SMS automation that customers do not block

Text messages get read. That is the whole appeal, and it is also why the tolerance for getting them wrong is so much lower than email. Somebody who dislikes your email unsubscribes. Somebody who dislikes your texts reports you, and carriers act on that.

This guide is what to automate, what not to, and the settings that decide whether it works.

Three kinds of text, and only one of them is easy

The categories are not degrees of the same thing. They have different legal bases and should have different rules in your software.

Conversational. Somebody texted you and you are replying. This needs no prior consent, because they started it. It should also be exempt from your quiet hours: a customer who messages at 22:40 is awake and waiting, and holding your reply until 08:00 breaks the thing they contacted you for.

Transactional. An appointment reminder, a confirmation, a change. This needs consent that specifically covers appointment messages, and quiet hours absolutely do apply, because nothing about a reminder needs to arrive at 03:00.

Marketing. A promotion, a reactivation campaign, a review request. This needs its own express written consent. The permission you collected for reminders does not cover it, whatever the software lets you do.

Most systems collapse these into one switch. A system that does is either waking people at 3am or refusing to answer them at 11pm, and both are worse than having the branch.

Your quiet hours are probably measured against the wrong clock

Nearly every system stores quiet hours against the business's timezone. That is the wrong reference point and it produces a specific, common failure.

A practice in California sends at 20:00 Pacific. A patient in Florida receives it at 23:00. The setting said 21:00, the send was inside it, and the person was asleep.

The clock that matters is the recipient's, and it can be inferred from their area code. Ask your vendor which one they use. If the answer is the business's, every customer you have outside your own timezone is getting messages at the wrong hour, and you will never hear about it from the ones who quietly stop replying.

There is a second layer on top: some states set a narrower window than the federal default. Florida and Oklahoma are the ones most often cited, both effectively 8am to 8pm. A system that applies one window to everybody is applying the wrong one to some of them.

A frequency cap is not optional

Set a maximum number of business-initiated messages one person can receive in a rolling day, and make it apply across every feature you have. The cap exists to stop a runaway: two automations that each seem reasonable, firing on the same person, on the same afternoon.

A ceiling of around half a dozen a day stops the disaster without interfering with normal use. It is a ceiling, not a target - a business regularly approaching it has a different problem.

Two details worth getting right:

• Replies to a customer should never count. The cap is about messages they did not ask for. Counting your answers against it means a long conversation ends with your software refusing to speak.
• The cap is per phone number, not per contact record. Two family members sharing a phone is one phone buzzing, and to the person holding it the distinction does not exist.

STOP and HELP have to work, and they are not yours to design

Carriers expect both on every number, and they check.

STOP must halt everything immediately, on every thread, for that number and that business. Scope matters in both directions: it must not be limited to the one conversation it arrived in, and it must not spill across to other businesses using the same platform.

HELP must reply with who you are, how to reach a human, and how to opt out. This is the one message that goes to somebody who has already opted out, which is exactly why the wording is worth having read by a lawyer rather than accepted from a default.

Both should be handled before the message ever reaches your AI. An automated reply to a STOP is the single worst message your system can send.

Before any of this: registration

In the United States, application-to-person texting from a normal 10-digit number requires registration through The Campaign Registry: a Brand for your business and a Campaign for what you send. Carriers filter unregistered traffic.

The failure mode is nasty, because it is silent. The message is accepted, you are billed, your software says "sent", and the customer receives nothing. The thread reads to them as though you ignored them.

Two things to insist on. Start registration weeks before you need it, because carrier review is measured in days and is entirely outside your control. And check that your software refuses to send from an unregistered number rather than handing it to the carrier to drop. A refusal you can see is a problem you can fix.

The four numbers that tell you it is going wrong

Watch these, in this order:

1. Opt-out rate. The clearest signal you have. A rising rate means the content, frequency or timing is wrong, and it is worth stopping to find out which before it becomes the next number.
2. Delivery failures by carrier error code. A sudden cluster is usually filtering rather than bad numbers, and filtering is a warning shot.
3. Reply rate on conversational threads. If it falls while your send volume rises, you are training people to ignore you.
4. Messages per contact per week. The number you will be most surprised by, because no single automation feels excessive.

What to automate, concretely

Worth automating: appointment reminders, confirmations, changes, a reply to a missed call, and answering inbound questions. All of these are things the customer is already expecting or has just asked for.

Worth thinking hard about: review requests, which need their own consent, and reactivation campaigns, which need the same.

Not worth automating at all: anything where being wrong is expensive and the software cannot tell. Cancellation of a clinical appointment, anything involving a price you may not honour, and anything that reads as a diagnosis. Those should produce a draft a person sends, which costs you seconds and saves the one message you would have regretted.

The short version

Separate conversational, transactional and marketing, and give them different rules. Measure quiet hours against the recipient's clock, not yours. Cap the daily total, exclude your own replies from it, and count per phone. Get STOP and HELP right before anything clever. Register before you send, and make sure your software refuses rather than letting the carrier drop it. Then watch the opt-out rate, because it will tell you before your customers do.