AI in email marketing has split the room into two camps, and both are wrong. One camp still writes every subject line by hand and treats anyone using AI as a shortcut-taker with no craft. The other camp asks a chatbot to build the whole HTML template and then wonders why the email looks broken in Outlook. I have run three agencies and thirty years of campaigns, and the truth sits in the middle, closer to boring discipline than either camp wants to admit.
This piece is for both of you. The sceptic who thinks AI is a fad. The enthusiast who thinks it replaces experience. Neither position survives contact with the data, and neither survives contact with a subscriber list segmented by Apple Mail Privacy Protection.
The State of AI in Email Marketing
Where the industry actually agrees
Start with the numbers, because opinions are cheap and benchmarks are not. Validity’s State of Email 2026 report, published on 14 April 2026 from a survey of more than 500 marketing professionals across the US, UK, Australia and New Zealand, found that teams with deep AI integration in their workflow are 75% more likely to achieve an email ROI above 45:1 than teams using AI superficially.
The same cohort is 54% more likely to meet WCAG accessibility standards and 52% more likely to comply with the European Accessibility Act. That’s Validity’s own commercial research, so treat the ROI multiple as directional rather than gospel, but the pattern (AI-mature teams pull ahead on both revenue and compliance) lines up with what I see in client accounts.
The adoption curve backs this up. Litmus’s own trend research states that 34% of email marketers already use AI for copywriting at least occasionally, making it the single most common AI-assisted task in the channel. Production speed has moved just as fast: Knak’s 2026 statistics roundup, which draws on Litmus’s own survey data, reports that 76% of marketing teams now produce an email in a week or less, up from 21% in 2023. The same source flags a gap worth sitting with: 87% of businesses apply AI somewhere in their email workflow, yet only 6% qualify as genuine high performers. Adoption is not the hard part any more. Doing it well is.
None of this is the famous $42-per-dollar email ROI figure you’ll see recycled on every listicle. That number has no traceable primary source I’m willing to stand behind, so I’ve left it out and you should too. 😉
Where AI genuinely earns its keep
Subject line generation and testing, send-time optimisation, predictive segmentation, churn and win-back scoring, dynamic content blocks, and rapid-fire copy variants for A/B and multivariate tests.
These are the jobs AI does better than a human working alone, because they’re statistical problems dressed up as creative ones. A model can hold thousands of send-time patterns in its head at once. You can’t, and you shouldn’t try.
The graph below sets out where I let AI run the show and where I pull it back. Read it as a working rule, not a law of physics; your account might justify a different split.
Why I won’t let AI build production HTML
Here’s my line, and I’ll defend it. AI is terrific for a prototype, a first draft, or a batch of A/B variants you’ll throw away after the test. It is a poor choice as your day-to-day email builder, for three reasons that have nothing to do with taste.
Rendering is a minefield that AI hasn’t walked before. Litmus’s own technical guide to coding dark mode emails explains that Outlook desktop and Gmail apply their own colour inversion with no coding workaround available, while Apple Mail and iOS respect the prefers-color-scheme media query properly. Two clients, two completely different rulebooks, and that’s before you count Windows Mail, Outlook.com, and the half-dozen webmail clients that behave like none of the above. Litmus’s dark mode hub confirms over 40% of subscribers now open in dark mode, so this isn’t an edge case you can defer. A model generating HTML from a prompt has no memory of which client broke last time, because it isn’t testing against 90-odd rendering engines before it answers you. A dedicated builder like Stripo.email has already fought that war and baked the fixes into its output.
Regeneration erodes consistency. Ask a generative model to tweak the same template five times, and you’ll get five subtly different structures, not one template refined five times. Nested tables, inline styles, and MSO conditional comments are unforgiving of drift. A tiny structural change that looked harmless in the chat window can silently break your fallback fonts or your Outlook button padding on send four.
It’s an expensive way to do a solved job. Every time you ask a model to regenerate broadly similar HTML, you’re spending tokens and time reproducing work a purpose-built tool already does reliably. Stripo, BEE, and similar ESP-agnostic builders exist precisely because email HTML is a narrow, well-understood problem with known traps. Use AI to draft the copy and structure, then build and test the final piece in a tool designed for the job.
The web designer can’t do it either, and neither can AI
This is the part that annoys web designers, so I’ll say it plainly: a skilled website developer, however good at modern CSS and responsive layout, generally cannot hand you a bulletproof HTML email on the first attempt. Email code is still closer to 2005 web development than to 2026 web development. You’re writing table-based layouts, inlining every style, and adding VML fallbacks for Outlook because Microsoft’s desktop client renders email through the Word engine, not a browser engine. None of that transfers from building websites, and none of it transfers from a language model’s training data either, because both are pattern-matching against the wrong reference point.
What actually works is scar tissue. 🤕
Years of watching a campaign look perfect in the preview pane and then collapse in one client you forgot to test. That experience is what tells you to check the fallback font stack, to test the dark mode logo swap, to know which client will ignore your border-radius. AI has no scars. It has training data, and training data is not the same thing as having sent to a list of 200,000 people and watched the bounce report.
The real leverage: letting AI talk to your platform directly
The part of this story that’s actually new in 2026 is the Model Context Protocol (MCP), an open standard that lets an AI agent like Claude connect directly to a marketing platform’s data and tools, instead of you exporting a CSV and pasting it into a chat window.
Klaviyo shipped its own MCP server in 2025 and then expanded the partnership. Klaviyo’s announcement from 7 May 2026 confirms the connector now spans Claude.ai and Claude Cowork, letting brands pull campaign data, flow performance and customer profiles straight into Claude and turn that into performance reports, campaign briefs and marketing insights without rebuilding a dashboard.
HubSpot has done the same: its official connector for Claude lets you generate insights and charts from CRM and engagement data, and create or update records directly from a chat window, though it’s worth knowing the connector responds inside a conversation you start rather than firing automatically on events in your account.
Using Cowork instead of another dashboard
This is where agencies like mine get real leverage. Instead of exporting Klaviyo data into a spreadsheet and manually building a client report every Monday, Claude Cowork can pull the numbers through the MCP connector and produce the report, the anomaly flags and the next campaign brief in one pass.
Klaviyo’s own framing is that Claude shifts “from analyst who tells you what to do to a teammate who actually does the work”, and for reporting and first-draft production work, that’s a fair description. It still needs a human to check the numbers before they go to a client, but it removes the copy-paste layer that used to eat a Monday morning.
Governance you can’t skip
None of this happens in a legal vacuum. The EU AI Act’s transparency obligations, covered under Article 50, become enforceable on 2 August 2026, requiring disclosure when subscribers are interacting with AI-generated content (even though they’re extending the deadline for the more complicated ones).
If you’re feeding subscriber data into any AI tool through an MCP connector, check the scope you’re granting. HubSpot’s connector, for instance, explicitly withholds sensitive data properties and won’t expose engagement data when sensitive data settings are enabled on the account. Read the permission screen before you click approve, and don’t assume a connector is read-only unless the vendor says so in writing.
The wake-up call
To the sceptics: refusing AI in 2026, instead of protecting your craft, just makes you slower than the marketer next to you who’s using it for drafts and reporting while you’re still building segments by hand. The data shows AI-mature teams pulling further ahead on ROI every quarter this trend gets measured.
To the AI-reliant: the fact that you can now generate a plausible-looking email in thirty seconds doesn’t mean you’ve earned the judgement to know when that email will break. Fluency isn’t expertise. A tool that can write convincing HTML has no idea it just built you an Outlook rendering problem, because it has never had to explain a failed send to a client.
What to do this week
- Audit where AI already touches your email programme and sort each use into “draft or test” versus “production”, using the split above as a starting point.
- Move any recurring production HTML work out of a chat window and into a dedicated builder like Stripo, BEE or your ESP’s native editor.
- Check whether your ESP has a native MCP server or Claude connector, and connect it read-only first.
- Review the data scope on any AI connector against your GDPR obligations before you enable it for the whole team.
- Keep a human sign-off step before any AI-assisted send goes to your full list, no exceptions.
AI in email marketing isn’t a debate you win by picking a side, but rather a discipline you build by knowing exactly which parts of the job you hand over, and which parts you keep, because you’ve earned the scars that tell you the difference.
Leave a Reply