AI Subject Line Optimizer

The Problem With Writing Subject Lines

Every marketer has been there. You've built the campaign, written the email, gotten approvals, and now you're staring at the subject line field with the low-grade dread of someone who knows this 60-character window will determine whether any of the rest of it matters. Subject line writing is one of the most persistent friction points in email marketing. It's high stakes, surprisingly hard to get right, and almost entirely based on gut feel. Unless you happen to be the kind of person who runs A/B tests on everything, tracks open rates obsessively, and has time to synthesize patterns across months of sends. Most marketers are not that person. Act-On had a meaningful asset sitting in its platform that was going largely unused for this problem: years of account-specific email performance data. The top performers, the duds, the patterns in between. What if we put that data to work?

How might we help marketers write better subject lines using their own performance history?

Act-On's First Generative AI Feature

This project was strategically significant beyond the feature itself. The AI Subject Line Optimizer was Act-On's first customer-facing generative AI feature, which meant the design decisions made here would establish patterns, set expectations, and build (or erode) trust in AI across the entire platform going forward. No pressure. The feature I designed does two things, but both of them only work because of one thing: it's using your data. Not industry benchmarks or generic best practices. It's looking at your top and bottom performing campaigns from the past year, analyzing what made the difference, and using that to evaluate your current subject line and generate alternatives that look more like your winners. Your data, working for you. Specifically, it analyzes True Open rate data from the top 5 and bottom 5 performing sends over the prior year and uses that intelligence to do two things:

Evaluate the marketer's current subject line against their own performance history
Generate personalized alternatives grounded in the patterns of their top performers

The Insight That Shaped Everything: Curiosity Over Promise

Early stakeholder conversations pushed toward outcome-based messaging. "Improve your open rates." "Write subject lines that perform." The kind of language that sounds compelling in a product brief and lands flat with the people who've seen a hundred tools make that same promise. Through iterative copy workshops, we landed on a different framing, one that turned out to be both more honest and more compelling. The psychological lever wasn't improvement. It was curiosity.

"See how your subject line compares to your past performance."

That single reframe changed the entire product voice. Instead of promising an outcome we couldn't guarantee, we were inviting a comparison the marketer was already curious about. It shifted the feature from "AI tells you what to write" to "AI shows you something you didn't know about your own data." That's a fundamentally different relationship, one that earns trust instead of asking for it upfront. That single line shaped everything downstream: the entry point callout, the loading experience, the modal copy, the suggestion rationales.

The Admin Opt-In Experience

The feature launched behind Act-On's Labs module, the platform's mechanism for experimental and beta feature opt-ins at the account level. I designed two distinct states that needed to coexist gracefully within the email composer. For admins who hadn't opted in: a callout with a direct "Opt in" CTA, explaining the value in plain terms. For end users whose admin hadn't enabled it: the same callout, but with a nudge to ask their admin. Not a dead end, but an invite. Both states use the same curiosity-driven headline: "How does your subject line stack up?" The copy lands the value without overselling it.

Admin (top) gets an Opt In button. Non-admin (bottom) gets a prompt to loop in their admin. Same message, different action.

The Entry Point in Context

Once enabled, the feature lives directly in the Subject section of Act-On's Review Email page, the pre-send checklist marketers work through before hitting send. The callout surfaces right alongside the subject line validation, where the question of "is this subject line any good?" is most alive. The placement was deliberate. The marketer has written the email, reviewed it, and is about to commit. That's the moment of highest intent and highest anxiety, and exactly when "let's see how this compares to your past performance" lands hardest.

The callout lives in the Subject section of the Review Email page, surfacing at the exact moment a marketer is deciding whether their subject line is ready to send.

Building It Together: The AI Conclave

One of the most important things about this project wasn't a design decision. It was a bet. Could we ship a fully realized AI feature in four weeks, from conception to users, if we threw out the handoff model entirely and put every discipline in the room from day one? That's what the AI Conclave was. One engineer, one product manager, one designer, one marketing stakeholder. All in the same sessions, ironing out details together in real time rather than passing decisions down a chain. No handoffs. No late-stage surprises. Just the right people in the room when the calls needed to be made. We shipped in four weeks.

Finding the Right Interaction Pattern

One of the earliest and most consequential design decisions was where and how the AI interaction would live within the existing email composer. I explored four main approaches:

Sidebar assistant: a persistent AI helper beside the composer
Inline suggestions: a dropdown that appeared in the subject line field itself
Modal generator: a dedicated interface launched from the composer
Smart completion: autocomplete-style progressive suggestions

The email composer is a complex, established workflow. Anything that competed visually with existing primary actions, or that introduced friction for users who didn't want AI assistance, would create more problems than it solved. The modal pattern won for three reasons: it preserved the existing page context entirely, it created a clear intentional entry point (you had to choose to open it), and it gave the AI interaction a dedicated space without creating visual noise in the composer itself. You don't accidentally end up in the AI experience. You go there on purpose.

Designing for Trust Without Designing for Clutter

AI suggestions without reasoning are black-box outputs. Users can accept or ignore them, but they can't evaluate them, which means they can't build the judgment about when the AI is useful and when it isn't. That's not a great foundation for trust. I pushed for transparent AI reasoning: each generated suggestion would include a brief rationale explaining why it was recommended. (Yes, I made them justify themselves. Briefly.) Not a wall of explanation. Just enough to connect the suggestion back to the data it came from. The challenge was keeping that transparency from creating visual overhead. The solution was progressive disclosure within the modal: the primary view shows the suggestion and a short rationale. More detail is available if you want it, but it doesn't front-load complexity for the user who just wants to see the options.

Left: the modal with no selection. Right: a suggestion selected, with the Replace Subject Line CTA activated. Each suggestion includes a one-line rationale tied to the account's own performance patterns.

The Loading Experience

AI takes a moment to think. That moment is a design surface. Rather than a generic spinner, I designed a five-step progressive loading sequence that narrates what the AI is actually doing: fetching top and bottom performers, analyzing subject line characteristics, determining best practices from the account's own data, comparing the current subject line, and summarizing findings. This wasn't just filling time. It was building understanding. By the time the suggestions appeared, the marketer had a mental model of how they were generated, which made them more likely to trust and engage with the results.

Each loading step narrates the analysis in plain language. The animated gradient background adds motion that signals active processing without feeling frantic.

The Data Threshold Guardrail

Not every Act-On account has enough email history for this feature to produce reliable suggestions. Generating low-confidence outputs and presenting them with the same confidence as well-grounded ones would be worse than not generating anything. It would train users not to trust the feature. I established a minimum data threshold as a design guardrail: the feature requires at least 10 qualifying sends (with 100+ recipients each) from the prior year before it will generate suggestions. Accounts that don't meet the threshold see a clear, honest explanation of why, not a broken experience. This was a deliberate tradeoff: fewer users would see the feature at launch, but the ones who did would see it working well.

Results

Within the first week of launch, marketers had selected and used over 200 AI-generated subject lines in real campaigns. For a Labs feature requiring admin opt-in, with a minimum data threshold limiting eligible accounts, that's a meaningful signal that the people who could use it wanted to.

What Shipped

At launch, the AI Subject Line Optimizer let marketers:

Access AI-powered subject line analysis directly from the email composer
See how their current subject line compared to their account's historical performance
Review AI-generated alternatives, each with transparent reasoning tied to their top-performing sends
Opt in (admins) or discover the feature (end users) through a purpose-built callout experience

The feature was the deliverable. The process was the asset. The AI conclave is now a repeatable format. The curiosity-driven, transparent-reasoning approach has become a reference point for subsequent AI work at Act-On. The patterns established here didn't stop at launch.

What I Learned

Curiosity beats promises. For AI tools especially, users have heard too many outcome guarantees. Framing that invites comparison and discovery, rather than claiming improvement, builds more trust and creates more genuine engagement. Transparency is a feature, not a footnote. Showing users why the AI made a suggestion isn't a nice-to-have. It's the thing that turns a black-box output into a tool users can evaluate and learn from. Guardrails are good design. Choosing not to show the feature when the data isn't strong enough is a UX decision, not just a technical one. The boundary conditions of an AI feature are part of the experience. Cross-functional design in real time changes outcomes. The AI conclave worked because design and engineering were in the same room when the decisions that mattered most were getting made. It's not a process overhead. It's just how good work happens.