Almost every AI feature sold to email marketers is a model trained on your engagement history. That raises a question the vendors are not eager to answer: which engagement history? Since September 2021, roughly half of all tracked email opens have been machine prefetches by Apple Mail rather than human behaviour. A model trained on that data is learning the schedule of a proxy server. Ask the question before you buy the feature.
Send-time optimisation is the clearest example
Send-time optimisation predicts, per recipient, the hour they are most likely to open, and schedules accordingly. Mailchimp describes its implementation as requiring enough data from your previous sends to calculate an optimal time, and gates it behind the Standard plan or higher. The mechanism is sound. The input is the problem.
If the model is trained on open timestamps, and Litmus put Apple Mail clients at roughly 46% of tracked opens in September 2025, then for around half your list the model is fitting a curve to when Apple’s privacy proxy happened to fetch an image. That is not a small amount of noise added to a good signal, it is a substantial fraction of the training data being generated by something with no opinion about when to read email.
This does not make the feature worthless. Click-trained models are unaffected, and for the portion of your list on Gmail and Outlook the open signal is still real. It does mean the honest expectation is a small improvement, not a transformation, and that you should ask your vendor a specific question: does the model exclude Apple Mail Privacy Protection opens, and is it trained on clicks or on opens?
Where AI genuinely earns its place
| Feature | Does it work? | What to watch |
|---|---|---|
| First-draft copy generation | Yes, as a drafting accelerator | Output converges on a recognisable register. Every business using the same tool sounds the same, which is the opposite of what a small brand needs |
| Subject line variants | Yes for generating options | Ranking them requires an A/B test you probably cannot power, and open rate is not a valid judge of that test |
| Send-time optimisation | Marginal | Ask whether it trains on opens or clicks. Open-trained models are contaminated by Apple Mail prefetching |
| Predictive churn and purchase scoring | Needs volume most small lists lack | Ask for the minimum contact and event history required. If your list is under a few thousand active buyers, the prediction is largely a prior |
| Product recommendations | Yes, with a real catalogue | Needs enough purchase co-occurrence data to beat "best sellers," which for small catalogues it often does not |
| Deliverability and content scanning | Yes, and underrated | Checks links, authentication, and rendering. Boring, mechanical, and the most reliable value on this list |
Note which row is the least glamorous and the most useful. Automated pre-send checks on authentication, broken links, image-to-text ratio, and client rendering catch the failures that actually cost money, and they do not require a model to have opinions about your customers.
The real risk is volume, not quality
The plausible failure mode of AI in email is not embarrassing copy. It is that reducing the cost of producing an email to near zero removes the only constraint that was keeping send frequency sane. Frequency is what generates complaints, and complaints are what mailbox providers filter on. Since February 2024, Google has told bulk senders to keep the user-reported spam rate below 0.3% and recommends staying under 0.10%, and a sender above 0.3% is ineligible for mitigation until they hold below it for seven consecutive days.
A tool that lets a one-person marketing team produce five campaigns a week instead of one has not made that team five times more productive. It has given them five times as many opportunities to cross a threshold that takes a week of clean sending to recover from.
The bottleneck in email marketing was never the writing. It was having something worth saying, and no model has solved that.
What to feed it, and what not to
Before pasting a customer list, support transcripts, or purchase history into a general purpose AI tool, check where that data goes and whether it is retained or used for training. If you process EU subscribers, GDPR Article 6 requires a lawful basis for that processing and your privacy notice has to reflect what you are actually doing. Features built into your email platform are usually on firmer ground here than a copy and paste into a chat window, because the data was already in scope. Usually is not always: read the vendor’s data processing terms.
A reasonable position
Use AI where the task is mechanical and verifiable: drafting, variant generation, pre-send technical checks, summarising reply threads. Be sceptical where it claims to predict human behaviour from a signal that is half machine-generated. And measure the same way you would measure anything else, against a holdout, because a feature that came with your subscription still has to prove it did something.
Key takeaways
- Ask what any predictive email feature trains on. Open-based models have been contaminated since Apple Mail Privacy Protection shipped in September 2021.
- Litmus put Apple Mail clients at roughly 46% of tracked opens in September 2025, so open-trained send-time optimisation is fitting a proxy server’s schedule for much of your list.
- The most reliable AI value in email is the least exciting: automated pre-send checks on authentication, links, and rendering.
- The genuine risk is frequency. Cheap production removes the constraint that kept sending sane, and Google’s spam rate ceiling is 0.3%.
- Check where your data goes before pasting customer records into a general purpose tool, and make sure your privacy notice matches.
Related reading
Sources
- Use Send Time Optimization, Mailchimp Help
- Email Client Market Share, Litmus
- Apple advances its privacy leadership with iOS 15, Apple Newsroom (2021)
- Email sender guidelines, Google Workspace Admin Help
- Article 6 GDPR: Lawfulness of processing, GDPR text
- AI
- Email Marketing
- Automation
Written by
Valter Brandt
Chief Marketing Officer
Valter Brandt is the Chief Marketing Officer of ThisCom, working with clients across the United States and internationally. He has led marketing strategy through the major shifts in social advertising, mobile, content marketing, programmatic media, and marketing automation.
All articles by Valter Brandt →
