b)(1)(v) makes it an abusive telemarketing practice to initiate an outbound telephone call that delivers a prerecorded message to induce the purchase of goods or services, unless the seller has obtained the recipient’s express agreement, in writing, that evidences the consumer’s authorisation to receive such calls, that includes the number to be called, and that was obtained without requiring the agreement as a condition of purchase. The provision also requires the call to disclose within the first two seconds after the greeting that the recipient may assert the do-not-call right, and to provide an automated interactive opt-out mechanism throughout the call.

A recorded pitch that a text-to-speech system generated is a prerecorded message. The fact that the audio was synthesised rather than voiced by a person changes nothing about the analysis, and a personalised variant generated per recipient is still a message delivered without a live agent. If your system dials, plays generated audio and captures a response, you are inside this provision and you need the written agreement.

Where a conversational agent falls

The harder case is a real-time conversational agent that listens, reasons and responds — not a recording, and not a person. The TSR text was written before this existed and does not address it by name, which means the position is genuinely unresolved and should be treated that way.

The stronger reading is that it falls within the prerecorded-message prohibition’s purpose and probably its terms: the recipient is being spoken to by a machine without a live human on the line, the opt-out rationale applies identically, and the FTC has consistently read the abusive practices provisions purposively. On the TCPA side the question is closer to settled in the same direction: the FCC’s declaratory ruling of 8 February 2024 confirmed that AI-generated voices are “artificial” voices under section 227(b), so calls using them require prior express consent. That ruling is covered separately in the FCC AI voice robocall page, and it is the safer thing to plan around, because a programme that satisfies the TCPA consent standard will generally satisfy the TSR’s.

What nobody should do is treat the absence of an explicit TSR provision as permission. The FTC has not said conversational agents are outside the rule, and its Operation AI Comply enforcement sweep announced in September 2024 demonstrates a willingness to use existing authorities against AI-enabled conduct rather than waiting for new ones.

Do-not-call, abandonment and the safe harbour

Three provisions apply regardless of who or what is speaking.

• Registry scrubbing. Section 310.4(b)(1)(iii)(B) prohibits calling a number on the National Do Not Call Registry absent an established business relationship or express written agreement, and § 310.8 requires sellers and telemarketers to pay the access fee for the area codes they call. Automation does not change the scrub.
• Entity-specific do-not-call. Section 310.4(b)(1)(iii)(A) prohibits calling a person who has previously asked not to be called by or on behalf of the seller. A conversational agent must recognise and honour that request in the moment, which is a product requirement: the model has to route “take me off your list” to a suppression write, not to a rebuttal.
• Call abandonment. Section 310.4(b)(1)(iv) treats it as abusive to abandon a call, defined as failing to connect the call to a sales representative within two seconds of the person’s completed greeting. The safe harbour in § 310.4(b)(4) permits abandonment of no more than three per cent of answered calls, measured per day per campaign, with a prompt recorded identification message and record retention. Any dialler you point a model at still has to hold that ratio.
• Curfew and disclosures. Section 310.4(c) limits calls to between 8am and 9pm local time at the called party’s location, and § 310.4(d) requires prompt disclosure of the caller’s identity, that the purpose is to sell, and the nature of the goods or services.

Misrepresentation and impersonation

Section 310.3(a) requires material disclosures before the customer consents to pay — total cost, material restrictions, refund policy — and § 310.3(a)(4) prohibits misrepresenting any material aspect of the performance, efficacy, nature or central characteristics of the goods or services. A model that improvises a claim about a product in order to keep a conversation moving is generating a potential violation per call, at whatever rate the model does that, and the seller is liable for it. Section 310.3(b) extends liability to a person who provides substantial assistance while knowing or consciously avoiding knowing that the primary violator is engaged in prohibited conduct — which is the hook a platform vendor should read carefully.

Impersonation is now separately actionable. The FTC’s Rule on Impersonation of Government and Businesses, 16 C.F.R. Part 461, took effect on 1 April 2024 and makes it unlawful to materially and falsely pose as a government entity or business, or to provide the means and instrumentalities to do so. A synthetic voice presented as a named company’s representative when it is not, or as a real named individual, is squarely the conduct that rule was written for.

Behind both sits section 5 of the FTC Act, which reaches deceptive acts and practices without needing a specific rule — the authority the Commission has used against overstated AI claims generally, covered in the AI-washing page. A TSR violation is a rule violation with civil penalties attached; a section 5 claim needs no rule at all, and the two are routinely pleaded together.

The practical design conclusion is that the risky component is not the dialler but the generation step. Constrain what the agent can say to approved content, log every utterance for the recordkeeping obligations in § 310.5, and treat an unconstrained model on a sales call as an uncontrolled statement-making machine that the FTC will attribute to you.