Subject-line generation lifted open rates 24% across 1.4M monthly sends.

Context

A direct-response e-commerce operation running large daily email sends across multiple brand properties in the UK and EU. Two in-house copywriters were the bottleneck on subject lines — every send needed 4–6 A/B variants, every variant needed to match a specific brand’s voice (the brands range from premium understated to playful direct), and every variant needed GDPR-compliant claims wording. The team was producing 80–110 subject lines per week and burning out.

The problem

Open rates had plateaued at 18–21% across the portfolio. The copywriters were generating most variants in the same lexical patterns simply because of fatigue. Manual brand-voice compliance was inconsistent; legal had flagged three claims in the last quarter that had slipped through. Hiring another copywriter was a four-month timeline and the team didn’t want to grow headcount.

What we shipped

A subject-line generator built per-brand, each one prompted with:

• The brand’s voice rules — curated by the copywriters themselves, not the AI guessing
• A corpus of historical subject lines — both high-performing and low-performing, so the model learns what works and what fails
• The campaign brief — product, audience, send context
• A compliance ruleset — banned phrases, claim restrictions, GDPR-required formats

Output: 12 variants per request with a rationale for each. Copywriters select, edit, and approve — nothing sends without human sign-off. A separate compliance pre-check runs on every variant before it’s surfaced to the copywriter, catching banned-claim phrasing before a human ever sees it.

Hard cost ceiling of $180/month total AI spend, set in software, never exceeded.

Results

The 24% open-rate lift is the weighted average across all brand properties, measured against a pre-deployment baseline over a matched six-month window. Zero compliance violations reached send in the first nine months of production — against three that had reached send in the equivalent prior period.

What we’d do differently

We initially under-invested in the per-brand voice rules. The first two brands hit lift targets fast because their voices were distinct. The third brand’s voice was subtler, and the AI produced bland variants for two weeks until we sat with the copywriter for an afternoon and rewrote the voice rules. Lesson: spend a full day per brand on voice setup, not a couple of hours.