Writing

• Marketing Workflows
• 1 October 2026

AI email marketing: personalisation beyond Hi {first_name}

AI improves email marketing by deciding four things per person that we used to decide per list: who should get a message, when it should land, which channel it should go through, and what goes inside it. The platforms do this with predictive models built on your own order and engagement data, so most of the gain comes from better timing and relevance.

The gap is in how few teams use it that way. Salesforce’s State of Marketing 2026 survey of 4,450 marketing decision makers found 75% have adopted AI, yet 84% admit they still run generic campaigns. So most of us have bought the engine and are using it to type “Hi Sarah” slightly faster.

How do AI-driven workflows improve email marketing performance?

Each platform has its own names for these features, but they map onto the same four jobs, and I’d judge any tool on how well it does each one with the data you actually have.

Who gets the email: prediction

Predictive fields are the easiest part of AI email marketing to switch on and then never build anything with. Klaviyo, an ecommerce email and SMS platform, calculates predicted customer lifetime value, churn risk, average time between orders and expected date of next order for each profile. You only get them once you have at least 500 customers who have placed orders, at least 180 days of order history, orders within the last 30 days, and some customers with three or more orders. Mailchimp’s equivalent, on Standard plans and above, offers likelihood to purchase within 60, 90 or 365 days, predicted CLV, predicted time to next purchase and a high, medium or low churn risk.

These fields are what make “personalisation” mean something. A replenishment email timed to someone’s expected next order date is personal in a way their first name never will be, and a high-CLV customer showing churn risk deserves a different message from a one-time buyer who was never coming back. If you want the method for turning these scores into segments, I’ve written it up separately in AI customer segmentation.

One field I’d leave alone: Klaviyo’s predicted gender, which the help page says is worked out from a customer’s first name plus census data, and returns likely male, likely female or uncertain. Building a segment on it is how you end up sending the men’s range to a Sam who has been buying dresses for three years.

When it lands: send time optimisation

Send time is where the evidence is clearest and the effort is lowest. Klaviyo’s Personalized Send Time picks an hour within your delivery window for each recipient, based on their engagement history, and reports results against a control group once the campaign has been finished for at least 48 hours. It needs the Marketing Analytics or Advanced KDP package, isn’t yet available in flows, and has to be scheduled at least a day ahead. In Klaviyo’s own beta results (October to November 2025, written up by product marketing manager Alicia Chang), the top 15% of campaigns saw click rates rise 35%, and Shady Rays, a sunglasses brand, reported an over 10% increase in placed order rates across more than 30 email campaigns. That is vendor data on its best performers, so I’d plan on something smaller.

Salesforce’s Einstein Send Time Optimization looks at about 20 factors across 90 days of engagement data, rebuilds its model weekly, and falls back to a generalised model for contacts without enough history. Braze’s Intelligent Timing is the most transparent about its fallback: without enough data it sends at 5pm in the user’s local time. It also excludes machine opens from the email opens it learns from.

Which channel: letting the model pick

If you send email, SMS and push, the channel decision can be automated too. Braze’s Intelligent Channel filter picks the channel with the highest likelihood of engagement for each user, looking at up to the last 100 messages per channel over six months, and only ranks a channel after a user has received at least three messages on it. There’s a lovely warning in the docs: if you haven’t switched on SMS click tracking, SMS can be chosen as someone’s best channel with a 0% engagement rate. The model is only as good as the tracking under it, which is the least glamorous sentence in marketing and also the most frequently ignored.

What goes inside: recommendations and generated copy

Product recommendations are the most familiar AI in email. Dotdigital’s “Best next” model uses collaborative filtering, grouping customers with similar buying habits and suggesting what lookalike customers bought that this one hasn’t, and it needs only order history rather than rich product data. It is a paid add-on, and Dotdigital reviews your data before switching it on.

Generated copy is where I’d be more careful. Adam Peruta and Carrie Riby of Syracuse University’s Newhouse School, working with Ipsos, rebuilt 10 human-made 30-second video ads from major brands with AI and tested all 20 with 3,000 US consumers. Ipsos reports the human-made ads were 14% stronger on short-term effectiveness and 17% stronger on long-term effectiveness, while only 25% of viewers confidently identified the AI versions. That study is about video ads rather than email, but the risk carries over: copy that looks fine in the approval round and underperforms where nobody is checking. I’d use generation for variants and subject line options, and keep a human writing the core message.

Does first-name personalisation still work?

It does, which is slightly annoying for the argument of this article. Navdeep Sahni and Christian Wheeler (Stanford) and Pradeep Chintagunta (Chicago Booth) ran field experiments with an online test-preparation company, MercadoLibre and Stanford University, reported by Chicago Booth Review in 2018. Adding the recipient’s name to the subject line lifted opens by 20%, from about 9% to about 11%, produced roughly 30% more sales leads and cut unsubscribes by 17%.

The catch is that the name tells the reader nothing new. The same email, with the same offer, arrives at the same time as it does for everyone else. The bigger money is in what McKinsey measured in Next in Personalization 2021: 71% of consumers expect personalised interactions, 76% get frustrated when they don’t get them, and personalisation most often drives a 10 to 15% revenue lift (the research is from 2021, so treat it as a direction rather than a forecast).

“Hi Priya, our autumn sale is on” is merge-tag personalisation. “Priya, you bought the 250ml last time and you’re usually back around now, want the refill?” is AI personalisation, because it depends on a predicted date, an inferred product and a timing model. The second one doesn’t even need the name.

Leading platforms for AI-driven email marketing personalisation

These are the platforms I’d shortlist, described from their own documentation. The right one depends more on your data volume and team size than on the feature list.

On the Salesforce side, Einstein Engagement Scoring produces five scores for each contact (email opens, clicks, retention, web conversion and overall engagement) and groups subscribers into personas. The thresholds in the last column decide whether any of this works for you. A charity sending a monthly newsletter to 8,000 supporters won’t hit most of them, and should put its money into cleaner data and better segments before any predictive add-on.

Which companies offer AI-driven email marketing solutions?

Nearly every email service provider now claims AI, so it helps more to sort them by the data they learn from.

• Ecommerce-first platforms such as Klaviyo and Mailchimp learn from order data through store integrations, which is why their predictions are about purchases: next order, lifetime value, churn.
• Enterprise marketing clouds such as Salesforce Marketing Cloud learn from engagement across journeys and CRM records, which suits B2B and large B2C teams with a lot of history.
• Customer engagement platforms such as Braze learn from app sessions and cross-channel behaviour, which is why their strongest features are about timing and channel choice.
• Mid-market suites such as Dotdigital sit between these, with recommendation models that run on order history alone.

When a vendor demos AI to you, ask what the minimum data volume is before the model switches on, what happens to contacts below that threshold, and whether you can run a holdout group against it. If the answer to the last one is vague, you’ll be taking the lift on trust, and we have all sat in the QBR where a dashboard claimed credit for Black Friday.

How to set up an AI personalisation workflow, step by step

This is the order I’d build it in, starting with the parts that need the least data. The broader method sits in how to build an AI marketing workflow, and this is the email version.

1. Fix the tracking first. Click tracking on every channel, order data flowing back into the platform, and machine opens flagged.
2. Switch on send time optimisation for campaigns, with the built-in control group left running. It needs no new content and gives you your first fair lift figure.
3. Check whether you meet the predictive thresholds. If you do, build three segments: high predicted CLV with rising churn risk, customers near their expected next order date, and one-time buyers past their usual reorder window.
4. Write one flow for each segment by hand. Replenishment, win-back and a VIP save are the obvious ones, and they should differ in offer and timing as well as wording.
5. Add recommendations to the flows once you have enough order history for the model to have something to go on.
6. Use generative AI for variants only, then test them against your human-written control.
7. Review monthly against a holdout group, using revenue per recipient as the main measure.

For step 6, give the model the segment logic rather than “write a subject line for our sale”:

You are writing subject lines for a UK skincare brand. British spelling.
Segment: customers whose predicted next order date is within 7 days,
who last bought the 250ml cleanser. Goal: replenishment purchase.
Write 6 subject lines under 45 characters. No first names, no emoji,
no urgency words like "last chance". Two should mention the product,
two should mention timing, two should be plain and practical.
After each, say in one line what it is testing.

For step 7, a prompt that stops the model cheerleading:

Here are results for 4 flows with holdout groups: [paste table].
For each flow, report revenue per recipient for treated vs holdout,
the difference, and whether the sample is large enough to trust.
Flag any flow where clicks rose but revenue did not.
Do not describe any result as a success unless the difference is
larger than the week-to-week variation shown in the holdout.

How do you measure AI email performance when open rates are unreliable?

Apple’s Mail Privacy Protection fetches messages and their images in the background, so the tracking pixel fires while the email is still sitting unread, as Twilio’s guide to MPP explains. The DMA’s Email Benchmarking Report 2025, produced with Marigold from seven UK email service providers, notes open rates “continued to climb post-Apple MPP” to 35.9%, while unique click rates reached 2.3%.

This matters twice over for AI. A send time model trained on machine opens ends up learning the schedule of Apple’s servers, which is why Braze excludes machine opens from Intelligent Timing. And if your success metric is opens, every AI feature will look good, so judge them on clicks, orders and revenue per recipient against a holdout. The wider ROI method is in measuring the ROI of AI in marketing.

What UK rules apply to AI-personalised email?

Under PECR, marketing email needs consent or the soft opt-in, and the ICO’s guidance on what else to consider for email marketing (updated 20 August 2025) points out that tracking pixels, which can record when someone opened an email, their location and their device’s operating system, fall under PECR’s separate rules on storage and access technologies. Data protection law also expects you to be clear, open and honest about how you use people’s information, and the predictions behind personalisation are part of that.

One change that matters for the third sector: under the Data (Use and Access) Act 2025, charities have been able to use a soft opt-in for emails, texts and direct messages on social media since 5 February 2026, for people who have expressed an interest in or offered to support their charitable purpose, provided strict requirements are met. As someone who leads AI adoption in marketing at ACTAsia, a charity, I’d say the sensible order is to get the soft opt-in wording right at every collection point before anyone builds a predictive model on the new list. There is more on how we approached AI there in the ACTAsia case study.

What happens when the inbox uses AI too?

Google announced on 8 January 2026 that Gmail’s AI Overviews would summarise email conversations for everyone, with an AI Inbox that prioritises important emails going to trusted testers first, starting in the US in English with more languages and regions to follow. Apple Intelligence summarises emails in Apple Mail as well. So a growing share of your subscribers will see a machine’s summary of your email before they see your email.

Litmus’s advice, in a 2024 post by Kimberly Huang, is a clear email hierarchy and subject lines that match the body copy, with Litmus email developer Carin Slater warning that a surprise reveal or bait-and-switch may not survive the summary. That happens to be the same thing that makes AI personalisation work: one specific message for one specific reason. A summariser can’t do much with “Don’t miss this!!” and a body of six unrelated offers, and frankly neither can a human. So the next step I’d take is small: pull last quarter’s campaigns, ask which of them could be summarised in one plain sentence, and then look at which segment each was actually written for.

FAQs

What is AI email personalisation?

AI email personalisation uses predictive models to change who receives an email, when it arrives, which channel carries it and what content it contains, for each subscriber. It relies on order and engagement data to predict things such as next purchase date, lifetime value and churn risk. Merge tags such as first name are basic personalisation; AI adds decisions based on predicted behaviour.

Is AI send time optimisation worth using?

For most senders with enough engagement history, yes, because it needs no new content and platforms such as Klaviyo measure it against a control group. Klaviyo’s beta found a 35% click rate increase in its top-performing 15% of campaigns, which is a best case. Judge it on clicks and orders rather than opens, since Apple Mail Privacy Protection inflates opens.

How much data do you need for predictive email marketing?

It depends on the platform. Klaviyo’s predictive analytics need at least 500 customers who have placed orders, 180 days of order history, orders in the last 30 days, and some customers with three or more orders. Salesforce’s Einstein Engagement Scoring needs at least 1,000 events in the 90 days before activation. Below these levels, focus on tracking and segmentation first.

Can AI write marketing emails?

It can write usable drafts and subject line variants, but test them against human-written controls. A 2026 study by Syracuse University and Ipsos found human-made video ads were 14% stronger on short-term effectiveness and 17% stronger on long-term effectiveness than AI-generated versions, even though most viewers couldn’t confidently spot the AI ones. Use generative AI for variants and keep humans on the core message.

Does adding a first name to the subject line still help?

Evidence says yes. Field experiments by Sahni, Wheeler and Chintagunta, reported by Chicago Booth Review in 2018, found adding the name raised opens by 20%, from about 9% to about 11%, and cut unsubscribes by 17%. The name doesn’t change the offer or the timing, so on its own it leaves most of the value of personalisation untouched.

Sources

1. Salesforce, “75% of Marketers Have Adopted AI Yet Still Use It To Send One-Way, Generic Campaigns” (State of Marketing 2026), 19 February 2026
2. Klaviyo Help Center, “Understanding Klaviyo’s predictive analytics”, updated 5 August 2025
3. Klaviyo Help Center, “Drive performance with right time delivery using Personalized Send Time”, accessed 1 October 2026
4. Klaviyo, Alicia Chang, “Personalized send time: AI for send time optimization”, 14 January 2026
5. Mailchimp, “About Predictive Analytics and Demographics”, accessed 1 October 2026
6. Salesforce Help, “Einstein Send Time Optimization”, accessed 1 October 2026
7. Salesforce Help, “Einstein Engagement Scoring”, accessed 1 October 2026
8. Braze Docs, “Intelligent Timing”, accessed 1 October 2026
9. Braze Docs, “Intelligent Channel Filter”, accessed 1 October 2026
10. Dotdigital Help Centre, “Best next predictive product recommendations”, accessed 1 October 2026
11. Ipsos, Adam Peruta, Carrie Riby, Nate Cummings and Nate Pascale, “AI Ads Are Good Enough, And That’s the Problem”, 18 May 2026
12. Chicago Booth Review, Ed Finkel, “Open This, [NAME]! Why Personalized Email Subject Lines Work”, 12 October 2018
13. McKinsey & Company, “The value of getting personalization right, or wrong, is multiplying”, 12 November 2021
14. DMA with Marigold, “Email Benchmarking Report 2025”, 4 June 2025
15. ICO, “Guidance on direct marketing using electronic mail: What else do we need to consider?”, updated 20 August 2025
16. ICO, “Charities given new flexibility to contact supporters under data law change”, 28 April 2026
17. Google, Blake Barnes, “Gmail is entering the Gemini era”, 8 January 2026
18. Litmus, Kimberly Huang, “AI-Generated Email Summaries: What Marketers Need To Know”, 25 June 2024
19. Twilio, “Guide to Apple Mail Privacy Protection (MPP) & iOS 18”, 1 September 2025