Three tenths of one percent. That is the spam complaint rate at which both Gmail and Yahoo say your mail stops being welcome, and it is the number that quietly decides whether anything else in this article matters. Three complaints in a thousand and you have a delivery problem. No subject line written by any model survives that.
Almost every pitch for AI email marketing is aimed at the middle of that pipeline. Which is worth putting first, because open rates are falling for most small senders and the tools being sold as the fix are all aimed at the middle of the pipeline while the problem usually sits at either end.
- An email programme has three stages: who receives it, what it says, when it lands. AI moves the number on the first, moves only the effort on the second, and has nothing to do with the third.
- Gmail requires bulk senders above 5,000 messages a day to Gmail accounts to run SPF, DKIM and DMARC, and to keep spam rates in Postmaster Tools below 0.3 percent, with 0.10 percent as the level to aim at.
- Yahoo sets the same 0.3 percent spam rate ceiling, requires a working one-click unsubscribe header for bulk senders, and asks that unsubscribes be honoured within 2 days.
- The FTC's CAN-SPAM guidance allows 10 business days to honour an opt-out and puts penalties at up to 53,088 dollars per individual email in violation.
- Segmentation is the one stage where a model reading your own order history changes the outcome rather than the workload, because it can hold more customers in view than you can.
- Better copy raises the ceiling on a campaign. Better targeting raises the floor on every campaign, which compounds.
- Send-time optimisation is the most oversold feature in the category for a small list, because the sample size needed to tell one hour from another is larger than most lists.
What are the three stages of an email programme?
Every marketing email passes through the same three gates, and each gate has a completely different relationship with automation. Separating them is most of the analysis.
| Stage | What decides it | Does AI move the number, or the effort? | What that means for you |
|---|---|---|---|
| Who receives it | Segmentation against your own order and browsing history | The number. A model can hold thousands of customer histories in view and propose groupings you would not find by hand | This is where to spend your attention, and it is the stage the tools advertise least |
| What it says | Subject line, body, offer, the reason to open | Mostly the effort. Output volume goes up sharply, quality goes up modestly, and the offer itself is unchanged | Useful, cheap, and not the lever you think it is. Twenty subject lines to choose from beats one you agonised over |
| When it lands | Authentication, complaint rate, list hygiene, unsubscribe handling | Neither. This is DNS records and list discipline, not language | Fix this first or the other two stages are theoretical |
| Whether it is legal | Accurate headers, a physical address, a working opt-out honoured in time | Neither, and generated copy can quietly break the subject line rule | One review of your template settles it permanently. The penalty is per email |
The fourth row is not a stage so much as a condition on all three, and it is included because a generated subject line is the one place AI can create a legal problem rather than a marketing one.
Where does AI actually move the number?
In segmentation, and the reason is a capacity difference rather than an intelligence one. You know your customers better than any model does. You cannot hold four thousand of them in your head at once, and that is the entire job.
Consider what a useful segment looks like for a shop. People who bought a consumable ninety to a hundred and twenty days ago and have not reordered. People who bought a gift item once, in November, and never returned. People who opened the last four emails and clicked nothing, which is a different problem from people who opened nothing. People whose average order value doubled this year. Each of those is a different message, and each of them exists in your order table right now.
Doing that by hand means writing queries or exporting to a spreadsheet and thinking hard, which is why most small senders have one list and send everyone the same thing. A model with access to your order data can propose those groupings, describe them in plain language, and tell you how many people are in each. The output is not a message, it is a list of lists, and it is the highest-value thing in this whole area.
The measured effect is the mechanism, not magic: sending a relevant message to four hundred people beats sending a generic one to four thousand, on revenue and on your complaint rate at the same time. That second part is why segmentation is also a deliverability tactic, which is the connection almost nobody draws. The same arithmetic decides whether an automated outreach agent produces meetings or complaints, on a list of people who never asked to hear from you.
Better copy raises the ceiling on one campaign. Better targeting raises the floor on every campaign after it, because it protects the complaint rate that decides whether the next one is delivered. One is a nicer email. The other compounds.
Where does AI only move the effort?
In the writing, and this is not a small thing, it is just a different thing from what is advertised.
Subject lines are the clearest case. A model will give you twenty in ten seconds, and choosing from twenty is genuinely better than polishing one, because the variation is where the value is rather than the craft. This is the single best use of AI in an email programme and it costs nothing.
Body copy is where the claim gets weaker. The email that works is the one with a real reason to open: a genuine restock, a genuine problem you solved, a genuine price change, a piece of information the reader wanted. A model cannot supply the reason. It can express a reason well once you have one, and it will express an absent reason at length, which is the failure mode. The specificity problem is identical to the one on product pages, which we work through in the piece on how much editing AI written shop copy needs before it ships.
The honest framing for stage two is this: AI removes the excuse not to send. If the newsletter has been stuck for three weeks because writing it feels heavy, that specific blocker is gone, and frequency is worth more to a small list than polish. Just do not expect the words to be the reason it works.
What decides whether the email lands at all?
Authentication and complaint rate, and both are settled long before you write anything. This is the stage where a couple of hours of unglamorous configuration outperforms every content decision you will make this year.
Gmail's requirements
Google's email sender guidelines set a floor for everyone and a higher bar for volume. All senders need SPF or DKIM on their sending domain, valid forward and reverse DNS records, and a TLS connection. Senders of more than 5,000 messages a day to Gmail accounts, a threshold in force since 1 February 2024, need both SPF and DKIM plus a DMARC policy, with DMARC alignment against the domain in the From header. The spam rate reported in Postmaster Tools must stay below 0.3 percent, and Google names 0.10 percent as the level to monitor toward. Bulk marketing mail must carry a one-click unsubscribe header alongside a visible unsubscribe link in the body.
Yahoo's requirements
Yahoo's sender best practices land in the same place, which is useful because it means there is one standard to meet rather than two. SPF or DKIM at minimum, both plus a valid DMARC policy for bulk senders, and the same 0.3 percent spam rate ceiling, calculated on mail delivered to the inbox. On list hygiene Yahoo is direct: sending to people who do not read your mail or who report it will harm your delivery metrics and reputation. It recommends double opt-in to cut invalid recipients, and asks that unsubscribes be honoured within 2 days, with the RFC 8058 post method highly recommended over mailto.
Note the gap between that 2 days and the 10 business days the law allows. The mailbox providers are stricter than the statute, and they are the ones who decide whether you reach the inbox, so the tighter number is the real one.
The legal floor
The FTC's CAN-SPAM compliance guide lists requirements that are mostly template work: accurate From, To, Reply-To and routing information, a subject line that reflects the message, clear disclosure that the message is an advertisement, a valid physical postal address, and a conspicuous explanation of how to opt out. Opt-outs must be honoured within 10 business days and the mechanism kept working for at least 30 days after sending. Hiring another company to run your email does not transfer the responsibility. Penalties run to 53,088 dollars per individual email in violation.
Two of those interact with generated copy directly. A subject line must reflect the content, and a model optimising for opens will happily write one that overstates. And the physical address requirement is a template field, so if you rebuild your email template with an assistant and it drops the footer, you have created a per-email liability without noticing.
Does AI written email hurt deliverability?
Not directly. No mailbox provider publishes a rule against machine-written text, and the signals they do publish are behavioural: complaint rate, engagement, bounce rate, authentication. A model did not write those.
Indirectly, there are two real routes and both are about behaviour rather than authorship. The first is volume. Generated copy makes sending easy, sending more often raises absolute complaints, and complaint rate is the metric with a hard ceiling on it. A shop that went from monthly to weekly because drafting got cheap can cross 0.3 percent without ever writing a bad email. The second is sameness. If your mail reads like everyone else's mail, engagement drifts down, and engagement is a delivery signal on a slow fuse rather than an immediate one.
The practical guard is to treat your complaint rate as a budget rather than a threshold. Look at it after every send. If it climbs while your frequency climbs, the frequency was the change, not the copy.
What about send-time optimisation?
Treat it as the most oversold feature in the category if your list is small, and the reason is arithmetic rather than scepticism about the technique.
To tell one send hour from another you need enough opens in each hour for the difference to be distinguishable from noise. A list of two thousand with a twenty percent open rate produces four hundred opens spread over a day, which is a handful per hour. Any per-hour ranking built on that is mostly random, and a tool that presents it confidently is presenting noise with a chart around it. Larger senders genuinely benefit here; a small shop is being sold a feature its data cannot support.
What does work at small scale is coarser and duller. Pick a consistent day and hour, keep it, and let your readers learn when you arrive. Consistency is a real effect at any list size, and it is free.
What about the automated flows rather than the newsletter?
They matter more than the newsletter for most shops, and they are the part of an email programme where AI is least useful, which is worth knowing before you buy a tool to write them.
An abandoned basket sequence, a post-purchase check-in, a reorder reminder for a consumable: these send themselves once configured, they fire on behaviour rather than on a calendar, and they carry most of the revenue in a small programme because the timing is perfect by construction. The writing is a one-off job. You draft four emails once, edit them properly, and they run for two years.
Which inverts the usual advice. Generated copy is most valuable where volume is high and each piece is disposable, and the flows are the opposite: low volume, high leverage, permanent. Write those four emails yourself, or draft them with a model and then edit them like you mean it, because a clumsy sentence in a flow gets sent ten thousand times.
The place AI helps in flows is not the words but the trigger. Working out that your consumable actually reorders at ninety days rather than the thirty the platform defaults to, or that basket abandonment on your site clusters at the shipping-cost step rather than at checkout, is a question about your own data. That is stage one wearing different clothes, and it is the same answer: point the model at the order history, not at the copy.
The order of operations
Do these in order and skip nothing, because each step makes the next one measurable.
- Authentication and unsubscribe. SPF, DKIM, DMARC on your sending domain, a working one-click unsubscribe, the physical address in the footer. An afternoon, once.
- List hygiene. Remove addresses that have not opened in six months. It feels like deleting revenue and it is buying delivery for everyone who remains.
- Segments before subjects. Get your model to propose five segments from your order history and their sizes. Send to one of them.
- Then use it for copy. Twenty subject lines, pick one. Draft the body, cut a third, add the specific fact that gave you a reason to send.
- Watch the complaint rate, not the open rate. Opens are increasingly unreliable as a metric. Complaints are what the providers act on.
Steps one and two involve no AI at all, and they are the two that determine whether the rest has any effect. That ordering is the whole point of splitting the pipeline into stages, and it is the opposite of how the category is sold.
Where this fits with everything else
Email is the channel where the merchant's own data is the raw material, which makes it the place where AI has the most to work with and the least room to invent. A model writing a product page can only guess at your product. A model reading your order history is working from facts you actually own.
That only holds if the data is reachable. A shop whose orders live in one platform, whose customer records live in another and whose email tool sees neither cannot do stage one at all, which is why most small senders are stuck optimising stage two. If your storefront and its back office are one system, segmentation stops being an export job, which is the case for an AI generated shop that keeps its own order data, and the credit costs of running one are the number to weigh against a stack of connected tools.
Set against the rest of the week, email sits high on the list. The wider ranking is in our piece on which hours of a founder's week AI genuinely gives back, and the tool-by-tool version is in the shortlist of AI tools worth their subscription. Email earns its place for one reason: it is the only channel where you own both the list and the data that tells you what to say to it.