When Your Marketing Team Has Three People, How Do You Write "One-to-One" Messages for a Million Users?

A few days ago, I was scrolling through my feed and saw a line that stuck with me.

A McKinsey survey found that 76% of consumers get frustrated when a brand "doesn't remember them."

At first, I didn't think much of it. But after mulling it over, it actually makes a lot of sense. Think about it yourself: you placed an order on an app last week, and this week the push notification you get is still "New User Exclusive." How do you feel? You probably turn it off — and tap "Don't remind me again" while you're at it.

The second half of that finding is even more sobering. The cost of that frustration isn't emotional — it's real money. Conversion rates drop, repurchase rates drop, users quietly walk away without so much as a goodbye.

So today I want to talk about one thing: how to make every user feel like "this message was written for me" — at scale.

Let's Define the Problem First: What Is "AI Marketing Personalization"?

Let me give you a definition first, then we'll build on it.

AI marketing personalization, plainly speaking, is using machine learning to decide — in real time, based on each person's current behavior — what content to show them, through which channel, and when. It's not sending the same email to a segment bucket. It's true one-to-one personalization at scale, with all thousand variations sent simultaneously.

You might ask: isn't this just personalized recommendations? I can do that with a rules engine too.

Not the same. What's the difference?

A rules engine is a fixed set of logic you write: "Three or more orders → VIP." But people don't live by rules. A user was ordering like crazy last week, then suddenly stops coming this week — and your rules engine won't catch it until next month's batch run. Rules look backward. AI looks forward.

Here's a concrete example. 8fit, a fitness app with 40 million downloads globally, had a very real pain point: they wanted to push paid memberships, but didn't want to offer the same discount to everyone. Giving a big discount to someone with high intent means leaving money on the table. Giving no discount to someone with low intent means they never buy.

They used Braze's Predictive Purchases to calculate a "purchase propensity score" for each user, from 0 to 100.

Note: this score isn't a gut feeling. It's calculated by a model based on that person's past behavior. What does it calculate? How likely they are to spend money in the coming period.

Then they did something really smart:

• High score? Less discount. They were going to buy anyway.
• Low score but the model says "just needs a nudge to convert"? That's where the big discounts go.
• Lowest score, almost zero chance of converting? Don't send anything at all.

The result? High-intent users converted at 3.75 times the rate of the random control group. And even better: they sent 100,000 fewer emails per week, with zero drop in conversions.

Think about that. The algorithm did the math for them. Sending costs for 100,000 emails, the cost of annoying users, the risk of unsubscriptions — all saved.

The Hard Part Isn't "Personalization." It's "Personalization at Scale."

Let me tell you a second story.

Dayuse, a hotel distribution platform specializing in day-use bookings, covers 30 countries. How many people are on their CRM team?

Three.

Three people, facing a global user base. English, French, Japanese; business travelers, couples, families with kids — all crammed into the same welcome email template. Users receive content that has zero relevance to them, and naturally, they don't come back to book again.

Have three people hand-write every single message? Not realistic.

They embedded Braze AI Agent Console into Canvas. Simply put: at every touchpoint, AI receives that user's favorites, order history, preferred room types, and preferred language — then generates a piece of copy on the spot that belongs to that one person alone.

Two numbers stood out to me:

• The "Favorites" campaign generated 2x the incremental revenue of the control group.
• Overall booking conversion rate increased by 90%.

The hard part was never "can you write a good personal message." The hard part is: how do you make this run automatically across a million users — without hiring a hundred people?

Segmentation: Rules Can't Keep Up Anymore

Here's another case study that blew me away.

Luxury Escapes, a travel platform with 9 million members globally, covering 30 countries. Their old approach was textbook standard: sort new users into different welcome paths based on "number of visits." Three visits → Group A. Five visits → Group B.

The problem? Their backend actually had tons of more granular behavioral signals: which room types they looked at, how long they stayed, whether they searched for dates. But none of it was usable. To change the rules, they'd have to rewrite everything from scratch — a massive engineering effort.

They flipped the approach: stop writing rules. Let AI be the one doing the segmenting.

Using Braze AI Agent Console, they simultaneously tracked 10 website behavioral signals, placing each new user into the most appropriate cohort in real time. The rules went from one variable ("visit count") to 10 variables computed simultaneously.

Here's the math:

• Revenue per user was 10% higher than the rules-based group.
• Average order value up 7%.
• Order volume up 6%.

That 10% revenue lift came entirely from conversion rate improvement. Not a penny more in ad spend, not one extra message sent — just more accurate segmentation.

Let me tell you, this is the fundamental difference between a rules engine and AI segmentation. Rules "shove people into fixed boxes." AI "makes the boxes follow the people." The user changes, the boxes change instantly.

The Real Pros Even Let Machines Run the Testing

You've probably heard of A/B testing. I prepare two versions of a headline, send each to 50% of users, see which one wins.

But traditional A/B testing has a flaw: it waits for humans to wrap things up. You wait a week for data to stabilize, hold a meeting to decide which version to use, then push it to full scale. During that week, your users are still seeing the losing version.

Cleo did something more aggressive. Cleo is an employer-provided family care benefits platform, covering members from trying to conceive all the way to caring for aging parents. Their welcome email series was performing decently, but the content was written to serve all audiences — so it was "safe." It touched on a bit of everything, and resonated with no one.

Their lifecycle manager Holly used Braze AI Operator to have AI write personalized versions of Liquid code, automatically accounting for edge cases the team hadn't considered. For example: the same member might have indicated "expecting to be a dad" or "caring for an elderly mother" at registration — and the entire welcome series should flow completely differently based on that.

Here's what happened:

• App open rate up 284%.
• Push notification engagement up 124%.
• Unsubscribe rate down 81%, and first-email unsubscribes down 97%.

I went back and re-read those numbers. A 97% drop in unsubscribes means the old welcome emails were basically being treated as spam. Swap in personalized versions, and nearly all of them were rescued.

This is what I mean by "let testing run itself." AI doesn't just run experiments for you — it pushes the winner live at the moment of every send, so users never even feel the "losing version" existed.

So How Do You Start? Let Me Break the Path Down for You.

Four case studies in, you might be wondering: I want to do this too — where do I begin?

I'll break it into five steps. Each one is the foundation for the next, so don't skip ahead.

1. Clean and unify your data first.

Behavioral data, purchase records, user preferences, real-time events — all flow into one pool. If your data is scattered across three or four systems with batch-processing delays, everything built on top of it carries errors. This is the foundation of foundations.

2. Let the model start predicting.

Once the data flows, machine learning models can calculate: who's likely to buy, who's likely to churn, whose lifetime value is highest. Turn those predictions into dynamic audience segments that update in real time with user behavior.

3. Connect predictions to actual outreach.

Lifecycle marketing, triggered messages, cross-channel sequences — all run based on real-time behavior. AI decides content, channel, and timing. You decide strategy.

4. Don't just set it and forget it.

Automated A/B testing makes every send smarter than the last. Engagement analysis tells you where users are responding and where they're dropping off. These signals feed back into the model, so the next round of decisions is always sharper than the last.

5. Let the team think big.

This is my favorite step. Once automation is running, your team no longer manually dispatches every message. Three people can manage personalized outreach to a million users. The rest of the team goes off to think about strategy, creativity — the things machines can't do.

How Do You Know If You're Doing It Right?

Here are the metrics I'd always keep an eye on.

Start with engagement data. Open rates, click-through rates, in-app engagement. These are your early warning system — before business results materialize, these numbers move first. When they trend in the right direction, conversions typically follow.

Then look at conversion and revenue. Conversion rate and revenue per message — compare against your control group or the baseline before personalization, and track over time.

Then retention and lifetime value. Repurchase rate, churn rate, CLTV — this is the long-term view. If your personalization only drove a one-off transaction and retention didn't move, you only did half the job.

Finally, the testing feedback itself. Which version is winning, which channel is growing, how fast the model is learning. These signals feed back into the model, making your marketing performance compound over time.

A Genuine Reflection

After writing up these four case studies, I have a very specific feeling.

When we used to say "personalization," we really meant "group-ization." Split people into a hundred groups, send each group one script — that was the ceiling. No matter how big the team or how detailed the rules, you couldn't keep up with the speed at which user behavior changed.

What AI does is take "grouping" out of human hands. It doesn't need you to write a thousand rules. It needs one model and clean data to make millisecond-level decisions for every individual.

What excites me most isn't any single number — it's that these four teams are in different industries, at different scales, with different team sizes. A travel platform, a fitness app, a hotel distributor, a family care service. But the same methodology worked for all of them.

That tells me it's not a proprietary trick belonging to any one company. It's a repeatable capability.

Whoever truly understands this shift from "groups" to "individuals" can leverage marketing to multiples they never dared imagine — without adding headcount.

As for which specific tools to use, how to integrate them, how to orchestrate them — those are engineering problems. Engineering problems always have solutions.

The hard part is mindset. It's whether you're willing to admit: those rules you've been writing for three years, that operations manual you're so proud of — in front of a new generation of users, they can't keep up anymore.

I don't have a standard answer either. But this question is worth thinking about seriously.