A tactical, from-the-trenches look at what AI can actually do for text message marketing right now, and where it still needs a human checking its work, from drafting variants to picking send times.
SMS is still the highest-engagement channel most marketers have, and I’m not saying that from a deck I skimmed somewhere, I’m saying it because it’s the channel where my clients’ numbers actually hold up quarter after quarter. Klaviyo’s 2026 benchmarks show flow-based texts pulling click rates near 10%, almost double standalone campaigns [1], and 71% of businesses say AI has already improved their SMS results [2]. But asking AI to “write a text” is close to the least useful thing you can do with it. The skill worth building is knowing when to reach for a general model like ChatGPT and when to trust your platform’s built-in AI instead, and never letting either one near a send button before someone checks the character count and the opt-out line.
A DTC skincare brand I worked with last spring flipped on their SMS platform’s built-in AI copy generator the same week they doubled send frequency. Two weeks later, their unsubscribe rate had tripled. The AI wasn’t broken. It did exactly what it was told: write more messages, faster, with punchier CTAs. Nobody told it to slow down.
I see some version of that mistake constantly. AI doesn’t diagnose a weak cadence for you, it just executes whatever cadence you hand it, faster and more confidently than you had going in. If your send frequency was already too aggressive, AI-generated variants mean five versions of “LAST CHANCE” instead of two, in about four minutes flat. A bad SMS cadence gets you unsubscribed faster than almost any channel I’ve worked in, email included.
I’ve been running SMS programs since brands were nervous about texting people at all, back when a 2% click rate felt like a small miracle. The channel itself hasn’t changed much. What exploded is the number of tools promising to automate your way to better performance, and nobody trained any of them on restraint, because restraint isn’t something you can prompt for.
Less AI isn’t the fix. Point it at variant testing and personalization instead of raw copy volume, and once those two habits stick, let it help with send timing too.
Email gives you a subject line, preview text, and as much body copy as you want to write. SMS doesn’t hand you any of that room. A standard text segment caps out around 160 characters in plain GSM-7, and drops to roughly 70 characters the moment you add an emoji or a curly apostrophe, because that switches the encoding to Unicode and splits the message into more, and more expensive, segments. Most platforms will still send longer messages as multiple linked segments, and you don’t notice you’ve blown past the limit until the invoice shows up.
I’ve had a client’s opt-out line quietly flip an entire flow into Unicode because someone pasted a right single quotation mark from Word instead of a straight apostrophe. A 158-character message that should have fit in one segment became two, for every send, for about six weeks before anyone caught it on the billing report. Small detail, real money.
General-purpose AI tools like ChatGPT and Claude don’t know any of this unless you tell them. Ask for “an SMS promoting our sale” and you’ll often get back 220 characters of enthusiastic copy that reads like an email teaser wearing a text message’s clothes. Prompt for character count and format every time, or save a template that bakes those in.
Then there’s the opt-in language, and this is where I get careful. Every legitimate SMS message needs clear sender identification, a working way to opt out, usually some version of “Reply STOP to unsubscribe,” and consistency with whatever the subscriber agreed to at signup. I’m not walking you through TCPA specifics here, the rules are jurisdiction-specific and they change, so talk to your platform’s compliance team or your own counsel before launching anything new. Tactically: bake the opt-out line into your AI prompt templates, and double-check it survives the character count, because I’ve seen “Reply STOP” get quietly trimmed off a draft that ran two characters over.
Here’s the honest breakdown, from actually running both sides of this on real campaigns.
General AI models are better brainstorming partners, hands down. They’re faster for generating a wide spread of angles (“write me 10 different hooks for a restock announcement”), adapting a brand voice from a style guide you paste in, and drafting longer-form flow copy before you chop it to SMS length. What they don’t know is your subscriber list, your past send performance, or your compliance history, so everything they generate still needs a human checking it against real data. I’ve had Claude write a genuinely great five-message winback sequence in about ninety seconds. I’ve also had it confidently suggest a subject-line-style teaser for a channel that doesn’t have subject lines, forgetting mid-conversation that we were writing texts.
Klaviyo’s AI SMS assistant, Attentive’s Magic Message, Postscript AI, and SimpleTexting’s AI Assist all plug directly into your account’s send history and subscriber behavior. Attentive AI is trained on more than 1.4 trillion data points pulled from 40+ billion messages across 70+ verticals, and its Copy Assistant for SMS taps that historical performance data to generate copy tailored to a specific audience and goal.[4] Postscript’s Infinity Testing goes further: it tests thousands of variants inside live flows, and Postscript reports an average 20% revenue increase on a brand’s highest-ROI flows.[3]
One thing platform AI still doesn’t do: read the results with judgment. Infinity Testing will tell you variant B beat variant A by a solid margin, and if your sample size is small, that margin can be noise dressed up as a signal. I wait for at least a few hundred sends per variant before trusting the winner, and I glance at the unsubscribe rate next to the click rate too, because a variant that wins on clicks and quietly spikes opt-outs isn’t really a winner.
Platform AI is locked to what it can see and do. For a genuinely fresh angle, a general model is the better start. For AI that already knows what worked for your list last quarter, the platform tool wins, no contest.
Most marketers A/B test SMS by accident. They write one message, tweak the emoji, and call it a test, when really it’s a guess with an extra step tacked onto it.
A better approach is to use AI to generate four to six variants that each change exactly one variable. Say you run SMS for a coffee subscription brand, we’ll call it Northline Coffee Co, and want to text your “paused subscription” segment about restarting with a discount. Instead of “write a text about restarting,” try: “Write 5 SMS variants under 160 characters offering 15% off a restarted subscription to Northline Coffee. Each variant should test a different opening line: urgency, curiosity, a direct offer, social proof, and a question. Include the opt-out line at the end.” That gives you a real test matrix instead of five versions of the same idea.
Rule of thumb: never test more than one variable per pair of messages, or you won’t know what actually moved the needle.
This is also where built-in platform tools earn their keep over a general model. Postscript’s Infinity Testing runs this kind of variant testing automatically inside your flows and keeps optimizing after launch, instead of you manually swapping in a new draft every week. I still check in on those flows every couple of weeks, because I’ve been burned once by a flow that kept running a variant which had technically won on clicks but was quietly costing us list health.
Personalization in SMS used to mean {{first_name}} and nothing else, which is table stakes now, and honestly was table stakes five years ago too. It’s still one of the ways brands get it wrong. Nothing screams “you’re on a list” faster than a text that opens with your name and has nothing useful to say after it.
Real personalization means the AI has context on what the subscriber actually did: browsed a product and didn’t buy, purchased once and never came back, hit a VIP spend threshold. We’ve written a full personalization playbook that goes deeper on building these behavioral segments, but for SMS specifically, the AI’s job is turning that data into a message that reads like it was written by someone who actually knows the customer.
For Northline Coffee Co, that might mean the “paused subscription” text mentions the actual roast they paused, not a generic “come back and save.” That single detail, pulled from order history rather than guessed at, is what separates AI personalization that converts from AI personalization that gets reported as spam.
One nuance worth knowing: personalization tokens eat into your character budget too. A first-name token might add two characters or fourteen, and if you’re already close to the 160-character GSM-7 ceiling, a long name can quietly push part of your list into two-segment sends without you touching the template. I build in a buffer of about fifteen characters under the cap for this. Learned that one the expensive way.
Timing matters more in SMS than in almost any other channel I’ve worked in, because a text interrupts whatever someone’s doing. Klaviyo’s 2026 SMS benchmarks, pulled from data across more than 183,000 customers, found that flow-based (automated, behavior-triggered) SMS messages get click rates nearing 10% on average, almost double what standalone campaigns pull, with top performers clearing 16%.[1] That gap makes sense once you say it out loud. A message triggered by something the customer just did, like abandoning a cart, lands at a moment that’s actually relevant to them, instead of whenever your send calendar happened to schedule it.
This is where platform AI earns its keep over general AI tools. Attentive’s Automated Campaigns feature uses anonymized insights from high-performing sends across its customer base to build segments and pick send times automatically. Postscript offers more than 40 segmentation filters and 85+ trigger options so you can slice your list by behavior, not just demographics. None of that is something ChatGPT can do, since it has no access to your subscriber data or send logs.
I’ve also learned to check the time zone settings on every new flow before it goes live, not after. I once inherited an account where a winback flow had been firing at 6:45am local time for four months because the platform defaulted to the billing address time zone instead of each subscriber’s actual location. Nobody complained loudly enough to flag it as an emergency, they just quietly opted out, which is worse, because you don’t get a warning before the unsubscribe number creeps up.
If you haven’t mapped out where SMS actually sits in your customer’s path to purchase, our customer journey mapping guide is a useful next stop, because send-time optimization only works if you already know which moment in the journey you’re optimizing for. Pair that with a real look at your segmentation strategy before you let any platform’s AI pick your triggers for you.
A quick, honest caveat before we go further: I’m a marketer, not a lawyer, and SMS compliance rules, including TCPA-related requirements in the US, are detailed, jurisdiction-specific, and they change more often than most marketing blogs bother to update. Nothing in this article is legal advice. If you’re setting up or scaling an SMS program, get your opt-in flows, consent language, and quiet-hours settings reviewed by your platform’s compliance team or your own counsel before you launch, not after your first big send.
What I can tell you from a copywriting seat: build your compliance guardrails into your AI prompts so they’re never optional. That means every prompt template includes the opt-out instruction, respects your platform’s built-in quiet-hours settings, and never asks AI to imply a purchase is required to redeem an offer when it isn’t. Klaviyo, Attentive, Postscript, and SimpleTexting each publish their own compliance guidance specific to their sending infrastructure. Read your platform’s version, not just a general blog post like this one. I keep a running doc of opt-out phrasing my clients have had approved by their own counsel and paste from that, verbatim, instead of trusting AI to phrase it fresh.
Here’s how I’d actually run this for Northline Coffee Co, or for your own list, starting Monday morning, assuming you’ve got at least a quarter of send history to work from.
I’ll admit the first time I ran this exact workflow for a client, I skipped step one, jumped straight to drafting because the deadline was tight, and the AI-generated variants looked great in isolation. They flopped, because that segment had already been hit with three campaigns the same week and nobody had checked before adding a fourth. I don’t skip step one anymore. Ever.
None of this requires a bigger team or budget. It requires treating AI as a tool that speeds up testing and personalization, not a replacement for actually looking at what your list does. Most of the accounts I’ve watched get SMS right this year don’t have the fanciest AI stack. They still read their own unsubscribe report every week, by hand.
It’s safe as a first draft, never as a straight-to-send button. AI is good at generating variants and matching brand voice, but someone still needs to check every message against your character limit, opt-out language, and platform’s rules before it goes out, especially on anything set to fire automatically. I’ve seen the opt-out line go missing from an AI draft more than once.
ChatGPT and Claude are stronger for brainstorming a wide range of angles and drafting brand-voice copy from scratch, which is where I start most campaigns. Built-in tools like Klaviyo’s AI SMS assistant, Attentive’s Magic Message, Postscript AI, and SimpleTexting’s AI Assist plug into your actual subscriber data and send history, so they can personalize, segment, and pick send times using information a general AI model never sees.
Aim to stay inside a single 160-character segment, or roughly 70 characters if you’re using emojis or curly quotes, since that switches the encoding from GSM-7 to Unicode. Most platforms will still send longer messages as multiple linked segments, they just cost more and can arrive slightly out of order on older phones. Shorter, punchier copy tends to perform better anyway, and it sounds more like a text and less like an ad.
AI can help you consistently include opt-out language and stick to approved templates, but it can’t replace legal review, full stop. TCPA and related SMS rules are detailed and they change, so treat AI as a consistency tool, not a compliance officer, and get your opt-in flows and consent language checked by your platform’s compliance team or your own counsel. I say that as someone who is not a lawyer.
It depends on what you need. Klaviyo’s AI SMS features are strongest if you’re already running email and SMS together in one CRM. Attentive leans into generative copy trained on a massive message dataset. Postscript’s Infinity Testing is built for variant testing inside Shopify flows, my go-to on Shopify accounts. SimpleTexting’s AI Assist fits smaller teams. Test the free trial on your actual list before committing.
I pulled the numbers in this piece directly from Klaviyo’s 2026 SMS Marketing Benchmarks report, SimpleTexting’s 2026 SMS Marketing Statistics survey of 1,000 consumers and 400 businesses, and the live product pages for Attentive AI and Postscript AI, all fetched in July 2026. The anecdotes and workflow are from my own campaign work; the platform comparisons are cross-checked against each vendor’s own documentation rather than secondhand reviews, since vendor feature pages change often enough that a six-month-old post about them is usually already wrong somewhere.
Weekly tips. Real examples. Practical help for busy professionals.
We care about your data, check out our privacy policy.