AI Dynamic Email Personalization: Campaign Results Across Four Documented Cases

AI dynamic email personalization sits somewhere between genuinely mature and still-being-figured-out. The infrastructure exists — most major ESPs now expose some form of predictive send-time optimization, behavioral content blocks, or generative subject line testing. But the documented campaign results are scattered, vendor-framed, or light on methodology. This record pulls together cases where the outcomes were specific enough to be useful and the source was traceable.

"AI dynamic email personalization" covers several distinct methods that often get lumped together. For the purposes of this record, the cases below are organized by the specific AI capability applied, not by the generic label.

What Counts as AI Dynamic Personalization in Email

Not every personalization tactic in email involves AI. Merge-field substitution (inserting a first name) is not AI. Rule-based segmentation ("send this version to subscribers who clicked last month") is not AI. The cases in this record involve one or more of the following:

• Predictive content selection — an ML model choosing which product block, offer, or content module to render per recipient at send time or open time
• Generative subject line or preview text — a language model producing or testing multiple subject line variants, with selection based on predicted open probability per segment
• Send-time optimization — a model predicting the individual send time that maximizes open likelihood per subscriber
• Dynamic offer personalization — real-time rendering of discount level, product recommendation, or content category based on behavioral signals at open time

Case Records

Case 1 — Retail: Predictive Product Recommendation Blocks

Marks & Spencer (UK retail) ran a campaign using Movable Ink's Da Vinci personalization engine to render individual product recommendation blocks at email open time. Rather than pre-selecting a product grid at send, the content was assembled per-recipient using browsing history, purchase history, and real-time inventory signals.

Case 2 — SaaS: Generative Subject Line Testing at Scale

Phrasee (now Jacquard) documented a campaign with a major UK financial services brand — not named in the public record — where the system generated and tested subject line variants using its language model, with variant selection weighted by predicted open probability per audience segment rather than a simple A/B split.

The reported result was a 2.6% absolute improvement in open rate over the brand's existing subject line process, measured across a 90-day window with approximately 4 million sends. The comparison baseline was the brand's internal copywriting team, not a static template.

Case 3 — Ecommerce: Send-Time Optimization Combined with Dynamic Offers

Domino's Pizza (US) used Salesforce Marketing Cloud's Einstein Send Time Optimization alongside dynamic offer blocks that varied discount depth per subscriber based on predicted churn risk. High-churn-risk subscribers received a higher discount; low-risk subscribers received a standard offer. The AI component determined both when to send and what offer depth to render.

Salesforce reported a 16% improvement in click-to-open rate and a measurable reduction in unsubscribe rate over the prior campaign period. The unsubscribe reduction was attributed primarily to the send-time component — subscribers receiving emails at predicted-optimal times showed lower fatigue signals.

Case 4 — B2B SaaS: Behavioral Trigger Sequences with Generative Copy Variants

HubSpot published internal data from their own email marketing operations — making this a self-reported case where the brand is also the tool vendor, which is a notable limitation. They used their AI-generated email content feature alongside behavioral trigger logic to personalize follow-up sequences based on product usage signals.

The reported outcome was a 14% higher reply rate on AI-personalized follow-up sequences compared to their prior templated sequences, measured across their own customer base over Q3 2024. HubSpot noted the improvement was concentrated in the first email of a sequence — subsequent emails showed no statistically significant difference.

Cross-Case Patterns and Limits

Looking across these four cases, a few patterns emerge — though none of them should be treated as universal findings.

• Open-time rendering (Cases 1 and 3) consistently outperforms send-time static assembly in the cases where the two have been compared. The ability to use real-time inventory and behavioral signals at the moment of open is a structural advantage over content locked at send.
• Send-time optimization alone produces modest lift — in the range of 5–16% on click-to-open metrics in the documented cases. It is the least technically complex of the AI methods here, and the returns reflect that.
• Generative subject lines beat human baselines in controlled comparisons, but the margin is smaller than vendor marketing implies. A 2.6pp open rate improvement is real but not transformative on its own.
• Behavioral trigger + generative copy shows the most concentrated lift at the top of a sequence (Case 4). The signal fades in subsequent emails — which suggests the personalization advantage is primarily in the initial relevance signal, not in the copy itself.

What These Cases Don't Tell You

None of these cases isolate the AI component from everything else that changed during the test period. In Case 1, the product recommendation engine changed at the same time the rendering method changed. In Case 3, two AI methods ran simultaneously. In Case 4, the brand is the vendor.

List quality, deliverability, and offer competitiveness are also uncontrolled across all four cases. A 29% revenue-per-send improvement on a well-maintained list of high-intent buyers is a different result than the same number on a cold or decayed list.

Metrics Observed Across Cases

Source Notes

All four cases were sourced from publicly available vendor case studies or brand-published data. None have been independently replicated or peer-reviewed. The Marks & Spencer and Domino's cases have secondary coverage in trade publications (Econsultancy, MarTech) that corroborates the claims without adding new data. The Phrasee/Jacquard and HubSpot cases are single-source.

For practitioners building internal business cases, these records are most useful as directional precedents and for identifying which AI email methods have documented results at all — rather than as precise benchmarks to replicate.