AI and Deliverability

AI now shapes email at both ends: senders use it to write and time their emails, and mailbox providers use it to filter, sort and summarise them. This article explains what each side does today, separates known facts from reasonable guesses, and shows how to use AI without hurting whether your mail reaches the inbox.

AI on both sides

Deliverability means whether your email reaches the inbox, rather than the spam folder or nowhere. AI touches it in two places. You may use AI tools before you send. The mailbox provider (Gmail, Outlook, Yahoo and so on) uses AI after you send.

AI on the sending side

Writing emails with AI

Large language models (LLMs, AI tools such as ChatGPT, Gemini or Claude that write text) can draft subject lines, body copy and versions for different groups of readers or languages.

• What helps: relevant content. If AI helps you send each reader something that fits their interests, more people open, click and reply. Those actions are what filters reward.
• What hurts: mistakes at scale. AI writes faster than you can check. An unchecked email can contain wrong prices, made-up claims or an odd tone, and readers complain.
• A possible risk: if many senders use the same tools, their emails may start to sound alike. Some people worry filters could link that style to spam. There is no public evidence that this happens today.

Choosing send times

Send-time optimisation tools look at when each reader usually opens email, then deliver at that hour. This can raise engagement. It also spreads your sending over the day instead of one big burst, which providers handle more easily.

One catch: these tools often learn from open data. Apple Mail Privacy Protection loads images automatically, so many recorded “opens” are not real reading. Check that your tool also uses clicks or other real actions.

Predicting who will drift away

Prediction tools can flag readers who are likely to stop engaging, complain, or go silent for good. That lets you act early: send less often, change the content, or ask if they still want your mail. Long-silent addresses matter because some get recycled into spam traps (old addresses a provider reuses to catch senders who never clean their lists).

AI on the receiving side

Providers have used machine learning (software that learns patterns from many examples) in spam filters for years. They share few details. Here is what is reasonably clear.

Reading the message

Filters look at what an email is trying to do, not just at single words. They try to spot:

• phishing, which means emails that trick people into giving away passwords or money
• fake urgency and pressure (“your account closes in 1 hour”)
• where links really lead, not just the text shown
• text hidden inside images

Watching behaviour

Filters also compare your sending with your normal pattern. A sudden jump in volume, or mail to addresses you’ve never written to before, stands out. They also predict whether each reader will want the message, based on how that reader and similar readers acted before. AI Filters & User-Level Trust covers this in detail.

Learning fast

When users start reporting a new kind of spam, providers can update their models quickly. Old tricks, such as swapping letters for look-alike characters or adding hidden text, are now well known to filters. A trick that works today may stop working within days.

AI inside the inbox

AI now works inside the inbox too, after the filter has done its job. Gmail, Apple Mail, Outlook and Yahoo Mail offer features that summarise emails, highlight the ones that look important, or sort mail into categories.

This changes what “reaching the reader” means. Your email can land in the inbox, and the reader may still see only a one-line summary. Some practical steps follow from that:

• Put the main point in the subject line and the first sentence.
• Write the key facts (dates, prices, actions) as real text, not only inside images.
• Keep one clear purpose per email, so a summary can’t miss it.

How these features choose what is “important” is not public. Expect them to favour mail from senders the reader already engages with.

Do filters punish AI-written email?

No major provider has said it filters email for being written by AI. Their published rules focus on whether mail is wanted, authenticated and safe.

Tools that detect AI writing do exist. Some look for statistical patterns, such as very predictable word choices. Some look for watermarks: hidden patterns that certain AI tools, such as Google’s SynthID, can add to their text. These tools are unreliable on short texts like emails, and lightly edited text is hard to detect.

So the risk is not “AI wrote it”. The risk is what careless AI use produces: more volume, less relevance, and errors that annoy readers. AI-written spam gets caught because it is spam.

• Let AI draft, and have a person review. A human check catches mistakes and keeps your voice.
• Use AI to be more relevant, not to send more. More email to the same people tends to raise complaints.
• Watch the law. Some new laws, such as the EU AI Act, require certain AI-generated content to be labelled. If you’re unsure whether that covers your emails, ask a lawyer.

Automating deliverability work

AI monitoring tools can spot unusual changes in your numbers early. Examples are a rise in bounces or complaints, or a drop in opens at one provider. Some can suggest a likely cause. Many routine actions are safe to automate. Others need a person.

Action	Automate?	Notes
Remove addresses that don’t exist	Yes	After a permanent “unknown user” bounce (5.1.1). Not every 5xx means the address is bad.
Remove people who complain	Yes	Use the complaint reports from feedback loops (FBLs).
Track temporary failures	Yes	Remove an address only after repeated failures over several days.
Slow down when a provider defers	Yes	Defer means “try again later” (a 4xx code). Lower your sending rate.
Alert on blocklist listings	Yes	A person should then look into the cause.
Increase volume	Partly	Only within a warm-up plan you set in advance.
Stop mailing a whole segment	No	Needs judgement about business impact.
Change SPF, DKIM or DMARC records	No	A DNS mistake can stop all your mail.
Contact a provider’s postmaster team	No	Needs a person to explain and follow up.

As routine work gets automated, deliverability jobs shift toward strategy, handling incidents and explaining results to the rest of the business. Beyond Deliverability: Email Operations covers that shift.

Risks to watch

Better impersonation

AI makes it easy to copy a brand’s writing style and build convincing fake login pages. That makes proof of identity more important. DMARC (a DNS record that tells providers what to do with mail that fails SPF and DKIM checks) is your main defence. Move it from p=none (monitor only) to p=quarantine or p=reject once your reports show all your real mail passes.

Privacy limits

Some AI features run on the reader’s own device. Apple, for example, runs much of Apple Intelligence on the device. Privacy laws may also limit how much tracking data you and providers can collect. Expect less data about individual readers over time, for you and possibly for filters.

More spam, tighter filters

AI makes convincing spam and phishing cheaper to produce. Providers respond by tightening filters for everyone. Senders with strong authentication, low complaints and engaged readers stand out more as that happens.

What to do now

Known today, so act on it:

1. Publish SPF, DKIM and DMARC. Gmail and Yahoo require all three for bulk senders (5,000+ messages a day to their users), and Microsoft (Outlook.com) has required them since May 2025. Check yours with the DNS lookup and the test inbox.
2. Move DMARC toward p=reject once your reports show your legitimate mail passes.
3. Keep spam complaints under 0.3% and offer one-click unsubscribe.
4. Have a person review every AI-written email before it goes out.
5. Put the key message in the subject line and first sentence, as plain text.

Reasonable expectations, so prepare for them:

• Inbox AI will play a bigger part in what readers notice.
• Filters will lean even more on each reader’s own behaviour.
• Rules on labelling AI content may spread.
• Tracking data will keep shrinking, so clicks, replies and purchases will matter more than opens.