What AI changed isn’t the message. It’s the trigger. The old model was calendar-driven: pick a date, pick a list, write a blast, send. The new model is behaviour-driven: a model watches what a customer does, predicts what they’ll do next, and fires a single well-timed message. Same channel. Completely different economics.
Here’s where that’s actually working.
Calendar-driven
Pick a date. Pick a list. Write a blast. Send to everyone at 11am and hope. Volume is the lever, and opt-outs are the cost of pulling it.
Behaviour-driven
A model watches behaviour, predicts intent, and fires one message at the moment it matters — to the people it matters to. Fewer sends, better outcomes.
AI hasn’t replaced SMS — it has changed when and why a message is sent. The gains come from triggering, not writing: predictive churn scoring, per-recipient send-time optimisation, intent-scored abandonment, and LLM-backed two-way replies. Every use case below is valuable because it sends fewer, better-timed messages. None of it works without delivery, DLT registration and genuine consent — the three things AI has changed not at all.
- The trigger is the innovation. Same 160 characters, fired at a moment a model chose rather than a date you picked.
- Send-time optimisation is the highest-ROI, least glamorous win — no new content required, just per-recipient scheduling.
- Conversational SMS needs webhooks, and the LLM must answer from your order system, not its own recall.
- Translation got cheap; DLT approval did not. Every language variant is a separate template to register.
- Generation speed solves the cheap problem. Knowing what to say and to whom was always the constraint.
- More messages is still worse. Cheap generation plus automated triggers produces volume, and volume produces opt-outs.
Nine use cases that earn their place
Ordered roughly by how reliably they pay off, not by how impressive they sound in a deck.
Predictive churn intervention
The classic retention SMS goes out on a schedule — “we miss you” to everyone inactive for 60 days. Most of those people were never leaving, and the ones who were had already gone.
A churn model scores customers continuously and fires the message at the point risk crosses a threshold — which might be day 12 for one customer and day 90 for another. You send fewer messages to better-chosen people.
Send-time optimisation per recipient
Batch-and-blast sends everyone the same message at 11am. A model learns that a given customer opens links at 7am and another at 10pm, and staggers delivery accordingly.
This is the single highest-ROI AI application in messaging, and the least glamorous. It requires no new content — just per-recipient scheduling against historical engagement.
Promotional SMS in India has restricted sending windows under TRAI rules. Optimisation operates within that window, never around it.
Conversational two-way SMS with an LLM behind it
Two-way SMS used to mean keyword matching — reply YES to confirm, STOP to opt out. An LLM behind the inbound webhook lets a customer reply in natural language, in Hindi or Tamil or Marathi, and get a useful answer.
These are the highest-volume, lowest-complexity support queries in most businesses, and they resolve without a human. The requirement is a provider with genuine two-way SMS through a single API and webhook delivery of inbound messages — not all of them have this.
The model answers from your order system, not from its own recall. An LLM guessing at a delivery date is worse than no reply.
Regional-language personalisation at scale
India has 22 official languages. Historically, running campaigns in eight of them meant eight copywriters, eight approval cycles and eight sets of DLT templates.
Machine translation with human review has collapsed the cost of that. A campaign can now run in a customer’s own language based on their circle or stated preference, at close to the cost of running it in one.
Fraud signals and adaptive OTP
OTP delivery is the highest-value transactional SMS most businesses send — a failed OTP is a failed login, a failed payment, an abandoned signup.
AI has changed this in two directions. Risk models decide whether a step-up authentication is needed at all, cutting unnecessary OTPs. And on the fraud side, models detect OTP-farming patterns — bursts of requests across number ranges — that used to be discovered on the invoice.
Cart and form abandonment with intent scoring
Abandonment SMS is old. Intent scoring is what makes it tolerable.
Not everyone who abandons a cart wants a message. A model separates the genuinely hesitant from the browsing and the price-checking, and messages only the first group. The result is fewer sends, higher conversion, and — critically — fewer opt-outs, which is the hidden cost of over-messaging.
The AI Is the Easy Part. The Infrastructure Isn’t.
AI-drafted campaigns with human approval
Yes, models write competent SMS copy. Within 160 characters, in a required tone, with a variant set for testing, in minutes.
This is genuinely useful and genuinely overrated. The constraint on SMS marketing has never been how fast you can write 160 characters. It’s been knowing what to say and to whom. Generation speed solves the cheap problem.
Use it for volume — variant generation for A/B testing, regional adaptations, seasonal refreshes. Don’t use it to decide strategy.
RCS with AI-selected rich content
RCS gives Android users branded sender identity, verified badges, images, carousels and suggested replies — SMS with a modern interface. Combine it with a model choosing which product cards to show a given recipient and you have something closer to a personalised storefront than a text message.
The practical approach is RCS where supported, automatic SMS fallback where not. Your provider should handle that fallback natively rather than making you build it. At around ₹0.18 per RCS message, the economics work when the content is genuinely personalised — and not when it isn’t.
Voice as the accessibility layer
For low-literacy and older audiences, and in several regional markets, an automated voice call outperforms a text — and AI-generated regional-language voice has made this dramatically cheaper to produce.
At roughly ₹0.08 per 15-second pulse, a voice drop costs less than a WhatsApp marketing message. For customer segments that don’t reliably read SMS, it’s often the only channel that works.
What AI hasn’t changed
Three things — and they’re the ones that determine whether any of the above works.
Delivery is still physics
No model improves a bad route. If your provider’s delivery rate is 88%, AI-optimised send timing on messages that don’t arrive is an expensive way to achieve nothing.
DLT is still mandatory
TRAI’s framework doesn’t have an AI exemption. Every template, however generated, requires registration. Every sender ID requires approval.
Consent is still consent
Better targeting is not a substitute for permission. Models that predict who is most likely to respond will happily point you at people who never opted in.
The consistent failure pattern with AI-driven messaging is that cheap generation plus automated triggers produces volume, and volume produces opt-outs. Every use case above is valuable because it sends fewer, better-timed messages. Used to justify more messages, AI makes your channel worse, faster.
What this needs from your provider
Most of these use cases fail on infrastructure, not on the AI. Specifically, you need:
Most of this is answerable in a single call. Ask for the webhook documentation rather than a description of it, ask whether scheduling is per-recipient or per-campaign, and ask what a delivery receipt actually contains — an operator response, or just “sent”. Providers who have built this will send you the docs. The ones who haven’t will send you a case study.
We sell the plumbing, not the magic
Nothing on this page requires a special AI product from your messaging provider. It requires a provider whose API doesn’t get in the way of the one you build or buy.
That means webhook-delivered inbound SMS, per-recipient scheduling, operator-level delivery receipts your models can actually learn from, and SMS, RCS and voice on one account with fallback handled for you. Plus DLT registration support, because the moment you start generating variants your template count goes up sharply.
And flat ₹0.13 per delivered SMS, ₹0.18 RCS, ₹0.08 per 15-second voice pulse — billed only on what actually arrives, which is the same delivery data your models depend on.
Frequently asked questions
Primarily for triggering rather than writing — predictive churn scoring, per-recipient send-time optimisation, intent scoring on abandonment, and conversational two-way replies. The largest gains come from sending fewer, better-timed messages.
Yes, and it’s useful for generating variants and regional adaptations at volume. But every generated variant still requires DLT content-template registration in India before it can be sent, so approval time, not writing time, is the real constraint.
Only if registered. TRAI’s DLT framework makes no exception for AI-generated content — each template must be registered against your principal entity and approved before sending.
Two-way SMS where inbound replies are delivered to your system via webhook and answered automatically, increasingly by an LLM grounded in your order or account data rather than by keyword matching.
No. AI changes when and why a message is sent, not the channel. SMS remains the only channel with universal reach, no app install requirement and no algorithmic filtering.
