Skip to content
be-human.ai
← Back to the journal

Human oversight

Texting Rules for AI Agents: Consent, Opt-Outs and Quiet Hours

Set consent capture, opt-out handling and quiet-hour rules before an AI agent sends any text message. A practical checklist for business owners.

Be Human typographic cover with the title: Texting Rules for AI Agents: Consent, Opt-Outs and Quiet Hours
Conceptual illustration created with AI · Be Human

Before an AI agent sends a text message, a company needs three rules in writing: proof that each recipient agreed to receive texts from that company, a single way to stop messages that works from any channel, and send windows that respect the recipient's local time. An agent that skips any of these can create legal exposure and damage customer trust within one campaign. This guide sets out what to enforce first and how to check it.

Key takeaways

  • Record consent as a dated event with a source and the exact wording the person saw, not as a yes or no flag in a spreadsheet.
  • Treat every stop request, in any language and on any channel, as a suppression that applies to all numbers linked to that person.
  • Set quiet hours and frequency caps in the sending system, so no prompt or instruction can override them.
  • Give the agent read access to suppression lists and no permission to edit them.
  • Run the same pre-send check on every new template and audience before its first live send.

Why texting needs rules before the agent's first message

A text message is one of the most direct channels a company has. It arrives on a personal phone, it is read quickly, and many recipients treat it as more personal than email. An agent that can draft and send messages at scale can make mistakes faster than a staff member could: texting a list that was never approved for texts, repeating a message after someone asked to stop, or sending at an hour that feels intrusive. Consent, opt-out and timing rules are therefore not a legal footnote added after launch. They decide what the agent is allowed to do at all.

Rules differ by country, carrier and type of message, so this article does not state what any one jurisdiction requires. It describes the controls a careful operator enforces and where legal review belongs. The NIST AI Risk Management Framework is a useful structure for naming these risks, assigning an owner to each one and deciding what evidence proves the control works.

How do you capture consent you can prove later?

Consent for texts should be captured at the moment a person gives their number, with the purpose stated in plain words. A record that says opted in is weak evidence. A record that shows who agreed, when, through which form or conversation, which sentence they saw, and which kinds of messages they accepted is strong evidence. Each consent record should hold the following fields:

  • The phone number, stored in one normalized format
  • The source, such as a web form, an in-person sign-up, a keyword reply or a staff entry
  • The exact consent text shown at that moment, with a version number
  • The message categories agreed to, such as appointment reminders or promotions
  • The timestamp and the recipient's time zone
  • Who or what recorded the entry, so agent entries can be told apart from staff entries

Imported lists are the most common failure. A list from an event, a trade show or an old spreadsheet may contain numbers with no trace of consent. The agent should not treat those numbers as eligible. Mark them as unverified, keep them out of every send, and reach them only through a fresh opt-in path that your legal reviewer has approved. Any change to the consent wording should create a new version, so that old records still show what people actually agreed to.

How should opt-outs work across every channel?

Opt-outs fail in the gaps between systems. A recipient replies STOP to a text, then emails support asking to be removed, then mentions it to a sales representative on a call. If those three events land in three places, the agent may keep texting someone who has already refused. The rule is simple to state: one suppression list, checked before every send, and updated from every channel where a person can object.

  • Recognize stop words and phrases in every language you serve, including Portuguese, and route unclear replies to a person rather than guessing
  • Write each opt-out to the central suppression list within the same workflow that received it
  • Apply the suppression to every phone number linked to that contact, not only the number that replied
  • If your sending provider or counsel requires a confirmation reply, send only that reply and nothing else
  • Keep the agent's access to the suppression list read-only, so that no prompt can remove a name
  • Review the list after every import, merge, integration change or change of sending vendor

For Brazilian companies, opt-out handling sits inside a wider set of obligations on personal data. The guide on LGPD and AI agents for Brazilian companies covers the questions to take to counsel before a messaging agent goes live.

How do quiet hours and frequency limits work in practice?

Quiet hours are a company policy, and they must be measured in the recipient's local time, not the server's time or the office's time. A business in one city may message customers in another, and an agent working from a default time zone will get this wrong unless the time zone is stored with each contact. Frequency limits matter just as much. A person who receives several messages in one day from the same business will feel it, even when each message was allowed on its own.

Illustrative scenario: A home services company's agent drafts a reminder for a Saturday appointment. The customer's phone is in a different time zone from the office. The agent schedules the send for the office's morning, which is late evening for the customer. The sending system holds the message until the customer's permitted window opens, and the log records the hold and its reason. A person then reviews the held message, as described in reviewing blocked agent actions.

To make quiet hours hold, apply these controls:

  • Store the recipient's time zone with the contact record, not inferred from the area code alone
  • Set the send window inside the sending system, not inside the agent's instructions
  • Cap messages per person per period, and count reminders and follow-ups together
  • Define any category that may bypass the window, such as a safety notice, and require a person to approve that bypass rule in advance
  • Hold messages rather than dropping them silently, so that a person can review what was held and why

What should a pre-send checklist include?

Before a new template or audience goes live, run the same checks every time. The point is repetition: a check that is done by memory will eventually be skipped. Work through this sequence and keep a record of each pass.

  • Step 1: Confirm the audience is built only from verified consent records, and note any unverified numbers that were excluded.
  • Step 2: Remove every suppressed number and every contact with an opt-out in the same message category.
  • Step 3: Check that the message names the sender and gives a clear way to stop, written in the language the recipient uses.
  • Step 4: Confirm that the message content belongs to a category the person agreed to receive.
  • Step 5: Check the send window and the frequency cap against each recipient's local time and recent message history.
  • Step 6: Have a named person approve the first live send of any new template, and log that approval with the date.
  • Step 7: Watch the replies for the first hour after sending, and pause the campaign if stop requests or complaints appear.

Where a template touches a high-impact decision or a sensitive category, add a second approver. The guide on where an agent should stop and ask explains how to decide which steps need a person.

How does Be Human handle texting rules in managed operations?

Be Human defines the agent's messaging permissions during the first engagement, before any live send. The consent record, the suppression list and the quiet-hour settings are configured in the sending system, and the agent receives only the access its role requires. Managed operations then include recurring review of the suppression list, checks after each import or integration change, and incident response when a stop request is missed. Approval boundaries specify which templates need a person's sign-off and which categories, if any, may bypass a send window. The operating record keeps every hold, block and approval, so the company can review a decision long after it was made. If you are not sure which texting task your team should hand to an agent first, the free Be Human AI diagnostic can help you find that starting point.

Frequently asked questions

Can an AI agent text contacts who never opted in? No. Contacts without a verified consent record should be excluded from every send, including contacts inherited from an earlier team or an old spreadsheet. A fresh opt-in path, approved by a person and logged, is the only safe way to bring them back into a messaging workflow.

Does the agent have to handle STOP replies itself? The agent should recognize a stop request and write it to the central suppression list, but it should not decide on its own how to interpret an unclear reply. Unclear messages should go to a person. Any confirmation reply should be the one your sending provider or counsel requires, and it should contain no marketing content.

Do quiet hours replace legal review? No. Quiet hours and frequency caps are operating controls that make a legal position enforceable in daily work, but they do not establish the legal basis for a message. Your counsel should confirm the basis for each category of message and the rules that apply in each country where your recipients live.

Should the agent be allowed to change its own send window? No. Changes to send windows, frequency caps or suppression rules should require a named person's approval and should leave a log entry. An agent that can loosen its own limits has no effective limit at all.

Your next step

See where an AI agent fits in your company first.

Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.

Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.

Less busywork. More room for you.

Bring us the work you wish you had time for.

Book a call to try a personalized AI agent free for 7 days ↗Or take the free AI diagnostic →