The Carrier Registered Your Campaign. It Did Not Register What The Model Would Say.
SMS has two independent gates. The carrier gate decides whether your traffic is delivered and holds you to a declared use case. The TCPA gate decides whether sending was lawful at all and carries per-message statutory damages. Putting a language model between them means neither gate is checking the thing that actually goes out.
What A2P 10DLC Actually Governs
A2P 10DLC is the framework US carriers use to sanction application-to-person traffic over ordinary ten-digit numbers. A business registers a brand, then registers campaigns under it, declaring a use case — marketing, two-factor authentication, account notification, conversational support — along with sample message content, opt-in language, and the opt-in mechanism. Throughput and deliverability are provisioned against that declaration, and carriers filter or suspend traffic that departs from it.
This is a private governance layer, enforced commercially rather than by a regulator, and teams often treat clearing it as evidence that the program is compliant. It is not. Carrier approval says your paperwork matched your samples. It says nothing about whether the people on your list consented under the standard the TCPA imposes.
Where AI Composition Introduces Drift
A template program has a finite, reviewable message set. An agentic SMS program has a policy and a model. The gap between those two is where compliance failures live, and they are not exotic — they are the ordinary behavior of a helpful assistant that has been told to move the conversation forward.
Use case drift
A campaign registered as conversational support begins offering discounts because the model read the customer's hesitation as a buying signal. The traffic is now marketing under a non-marketing registration, and the consent record behind it was likely gathered for support, not promotion.
Missing or malformed disclosures
Sender identification and opt-out instructions are required content in practice, and a model composing freely will drop them when the message runs long or the conversation feels informal. Templates cannot forget; generated text can.
Opt-out interpreted rather than executed
If revocation detection lives in the prompt, a model can respond conversationally to 'stop texting me' instead of suppressing the number. The next message in the sequence is then a knowing violation, which is the trebling trigger.
Quiet hours and frequency
State telemarketing rules impose calling and texting time restrictions, and registered campaigns declare message frequency. An agent that replies whenever a customer replies has no concept of either constraint unless the send layer enforces it.
Unprovable message content
The strongest TCPA defense is a complete record of consent plus the exact text sent. Programs that log the prompt and the parameters but not the rendered outbound message have kept the wrong artifact and cannot reconstruct what the plaintiff received.
The Consent Standard Underneath
For marketing messages sent with an automated system, the TCPA requires prior express written consent: a signed agreement, clear disclosure that the consumer will receive marketing texts from a specifically identified sender, and an express statement that consent is not a condition of purchase. Non-marketing informational messages sit under a lower standard, which is precisely why use case drift is dangerous — a program built on informational consent that starts selling has moved to a standard its consent record was never designed to meet. State mini-TCPA statutes in Florida, Oklahoma, and elsewhere add their own consent and timing requirements with independent private rights of action, so a national program is complying with the strictest applicable rule, not the federal one.
Controls for an AI SMS Program
- ☐Store the opt-in timestamp, the exact disclosure text shown, the source URL, and the IP or device identifier
- ☐Version disclosure copy so an old subscriber's record resolves to the wording they saw
- ☐Keep marketing consent separate from transactional consent and never infer one from the other
- ☐Confirm consent is not presented as a condition of purchase anywhere in the flow
- ☐Detect opt-out keywords and free-text revocation before the message ever reaches the model
- ☐Enforce sender identification and opt-out instructions in the send layer, not in the prompt
- ☐Validate every outbound message against the registered use case before dispatch
- ☐Enforce quiet hours, frequency caps, and per-number suppression outside the model
- ☐Register a campaign per use case; never run marketing under a support or notification registration
- ☐Update sample messages when the agent's behavior or prompt materially changes
- ☐Match the opt-in language you registered to the language actually on your form today
- ☐Monitor carrier filtering and error codes as a compliance signal, not just a deliverability metric
- ☐Log the rendered outbound text of every message, not the prompt or the template id
- ☐Retain suppression list history showing when each number was removed and why
- ☐Keep records for at least the applicable limitations period for TCPA claims
- ☐Run a periodic sample review of live conversations against the declared use case
Your opt-in form is the evidence
Consent defenses are won or lost on what the sign-up page displayed, and that page drifts every time marketing edits it. RatedWithAI scans your live pages and reports what is actually rendered to a visitor.
Scan Your Site for Free →Frequently Asked Questions
Our messaging provider handles STOP automatically. Is opt-out covered?
Standard keyword handling at the provider covers the common case and not the hard one. Revocation communicated in ordinary language, or sent to a different number belonging to the same brand, will not trigger keyword suppression. Build your own free-text revocation detection ahead of the model and suppress the number across every campaign you run, not only the one that received the reply.
Does an existing business relationship let us text a customer marketing offers?
Not on its own for marketing messages sent with an automated system. That category requires prior express written consent regardless of relationship. An existing relationship is relevant to some informational messaging and to certain state rules, but treating it as a general permission for promotional texting is one of the most common and most expensive misreadings of the statute.
We only send when a customer texts us first. Do these rules apply?
Inbound-initiated conversational messaging sits in a much better position, because the customer's message supplies context and often consent for a reply. The risk begins when the agent extends beyond that exchange — following up days later, introducing an offer, or re-engaging a dormant thread. Each of those is a new outbound message assessed on its own consent basis, not on the original inbound one.
Can we point to the model's system prompt as evidence of our compliance policy?
It helps establish intent, and it will not establish behavior. A prompt is an instruction, not a control, and the plaintiff will have the message that was actually delivered. Where a requirement is absolute — opt-out, sender identification, quiet hours — enforce it in code between the model and the carrier, and keep the prompt as a second layer rather than the only one.