Using LLMs to generate personalized messaging journeys 

As marketers, we’ve all officially (hopefully) evolved from static drip campaigns to dynamic, behavior-driven journeys that adapt in real time. But personalization is more than just reacting to clicks and triggers. It’s about making every message feel like it was written for the person reading it.

Now, large language models (LLMs) are making that vision possible. By blending data, automation, and generative AI, we can design onboarding experiences that resonate with customers in the way they prefer to communicate.

Recently, our team ran an experiment to explore how LLMs could tailor Customer.io’s onboarding emails to match each trialer’s communication style, and what we learned might change how you think about AI personalization altogether.

Picture this: a new user signs up for a Customer.io trial. They’re eager to explore, ready to see what’s possible, and immediately receive our standard onboarding series. Historically, that sequence has been effective, triggered by behavior, timed to engagement, and optimized over dozens of iterations. However, while it responds to people's actions, it doesn’t adapt to how they prefer to be addressed.

Some people love direct, data-rich messaging. Others respond to stories and inspiration. Some appreciate a warm, conversational tone. So, what if we could automatically adapt the same email content to match those styles?

That’s the experiment we set out to run: an AI-assisted onboarding experience that uses LLMs to make every message feel more personal, relevant, and human.

Our goal was straightforward: to increase onboarding conversion among self-serve trialers, targeting approximately 440 marketing-qualified leads (MQLs) per month. Specifically, we aimed to increase “a-ha moment” completion by 9% within 14 days and achieve a 65% customer activation rate within 45 days of account creation.

Our team focused on self-serve trialers. While our existing onboarding already personalizes messages based on product behavior, we hadn’t yet personalized based on firmographic or psychographic insights. We suspected that trialers were dropping off not because the message wasn’t helpful, but because it wasn’t resonating with how they think, work, or communicate.

If we tailor the communication style of our onboarding emails to different types of people, engagement and conversion will rise. The primary metric was onboarding conversion rate, with secondary measures of email engagement (opens and clicks) and unsubscribe rates as counter-metrics.

Our core value proposition remained consistent: Customer.io helps you send data-driven, behavior-based messages that your customers actually want. What changed was how that message was expressed. Each version adapted the tone, structure, and style to match the trialer’s communication preferences.

We used a two-agent AI workflow:

1. Agent 1: The Enricher inferred communication styles from publicly available data, including job title, company size, and industry.
2. Agent 2: The Expert Copywriter rewrote the original onboarding email in the appropriate tone, adjusting phrasing, pacing, and vocabulary.

The system assigned each user one of six communication style categories:

• Data-Driven Concise — prefers facts, brevity, and measurable ROI.
• Visionary Inspirational — responds to bold ideas and big-picture thinking.
• Formal Traditional — values professionalism and structure.
• Collaborative Relational — appreciates teamwork and inclusive language.
• Friendly Informal — enjoys a personable, casual approach.
• Technical Detailed — seeks depth and specificity.

Our team then split the audience 50/50 into two test groups:

• Version A: text-only customization by style.
• Version B: text plus light design variation (layout and visual tone).

Email was chosen as the test channel since it remains the core driver of trial activation.

The team developed a pair of connected AI agents that worked like an assembly line for personalization. The first identified a person’s likely communication style based on enrichment data, while the second adapted the messaging to match. Each component required careful tuning: the enrichment model had to be precise enough to avoid mismatching users, and the stylist model needed just the right prompt engineering to produce on-brand, accurate, and human-sounding copy.

To put it lightly, this was not a “set it and forget it” process. Matching people to public profiles proved challenging, and prompt engineering turned out to be its own art form. Our team spent nearly four weeks refining the prompts to make sure the AI didn’t stray too far from our tone and to prevent it from rewriting content that didn’t need changing. The end result was a workflow that produced concise, well-targeted message variants ready for human review and deployment.

Here’s a simplified version of how the experiment worked in action.

• Data-Driven Concise:
  Subject: Build your first campaign in under 5 minutes
  Body: You’re three steps away from measurable ROI. Start your first campaign now and track results instantly.
• Visionary Inspirational:
  Subject: Shape the future of your customer communication
  Body: You’re about to unlock a more meaningful way to connect. Let’s build your first campaign and start inspiring action.
• Collaborative Relational:
  Subject: Let’s build something great together
  Body: We’ll guide you step by step. Together, we’ll create your first campaign and set you up for long-term success.

Each variation keeps the core idea intact but feels completely different to the reader. One speaks to efficiency, another to ambition, another to connection, all automatically generated and QA-checked before deployment.

After several weeks of iteration, testing, and validation, we’ve gathered a set of insights that go well beyond the experiment itself. Style inference, we discovered, isn’t an exact science; even small gaps in data can lead to misclassification, where someone who’s deeply analytical might be mistaken for an inspirational communicator.

The subtlety of human tone makes this especially tricky. Prompt design also proved to be more art than science. Tiny changes in phrasing, even swapping one adjective for another, could shift the entire voice of the message, turning a confident tone into something cold or overly casual. Human review remained indispensable throughout; every piece of AI-generated copy went through an editor to ensure accuracy, brand consistency, and emotional balance.

Most importantly, we learned that the novelty of AI-crafted personalization must still feel natural. Audiences can spot when text feels synthetic, so our goal became invisible enhancement, not showcasing that AI was involved. Ultimately, we recognized that scaling personalization successfully requires more than just words. It depends on robust design systems, segmentation logic, and operational readiness to support multiple journeys running in parallel.

Personalization isn’t just about saying the right thing; it’s about saying it in the right way. This experiment demonstrated that communication style is an underutilized dimension in lifecycle marketing. When tone aligns with personality, engagement naturally rises.

It also reminded us that AI can’t replace marketers; it enhances them. With LLMs embedded in your workflow, you can scale creative testing faster, uncover insights about what resonates with your audience, and evolve your messaging strategy in real time. The marketer’s job shifts from writing one perfect email to orchestrating a system that writes multiple good emails, each tuned to the reader’s perspective.

If you’re ready to experiment with AI-driven personalization, here are three of the same prompt patterns we tested, ready for you to use with your own campaigns.

Prompt 1: Style Matching Prompt

Rewrite the following email in a [communication_style] tone. Use the existing structure but adjust the language and rhythm to fit the tone. Email: [Paste your base message here]

Prompt 2: Persona-Driven Rewrite

You are a lifecycle marketer writing to a [job title] at a [company size] company in the [industry] industry. Rewrite the email below to align with their likely goals and communication style. Base Email: [Paste your content here]

Prompt 3: Subject Line Variations by Style

Generate 3 variations of this subject line that align with: - Data-Driven Concise tone - Visionary Inspirational tone - Friendly Informal tone

These prompts are simple to run but powerful when combined with segmentation and testing inside Customer.io. Start small, test one message in three tones, and measure engagement lift before scaling further.

Our AI-personalized onboarding is still running, but we already see new possibilities ahead, and we're not the only ones. Our friends at Notion also noticed the power of personalized onboarding to improve customer experiences.

Soon, our team plans to extend style-based personalization beyond email to include in-app messages, push notifications, and even chat experiences. We’ll also test how layout and design can align with communication style, such as minimalist visuals for data-driven users or vibrant imagery for visionary audiences.

We’re also investing in improving the accuracy of our enrichment models, expanding the percentage of users who can be matched confidently to a communication style. And perhaps the most exciting frontier is shifting from AI-generated copy to AI-assisted decision-making, letting models assign segment values, flag patterns, or suggest optimizations based on how users progress through their journey.

If you’re a marketer curious about using AI in your onboarding or lifecycle flows, start small. Test a few AI prompts. Measure engagement. Learn fast. You’ll be surprised at how quickly “personalization” stops feeling like a buzzword and starts feeling like a genuine conversation.

Start creating unique, personalized customer communications today, with a free trial of customer.io.