Email — the default floor.
Everyone has it; nobody resents it. It carries the considered message — the one that can wait an hour and still land. The baseline every other channel has to beat.
AI in CRM communications
I shipped LLM-driven personalization before it was a category.
Most “AI personalization” is the same mass send with a first-name token glued on top. I mean something narrower and harder: a different message for a different person — because they are a different person. Done with judgment, that is the line between noise and revenue.
I wired large language models into a live CRM in 2022, while the market was still arguing whether ChatGPT was a toy. I have done it for an edtech platform with millions of users and across a marketplace running in 17 countries — and, just as importantly, I know how to keep it from going wrong.
The thesis
Hyper-personalization is craft, not spray-and-pray.
A biology student in Massachusetts whose first language is Italian, signing up on a Saturday morning, is not the same person as a law student in Berlin signing up late on a Tuesday. Treat them the same and you waste both. Micro-segmentation isn’t a nice-to-have — it’s the path to real optimization.
So the goal was never more messages. It’s a better system: less noise, more meaning, measurable impact on revenue. AI is what finally makes that affordable at scale — an editor who can actually reason now sits behind every single send, not just the templates.
The hard part is not the model. It’s the judgment around it — what to say, to whom, when, and whether to send at all.
Match the channel to the person
Personalization isn’t only what you say. It’s where the person already is.
A perfect message in the wrong channel is still a miss. Before the model writes a word, the system decides where this person is reachable — and where they actually pay attention. The rule is boring and it works: meet people where they already are, not where it’s convenient to broadcast.
In-product — for the people already inside.
Power users live in the app, not the inbox. A nudge at the moment of use beats an email they’ll read tonight, if at all. You catch intent while it’s warm.
In-store / operator screens — for the people running the business.
For an operator mid-shift, the relevant surface is the tablet on the counter, not a marketing email. The message has to fit the moment they’re actually working in.
Messaging — mobile-first, and only when they’ve opted in.
Where people read in seconds, WhatsApp and similar channels beat email by a wide margin when the audience is eligible and opted in. High signal, low tolerance for noise — so it’s reserved for messages that earn it.
Same person, different surface depending on the moment — that’s the layer most “omnichannel” pitches skip. They blast every channel at once and call it coverage. The point isn’t to be everywhere. It’s to be in the one place this person will actually notice.
Before the trend · 2022
I put an LLM inside the CRM before ChatGPT was on most roadmaps.
At an edtech platform serving millions of students, I designed and launched an LLM-powered content engine — OpenAI models wired directly into the lifecycle layer. It generated hyper-personalized study recommendations and CRM messages at scale no human team could hand-write. This was 2022: building on OpenAI before ChatGPT became the default slide in every roadmap.
I also stood up the CRM stack underneath it — Customer.io and Segment — so the personalization had clean data and clean triggers to run on. The engine wasn’t a demo. It ran.
Then at scale
The same discipline, across 17 countries.
Not a startup-only trick. At Uber Eats I owned the EMEA CRM layer — 250,000 merchants across 17 countries, ~2M comms a month — and ran the same loop: rigorous A/B testing, AMPscript fixes, AI-assisted personalization, channel orchestration. The numbers that moved are the ones a CFO cares about.
+49%
CTR 1.53% → 2.28%
+20%
open rate 43.9% → 52.7%
95%+
deliverability from ~84% after a platform fix
>80×
WhatsApp CRM beta ROI vs email
~2,000
automations governed
The full record is on experience, and the measurement discipline behind each number is on how I measure.
Behaviorally triggered, not blasted
The best message is the one a person’s own behavior asks for.
A calendar blast treats everyone as if they did the same thing this week. They didn’t. The systems I build read behavior and fire when it matters — so the message arrives because the person just did something, not because it’s Tuesday. A few patterns I reuse, adapted to each business:
Dropped off mid-action → bring them back.
Someone starts and doesn’t finish — a cart, a setup, an application. A timely, specific nudge recovers a meaningful share of that lost intent. Not a guilt trip; a hand back to where they were.
New and unsure → teach, then convert.
A first-week user doesn’t need a discount; they need to reach the first moment the product is obviously worth it. Education sequenced to adoption beats a coupon every time.
Doing well → raise the ceiling.
When the numbers say someone is ready — usage up, results showing — that’s the moment to offer the next tier, not a random promo week. Performance triggers the upsell, so it lands as help, not a pitch.
Real-world events → automate the obvious.
Holiday hours, a status change, a renewal date — predictable moments that should never need a human to remember them. The system handles the obvious so the team can spend its judgment on what isn’t.
Thin data? Serve the fallback — don’t guess.
Personalization is only as honest as the data under it. If the signal is thin or stale, the system does not invent a guess dressed up as insight — it serves a clean, generic version and waits. A confident wrong message costs more than a plain, correct one. The fallback is a feature, not a failure.
Frequency capping is the quiet half of the job.
That every journey can fire does not mean every journey should fire. One contact-governance layer sits above all of them — caps per person, per week, priority when two triggers collide, and a hard stop so nobody gets three messages in a day. Orchestration over frequency: the system protects the person’s attention as carefully as it spends it.
How I keep it honest
The reason teams don’t ship AI personalization isn’t the model. It’s the fear.
Fear of a tone-deaf message. A privacy leak. A hallucinated promise sent to a real person. So I build the guardrails before the cleverness. I’ve done exactly that for DOC.UA, a Ukrainian healthcare service — where the messages are appointment reminders, not ad copy, and the stakes are real. Four rules I bake into the system itself:
No diagnosis.
The message never infers a health state from a doctor’s specialty. “You’re booked with an oncologist” — never “Don’t worry, the oncologist will help.”
No Big Brother.
Even with exact age and home address on hand, the message doesn’t show them. Data is for context and navigation, not for proving how much we know about you.
Restraint where it counts.
Sensitive specialties — psychiatry, oncology and the like — get a deliberately neutral, administrative tone. No festive emoji, no forced cheer.
Address people correctly.
Children → speak to the parents. Teenagers → first name only. Adults → full, formal address. The system applies the rule every single time.
Live run · names, clinics and contacts removed
Child (age 4) → the message speaks to the parents
Dear parents 🧑🧑🧒 We confirm your child’s appointment. 🗓 16 Aug · 🕒 13:00 · 🏥 clinic, Kyiv · 👨⚕️ specialist on staff.
Adult → full, formal address
Welcome, Olena Ivanivna! 👩⚕️ Your appointment is confirmed. 🗓 16 Aug · 🕒 11:30 · 👩⚕️ cardiologist · 🏥 clinic, Odesa.
And the output is structured — forced JSON, validated by the system — so personalization stays governed and auditable, not improvised one risky send at a time. That’s the part most “AI for marketing” pitches skip. It’s the part that lets you actually turn it on.
How I think about it
A few rules I don’t break.
Talk to your users. The personal stories of a dozen real people often beat the aggregated data of thousands — they tell you what to even measure.
Never give the model one goal. Give it a primary goal and a few intermediate ones, or it will cheerfully optimize the wrong thing.
The win is never volume. It’s a system that sends less and means more — the same instinct behind how I think about CRM.
The questions everyone asks
AI in CRM, answered straight.
Does AI in CRM actually work, or is it hype?
It works when it’s pointed at the right job. AI doesn’t make a bad strategy good — it makes a good system affordable at scale, by putting an editor that can reason behind every send instead of just the templates. Used well, it lifts the numbers a CFO cares about: click-through, retention, revenue per message. Used as a buzzword, it just produces more noise, faster. The difference is judgment, not the model.
How do you do AI personalization without being creepy?
You use data for context and navigation, never to show off how much you know. The rule I bake into the system: even with someone’s exact age and address on hand, the message doesn’t parade them back. No inferring a health state from a doctor’s specialty, no “we noticed you were near…”. Personalization should feel like good service from someone who pays attention — not surveillance. The moment it makes a person flinch, it has already cost you more than it earned.
What about privacy and governance — how do you keep AI from going wrong?
Guardrails first, cleverness second. The output is structured — forced JSON the system validates — so every message is governed and auditable, not improvised one risky send at a time. There’s a fallback for thin data, a frequency cap so nobody gets buried, and deliberately neutral handling for sensitive contexts. I’ve been doing this since 2022, including for a healthcare service where a tone-deaf message isn’t a metric — it’s a real person. That discipline is the part most “AI for marketing” pitches skip, and it’s the part that lets you actually turn it on.
How should a team start with AI in CRM?
Not with the model — with the data and one clear job. Pick a single high-value moment where a better message obviously helps (an abandoned action, a shaky onboarding) and make sure the data under it is clean and the trigger is real. Set a primary goal and a couple of guardrail metrics so the system can’t optimize the wrong thing. Ship that one loop, measure it honestly against a holdout, then expand. Teams that start with “let’s add AI everywhere” end up with noise; teams that start with one governed loop end up with a system.
AI vs a rules engine — when do you use which?
Use rules when the logic is known and stable; use AI when the variation is too large to hand-write. A renewal reminder or a holiday-hours notice is a rules job — predictable, and you want it predictable. Writing a genuinely different message for thousands of distinct people, in the right register for each, is where a language model earns its place, because no team can hand-author that volume. Most good systems are both: rules decide what fires and when; AI decides how it’s said — inside guardrails the rules enforce.
Do you need large volume or a big budget for this to pay off?
No — you need CRM that’s already tied to revenue. The approach works the same at startup scale and at millions of users; what changes is the size of the prize, not the method. If your messages already influence what people buy or whether they stay, a smarter system pays for itself quickly. If CRM is just a newsletter you send out of habit, AI won’t fix that — restraint will. Start where a better message has a clear path to money.
If you want AI in your CRM done with judgment, not hype — let’s talk.
Think this maps to your situation?