Case study · Klaviyo

Subject Line Assist

Klaviyo's AI could write brand-aligned subject lines automatically. Only 1.74% of users were using it. I redesigned the interaction model from scratch and drove an 8× increase in adoption.

Company

Klaviyo

Focus

Interaction Design · AI UX · Experimentation

Impact

1.5% → 24% account-level adoption

Background

Great AI. Nobody using it.

Klaviyo's Subject Line Assistant used your brand's historical emails, tone, and segment data to generate personalized subject lines automatically. The AI was good. But across every customer tier — from scrappy startups to enterprise accounts — barely anyone was using it.

Average adoption: 1.74% of messages. Flat. Across all segments.

LENT

2.8%

Large enterprise

UENT

2.0%

Upper enterprise

LSMB

1.6%

Large SMB

USMB

1.3%

Upper SMB

1.0%

Mid-market

User tickets told the same story. About 5 tickets and 10 chats every month — all saying the same things:

Too many clicks

The SLA required clicking an icon, filling in context fields, and clicking generate — multiple screen changes just to see a suggestion.

Disconnected from writing

Users had to leave the subject line field entirely to access the assistant, breaking their flow at the exact moment they needed help.

Suggestions felt generic

Output didn't reflect brand voice or campaign context. Users didn't trust it, so they didn't use it again.

Hard to learn

"Takes time to learn to generate subject lines that are brand relevant" — users gave up before they ever got good results.

Design thinking

The interaction model was the problem.

My goal: reduce clicks to 3 or fewer, integrate brand services, and get adoption to ~5% of messages. But before picking a direction, I explored three fundamentally different theories of how AI should enter the writing experience.

Keyword-based prompting

Users guide AI with keywords — adds intent, but still an extra step. Creates uncertainty around what the AI will do.

Toolbar-based controls

AI triggered via a sparkle icon — more efficient, but still reactive. Users have to remember to activate it every time.

Performance-informed

Surfaces past subject lines with open rates — builds trust but raises data clarity concerns and echo-chamber risk.

Key insight

Every concept still required users to do something extra. The pattern became clear: any friction — a click, a modal, a context field — was enough to make users skip it. The guiding principle: guidance without friction.

Experimentation

Three experiments. One clear winner.

I designed and shipped an A/B/C experiment across ⅓ of all Klaviyo users — free and paid — to isolate the impact of proactive vs. reactive AI and inline vs. modal interactions. Each user account was assigned to one variation so experiences stayed consistent.

Control · A

Modal

Existing multi-step experience — click icon, fill context, click generate

Variant · B

Reactive

Sparkle icon triggers inline suggestions with a regenerate option

Variant · C · chosen

Proactive

Suggestions surface on focus, refresh every 3 words with a 2-second pause

Control — existing modal-based Subject Line Assistant

Variant B — reactive AI via sparkle icon

Variant C — proactive inline AI surfacing suggestions as users type

Experiment takeaway

Proactive inline AI drove ~6× higher message-level adoption and ~8× higher account-level adoption vs. the modal baseline. Interaction design — not the AI model — was the constraint the whole time.

Final design

Type. See. Choose.

Variant C became the shipped experience. Proactive adaptive autocomplete — embedded directly in the subject line field, pulling from the user's last 5 emails, brand tone service, and segment attributes (lifecycle stage, product preferences). No modal. No extra clicks. The AI shows up where you already are.

Key design decisions

Discoverability

Suggestions appear when the field is focused — no activation step, no icon to hunt for. Zero additional clicks to see the AI.

Timing

Suggestions refresh every 3 words with a 2-second pause — responsive to context without interrupting the writing rhythm.

Control

A visible loading state lets users cancel a refresh and keep the current suggestions. AI supports — never blocks.

Exploration

Regenerate gives users a new set of suggestions without losing their place — iteration without commitment.

Brand relevance

Suggestions pull from historical subject lines, the brand tone service, and segment attributes — so output feels like yours, not generic AI copy. When no historical data exists, the system falls back to segment context (e.g. "Holiday Specials for Our Loyal Shoppers!") rather than going generic.

Impact

Results

1.5% → 8–9%

Message-level adoption (~6× increase)

3% → 24%

Account-level adoption (~8× increase)

Reflection

The AI was already good enough. What changed was where and when it showed up. Moving from a modal users had to find — to suggestions that appeared in context, without a single extra click — was the entire unlock. The PRD originally targeted 3× adoption (~5%). We hit 8×. Sometimes the best design decision is removing the decision entirely.

↑ Back to top