Almost every guide to AI and email says the same thing: use it to write subject lines faster. That advice is not wrong, exactly. It’s just aimed at the least important part of the job. The two things that actually changed about email marketing automation in the last two years have nothing to do with who writes the copy.
The first is that the mailbox providers turned their guidelines into enforced requirements, and automation is what pushes small senders across the threshold where those requirements bite. The second is stranger: an AI now reads your email, and in some inboxes it writes the preview your subscriber sees instead of the one you wrote.
Get those two right and mediocre copy still performs. Get them wrong and brilliant copy never arrives.
Automation raises your volume, and volume changes the rules
Google and Yahoo announced near-identical sender requirements in October 2023, effective 1 February 2024, with the one-click unsubscribe deadline landing 1 June 2024. Microsoft followed with enforcement from May 2025. The trigger is roughly 5,000 messages a day to that provider’s consumer mailboxes — and Google counts all subdomains of your primary domain together.
Here’s the part that catches people out: once you’re classified as a bulk sender, the status doesn’t lapse because you had a quiet month. And a welcome sequence, an abandoned-cart flow and a weekly newsletter running simultaneously will cross 5,000 a day at a list size that feels modest.
| Requirement | Applies to | Where it usually breaks |
|---|---|---|
| SPF and DKIM, both passing | All senders | A second sending tool added later and never authenticated |
| DMARC record, at least p=none, with alignment via SPF or DKIM | Bulk senders | Record published but not aligned to the From domain |
| Valid forward and reverse DNS (PTR) | Bulk senders | Dedicated IPs and self-hosted setups only |
| TLS on connection | All senders | Rarely — handled by your provider |
| One-click unsubscribe (RFC 8058) plus a visible link, honoured within two days | Marketing and promotional mail | A footer link with no header — the most common single failure |
| Spam complaint rate below 0.3% | All senders | List quality, which drifts |
Two clarifications worth having. Transactional mail — password resets, order confirmations, shipping notices — is exempt from one-click unsubscribe, and you should not add the header to it, or people will accidentally unsubscribe from their own receipts. And the rules target personal mailboxes, not business inboxes on Workspace or Microsoft 365, though there’s no advantage to being non-compliant either way (Google’s sender guidelines FAQ).
Most senders who fail an audit fail on DMARC alignment or a broken one-click unsubscribe header. Both are twenty-minute fixes. Neither can be fixed by better copy.
The complaint-rate maths that punishes scale
This is the number to internalise before you automate anything.
Google’s ceiling is 0.3%. Practitioners treat 0.1% as the operating target and 0.3% as the point where you’re already in trouble. And the arithmetic is brutal at small volumes: deliver 10,000 emails and it takes 30 people clicking “report spam” to hit the ceiling. Thirty. Out of ten thousand.
Cross it and you lose eligibility for Google’s delivery mitigation until your rate stays below 0.3% for seven consecutive days — which means you can’t get help precisely when you need it, and you’re waiting a week with reduced inbox placement in the meantime.
Now put that beside the standard AI pitch, which is that you can produce more campaigns in less time. More sends to the same list, with the same relevance, produces more complaints. AI makes the cheap part cheaper and does nothing at all for the constraint. This is the same failure pattern as automating a bad process, which is why the order of operations in automating work with AI tools starts with the workflow rather than the tool.
The practical rule: before you increase frequency, set up Google Postmaster Tools and Yahoo’s feedback loop, and watch the rate for a month. If you can’t see the number, you can’t safely send more.
An AI reads your email before your subscriber does
Here’s the change almost nobody has built for.
Apple Intelligence, from iOS 18, replaces the standard inbox preview with a one- or two-line AI summary. It happens before the open, and it is on by default for users on supported devices. Gmail’s Gemini summaries and Outlook’s Copilot summaries have mostly operated after the open, and Yahoo has its own version.
Sources genuinely disagree on how far Gmail has gone. Some report that Gemini summaries remain post-open and don’t touch the inbox preview; others report that Gmail moved into a broader “Gemini era” in early 2026 with AI Overviews and an AI Inbox mode that filters and prioritises messages before the user sees them. Both accounts are current. The honest position is that Apple’s pre-open summary is confirmed and widespread, Gmail’s is expanding and inconsistently documented, and the direction of travel is not in doubt (Litmus, Dyspatch).
What this does to your email is specific and predictable. Summarisers extract concrete things — offers, deadlines, dates, discount codes, actions. They flatten narrative, brand voice, slow builds and clever misdirection. Your carefully written preheader may simply not be what the subscriber reads.
Four consequences that should change how you build campaigns:
- Image-only emails summarise badly. Testing by inbox-monitoring vendors found summaries of image-only emails were substantially weaker than the same emails with live text. Live text and real alt text are now a deliverability-adjacent concern, not just an accessibility one.
- Front-load the point. The first one or two sentences of the body are what the summary is built from. Burying the offer under a story is now a structural mistake.
- One core message per email. A summariser given four competing offers will pick one, and it may not be yours.
- Make codes and deadlines unambiguous. Summarisers have been observed pulling the wrong number out of an email — a struck-through price, an old code — and presenting it as the offer.
The uncomfortable second-order effect: summaries create informed non-openers. Subscribers who get the value from the preview and never open. Some of your engagement isn’t disappearing, it’s becoming invisible (Salesforce).
Open rate is now broken twice
It was already unreliable. Apple’s Mail Privacy Protection, live since 2021, pre-fetches images and registers opens the subscriber never made, which inflated open rates for anyone with a meaningful Apple audience.
AI summaries push in the opposite direction, suppressing genuine opens from people who got what they needed from the preview. You now have one metric distorted upward by one mechanism and downward by another, in proportions you can’t measure.
So stop optimising for it. The metrics that still mean something:
| Metric | Why it survives |
|---|---|
| Clicks per delivered email | Requires a deliberate action; unaffected by pre-fetching or summaries |
| Revenue per recipient | The only number that settles arguments about frequency |
| Complaint rate | Directly governs whether you’re delivered at all |
| Unsubscribe rate per send | The early-warning signal that precedes complaints |
| List growth net of churn | Catches the send-more-to-earn-more trap |
Keep reporting open rate if your boss wants it. Just don’t make decisions with it.
Where AI genuinely belongs in email marketing automation
Ranked by return, which is roughly the inverse of how it’s usually marketed:
| Use | Why it pays |
|---|---|
| Segmentation and suppression | Sending less to the wrong people protects your complaint rate. Highest return, least discussed |
| Re-engagement triage | Deciding who to sunset. Removing dead addresses improves every downstream number |
| Subject and preview variants for testing | Volume of options is exactly what a model is good at — but the test decides, not the model |
| Send-time and frequency modelling | Real gains, but only with enough history for the model to learn from |
| Drafting body copy | Saves time, moves no needle on its own |
| “Personalised” generated copy per recipient | Lowest return, highest risk of an obviously wrong sentence going out at scale |
Segmentation sits at the top for a reason that follows directly from the maths above. Every campaign you don’t send to a disengaged segment is complaints you don’t accrue. AI is genuinely useful at spotting behavioural clusters in list data that a manual rule would miss — and that work is invisible, unglamorous, and worth more than a year of subject-line optimisation.
If you’re using AI for the copy itself, the drafting discipline in writing content faster with AI applies, and turning one piece of content into ten is usually a better use of the same hour than generating a campaign from nothing. For ecommerce specifically, email sits downstream of the product data work covered in AI ecommerce tools that increase sales. Platform-side, the Klaviyo AI review covers what the built-in features actually do.
Three things to keep AI away from
Consent. No model decides who is on your list. Permission is a legal fact with a source and a timestamp, and “the AI found these addresses” is the opening line of a compliance problem.
Your suppression list. Unsubscribes and complaints propagate across every tool you send from. If you run two platforms and only one has the suppression, you will mail someone who already opted out — which is both a legal exposure and a guaranteed complaint. Centralise it and audit it manually.
Cold outreach. Generating a thousand personalised cold emails is the single fastest route to the complaint threshold, and it puts your sending domain — the one your actual customers receive mail on — at risk. If you do cold outreach at all, do it from a separate domain.
The order to do this in
- Authenticate everything. SPF, DKIM, DMARC with alignment, for every tool that sends on your behalf. Not just the main one.
- Verify the one-click unsubscribe header exists on marketing mail and is absent from transactional mail. Send yourself a test and check the headers.
- Turn on Postmaster Tools and the Yahoo feedback loop. Record a baseline before you change anything.
- Clean the list. Sunset addresses that haven’t engaged in six to twelve months. This will feel like losing something. It isn’t.
- Rebuild templates for summarisation. Live text, real alt text, one message, offer in the first two sentences.
- Then automate. Welcome, abandonment, post-purchase, re-engagement — in that order.
- Then add AI, starting with segmentation.
Six of those seven steps happen before AI enters the picture, which is the honest summary of this whole topic. The tooling budget question is covered in cutting business costs with AI, and if you’re assembling a wider stack, the best AI tools for small business owners is the map. One thing that transfers directly from search: as with SEO content written with AI, the machine reading your work first has changed what “good” looks like — and the answer in both cases is clarity, not cleverness.
Frequently asked questions
Do the bulk sender rules apply to my small list?
The threshold is roughly 5,000 messages a day to a single provider’s consumer mailboxes, counted across all subdomains of your primary domain. Multiple automated flows running at once reach that faster than a single newsletter does. The baseline requirements — SPF or DKIM, TLS, and a spam rate under 0.3% — apply to every sender regardless of volume.
What spam complaint rate should I actually target?
Under 0.1%. The published ceiling is 0.3%, but that is the enforcement line, not a target — and at 10,000 delivered emails it takes only 30 complaints to reach it. Crossing 0.3% costs you access to Google’s delivery mitigation until your rate stays below it for seven consecutive days.
Will AI summaries lower my open rates?
Probably, where subscribers use Apple Mail on supported devices, because a useful pre-open summary creates subscribers who get the value without opening. Apple’s Mail Privacy Protection simultaneously inflates opens by pre-fetching images. The sensible response is to stop treating open rate as a decision metric and use clicks per delivered email and revenue per recipient instead.
How do I write emails that summarise well?
Put the core offer in the first one or two sentences of the body, keep one message per email, use live text rather than a single hero image, write real alt text, and make discount codes and deadlines unambiguous so a summariser can’t pull the wrong number. Assume the summary is what your subscriber reads.
Is AI-written email copy penalised by spam filters?
No. Filters respond to authentication, sending reputation, complaint rates and engagement, not to how the words were produced. What gets penalised is the behaviour AI makes easy — sending more, to less relevant people, more often.
Can I use AI to build or find an email list?
No. Consent is a legal fact that needs a documented source and timestamp, and addresses gathered or inferred by a tool have neither. Scraped and generated lists produce complaint rates that damage the sending domain your real customers receive mail on.
