AI text replies for business: how to use them safely

An AI text response is a reply to a customer's text that a language model drafts. It either sits as a suggestion a person approves or sends on its own. Use suggested replies freely. Let replies send automatically only for narrow, factual questions, and keep STOP handling, quiet hours and a fast human handoff outside the model's control. The rest of this post covers where the line between suggested and automatic should sit, which guardrails matter, and what an AI text response costs per segment.

We make Ready, so here is what it does and doesn't do in this area. Ready no longer lets accounts create AI reply agents. One set up earlier may still answer inbound texts on its own until you pause its automation in Automations, and the Take over button in the Inbox thread stops it in that conversation. A new account can't set up an AI reply agent, so a person writes the replies. The only texts Ready sends on its own are fixed ones you write, such as the out of hours auto-reply and keyword replies. It has Ask AI, an in-app assistant that drafts texts, checks them for spam words, answers questions and takes actions on your account. If you build your own agent, Ready's API and MCP server can carry its messages, and the compliance layer still applies to every text. Where this post describes automatic replies, it is describing a system you would build or buy, not a Ready feature.

What an AI text response is, and the two ways teams use it

Almost every product in this category falls into one of two modes.

Suggested replies (human in the loop)

The model reads the conversation and drafts a reply. A person edits it, approves it or throws it away. Nothing goes out without someone pressing send.

This is the low-risk version. A person stays on every message, so the AI's job is to save typing, not to make decisions. If the draft invents a discount or misstates your hours, the person catches it before the customer sees it.

Automatic replies (no human per message)

The model reads an inbound text and sends its answer directly. This is where the speed comes from. It is also where the risk comes from: every message the model sends is a message your business sent, under your registered 10DLC brand, with your name on it.

A useful way to decide between the two modes: if a wrong answer would cost you money, a customer or a compliance problem, a person approves it. If a wrong answer would only be mildly annoying, automation might be fine.

Where AI replies go wrong

The failure modes are predictable, which makes them easier to guard against.

• Invented facts. The model makes up a price, a policy, a delivery date or an appointment slot that doesn't exist. Customers screenshot texts, so the invented answer becomes a promise you have to deal with.
• Arguing with STOP. A model trained to be helpful may answer "STOP" with "Sorry to see you go! Can I ask why?" That is the opposite of what the rules require. An opt-out has to be honored, not negotiated.
• Scope creep. A reply bot set up to answer "what are your hours?" ends up sending promotional follow-ups to people who only asked a question. That turns a conversational exchange into marketing, and marketing needs its own consent.
• Off-hours sends. A customer texts at 10:40 PM and the bot replies instantly. Quiet-hours rules are based on the recipient's local time, and an automatic reply doesn't get an exemption because it was fast.
• Filtered content. AI drafts tend toward breathless marketing language, shortened links and emoji. Carriers filter on content even after a campaign is approved (we covered that in why approved campaigns still get filtered).
• No exit. The customer is upset or confused, and the bot keeps cheerfully answering. Nobody on your team finds out until the customer goes elsewhere or files a complaint.

Guardrails that belong in any setup

Whichever mode you pick, these guardrails do the actual work. The key point is that most of them should sit outside the model, enforced by code or by your messaging layer, not by a line in the prompt asking the model to behave.

1. Opt-outs are handled before the model sees anything. STOP, UNSUBSCRIBE and similar keywords should never reach the AI as a "message to answer." On Ready, inbound STOP is honored automatically and the opt-out carries across campaigns. A send to an opted-out number through the API is refused up front and not charged, even if your agent tries.
2. Quiet hours are enforced at send time. On Ready, a 1:1 or API text sent during the recipient's quiet hours is held only if the window opens within 6 hours; otherwise it is refused and your system has to resend it. A held text is charged only when it sends. That protects you from a bot answering at midnight. Tell the model this can happen, so it doesn't write "I'll reply right away" in a message that may be held or refused.
3. The model answers from a fixed source. Give it your hours, policies, pricing and FAQ as reference text, and instruct it to say "let me get someone to confirm" when the answer isn't there. Most invented facts come from a model asked to answer without a source.
4. An allowlist of topics for automatic replies. Hours, address, "did you get my form?", order status from a real lookup. Anything else goes to a person.
5. Hard handoff triggers. Anger, legal words ("lawyer", "sue", "report"), refund disputes, health information, and any message the model gives low confidence on all route to a human (more on this below).
6. No new marketing from a reply flow. A reply bot answers the question it was asked. Promotions go out as campaigns to contacts who agreed to marketing texts.
7. Content checks before anything goes live. Run your reply templates and typical drafts through a spam-filter check. In Ready, Ask AI checks a text for spam words when you ask. See AI text message generator for business texts for what tends to get filtered.
8. Logs a person actually reads. Review a sample of automatic replies every week. If nobody reads them, nobody will notice a drift until a customer does.

Consent and compliance: what the AI can't decide

A model can draft a reply. It cannot decide whether you're allowed to send it.

Replying to someone who texted you first is generally treated differently from texting someone out of the blue. Answering the question they asked is a conversational reply. Following up later with offers is a different message type, and it needs marketing consent.

Cold contact isn't an AI use case. Cold outreach belongs on the phone and in the mail. SMS goes to people who opted in or replied. An AI that writes faster doesn't change that.

Consent records are your job. When you send through the Ready API, you set consent_attested: true on POST /v1/messages or POST /v1/messages/bulk. That field is your promise that the recipient agreed to texts. Ready records the sender's attestation for each bulk send and each API send that sets consent_attested: who sent, to how many, and when. That builds an audit trail of your sends. It is not a per-contact consent record. Ready doesn't store each recipient's opt-in time, the wording they agreed to or the page URL, and a keyword reply doesn't create one. Your form, checkout or CRM has to keep that record. (A contact who signs up on a Ready-hosted opt-in page is the exception: Ready stores that sign-up, with the time, IP address and exact wording, and support@tryready.com can send you a copy.)

The stakes are real. TCPA damages are commonly cited at $500 to $1,500 per text. An automatic reply flow sends many texts quickly, so a consent mistake multiplies quickly too. The guardrails above reduce your risk; they don't make you lawsuit-proof. Compliance is ultimately the sender's responsibility, and that includes a sender whose replies are written by a model.

If you're adding marketing to a list, the standalone TCPA & DNC Litigator Scrub runs $0.005 per contact. It checks each number against known TCPA-litigator and DNC-complainer lists and flags matches before you send. The math on scrubbing vs one lawsuit walks through the numbers.

Handing off to a human

The handoff is the part most AI text setups get wrong. A good handoff has three properties.

It's triggered by rules, not just the model's judgment. Keyword triggers (refund, cancel, lawyer, emergency), sentiment triggers and "the question isn't in the source material" should all force a handoff. Don't rely only on the model deciding it's out of its depth.

The customer is told. A short line works: "Good question, I'm passing this to someone on our team who'll text you shortly." That beats an AI pretending to be a person, and it sets an expectation your team then has to meet.

The human has the full thread. On Ready, inbound replies land in the Inbox, and for GoHighLevel-connected accounts they also sync into GHL. Whoever picks up sees what the customer asked and what was already said. If a call is faster than more texting, the Ready Dialer handles click-to-call from the dial pad in the dashboard or GoHighLevel. It's billed at $0.025/min, with each started minute billed as a full minute. It is a manual dialer: no auto-dialing and no voicemail drop.

Pick a target time for human pickup and staff for it. A handoff that sits for six hours is worse than no AI at all, because the customer was told someone was coming.

Building your own reply agent on Ready

If you want automatic replies, you'd build or buy the agent and use Ready as the messaging layer. There are two routes.

• The REST API. Your agent sends with POST /v1/messages (header X-API-Key, plus an Idempotency-Key UUID so a retried request never sends twice), and reads conversations with GET /v1/messages. Before it texts a number, it can check GET /v1/opt-outs/:phone. Request shapes are in the messages docs.
• The MCP server. Ready runs a remote MCP server at https://api.tryready.com/mcp. Agents in Claude, ChatGPT custom connectors, Claude Code or Cursor can send texts and manage contacts and conversations. A test key simulates send_sms and refuses every other tool that changes or bills something, so you can watch an agent's behavior before it touches a real phone. Details are in the MCP docs and in SMS MCP server: let AI agents send texts safely.

Either way, the opt-out refusal, quiet-hours holds and landline refusal sit in Ready's send path, not in your prompt. A badly behaved agent still can't text someone who opted out.

What an AI text response costs

AI doesn't change the per-text price. A drafted reply bills exactly like a typed one. On Ready Standard, that is $0.0075/segment plus a $0.0045 carrier fee per segment. Replies you receive are billed the same way. That all-in figure excludes the optional compliance surcharge: scrubbing and landline validation are each $0.003 extra per segment when switched on for an account, and both are off by default.

Segments matter more than people expect with AI drafts, because models write long:

Worked example (assumptions stated): 1,000 conversations a month. Each has one inbound text of 1 segment and one AI-drafted reply of 300 characters.

• Without emoji: 1 + 2 = 3 segments per conversation, so 3,000 segments. That's 3,000 × $0.0075 = $23, plus 3,000 × $0.0045 in carrier fees.
• With one emoji in each reply: 1 + 5 = 6 segments per conversation, so 6,000 segments. That's $120 plus 6,000 × $0.0045.

One emoji doubles the bill in this example. Telling the model "plain text, under 160 characters, no emoji" is a cost control as much as a style rule.

Ready isn't the lowest per-segment price on the market. Raw CPaaS APIs and GoHighLevel's LeadConnector cost less per message. What you get with Ready is registration handled in-app (most approvals take 2 to 7 days), the inbox, and the compliance layer in the send path. Your first 20 one-to-one texts are free, so test your reply flow with those before you register.

The practical takeaway

Start with suggested replies, with a person approving each one. Move narrow, factual topics to automatic only after you've read a few weeks of drafts. Keep STOP, quiet hours and opt-out checks in the messaging layer rather than the prompt, and make the human handoff fast and visible. For more on which texts to hand to software at all, see automated texting service: which texts to automate.

If you want to try the drafting side, ask Ask AI, the assistant in Ready, to draft a reply or check a text for spam words: create an account.

FAQ

Does Ready send AI replies to my customers automatically?

Not for new accounts. Ready has Ask AI, an in-app assistant that drafts texts and checks them for spam words, and new AI reply agents can't be set up. Ask AI never sends a reply on its own. If you build your own agent, it can send through Ready's API or MCP server, and that agent is your system.

Is an AI text response legal under the TCPA?

The TCPA covers the message and the consent behind it, not who or what wrote the reply. A reply to someone who texted you first is generally lower risk than outreach. Marketing still needs documented consent, and compliance stays the sender's responsibility.

What happens if a customer texts STOP to an AI reply flow?

On Ready, inbound STOP is honored automatically and the opt-out carries across campaigns. Any later API send to that number is refused and not charged. Your agent should also be built to treat STOP as final rather than reply to it.

Do AI-written replies cost more to send?

Not per segment. They do tend to run longer, though, and an emoji switches the text to 70-character segments. A 300-character draft is 2 segments as plain text and 5 segments with one emoji.

When should an AI reply hand off to a person?

Hand off on refunds, complaints, legal language, health details, anything outside the reference material, and any sign of frustration. Tell the customer a person is taking over, then have someone pick up quickly with the full thread in front of them.