Who you are allowed to message
Consent, bot disclosure, and the people in a group thread who never agreed to anything. What the rules say, with the sources, for anyone building an agent that texts.
Three questions decide whether an agent that texts people is on solid ground, and only one of them is about registration. May you message this person. Must the agent say it is software. And what about everyone else in the thread.
This page reports what published sources say and links them. It is not legal advice, we are not lawyers, and consent under the TCPA is a legal question rather than a procedural one. If you are about to send at volume, this is the point to involve someone qualified.
Consent, and what it costs to get wrong
Automated texts to US mobile numbers sit under the TCPA, and the same consent rules apply to AI-generated messages as to any other automated ones. Marketing requires prior express written consent. Opt-out keywords must be honoured, within ten business days at the outside.
The exposure is what makes this worth reading twice. Statutory damages run $500 per violating message, rising to $1,500 for knowing or willful violations, each message counted separately and with no cap. A campaign of a hundred thousand messages sent without adequate consent is a nine-figure class action, which is a category of mistake a small team does not recover from.
One thing that has changed and is still misreported: the FCC's one-to-one consent rule, which would have stopped a single consent being sold to multiple callers, was vacated by the Eleventh Circuit in January 2025 and formally deleted by the FCC in September 2025. Guidance written between those dates and now is unreliable on this specific point.
How the first message actually gets sent
Here is a mechanic nobody documents, and it sits exactly where consent does.
Four agents in this directory publish a link that opens Messages with the first message already written. Not a number to text, a message to send.
| Agent | What it pre-writes |
|---|---|
| Arlo | "hey Arlo" |
| Ember | "Hey Ember 🔥" |
| Shuffle | "hey shuffle, im new here... what can u actually do?" |
| Sidekicks | "Hello 👋" |
Read Shuffle's again. Lowercase, an ellipsis, a casual register: written by the company to read as though the user typed it themselves.
The mechanic matters because of the direction it establishes. The person sends the first message, so every subsequent exchange is a reply to inbound contact rather than an unsolicited approach, and that distinction is central both to how carriers assess a registered campaign and to how consent is argued. Apple's own Messages for Business programme is built on the same principle from the other end: conversations there are customer-initiated by design, which is a legal architecture as much as a product one.
None of that makes a prefilled message insufficient. It does mean the consent story of these products rests on a sentence the vendor wrote, and that if you are building one you should be able to say clearly what your user agreed to and where that is recorded.
Must the agent say it is a bot
Two Californian laws, and the newer one is the one to read.
SB 1001, effective July 2019, prohibits using a bot to communicate with someone in California with intent to mislead about its artificial identity, where the purpose is to incentivise a sale or influence a vote. Clear and conspicuous disclosure is a safe harbour, and liability falls on whoever deploys the bot rather than whoever built it.
SB 243 took effect on 1 January 2026 and is more consequential for this market. It covers companion chatbots, meaning AI systems that simulate human-like relationships with their users, and requires disclosure where a person could reasonably believe they are talking to a human. It carries disclosure, safety and annual reporting obligations, and unlike its predecessor it includes a private right of action.
Look at what this directory contains and the relevance is obvious. One agent describes itself as "your smartest friend in iMessage, the friend who never leaves you on read." Others do coaching, accountability and social introductions. A product designed so that "nobody in the group needs to know what it is" is describing, in its own marketing, the situation the statute addresses. No vendor page we read mentions either law.
The people who never agreed to anything
This is the part the category is built to avoid, and it is where the directory's own research is most useful.
An agent in a group thread reads the messages of everyone in it. An agent that texts your friends on your behalf processes the data of people who never saw any terms. Of the twenty two privacy policies we have read, eight acknowledge this at all.
Jarvie is unusually direct: "In a group, Jarvie processes the thread's messages to respond, and its replies are visible to everyone." 222 goes further and states it collects information about people from existing users' phone contacts and from publicly available sources, which means it holds data on people who have never used it. Poke notes it may receive personal data from other customers, "for example, a customer's email inbox may contain your contact information."
Most simply write as though the account holder is the only person in the conversation. If you are building a group agent, that is the gap to close before someone else notices it, and the full comparison is on what these agents do with your data.
What to do about it
- Be able to show where consent came from. Campaign registration asks; a plaintiff will too. See A2P 10DLC registration and the readiness checker.
- Make opt-out work on the first message and every one after it. Not a later feature.
- Say the agent is software where a person could reasonably think otherwise. It is a safe harbour under one law and an obligation under another, and it costs you a sentence.
- Decide what your agent tells third parties. If it messages someone who never signed up, they should be able to find out what you hold and get rid of it.
- Agents that place calls raise recording and disclosure questions on top of all of this, which vary by state and which we have not researched here.
The cheapest version of all of this is deciding it before launch. The expensive version is deciding it after a demand letter.