Opinions
Product Updates

Every ESP Has The Plumbing. So What’s Running Through It?
Every major Email Service Provider has opened its plumbing to AI agents. So how smart are they? And what happens when nobody even opens an ESP anymore?
A marketer types "build my BFCM promo for lapsed buyers" into Claude and a campaign draft appears in Klaviyo without anyone actually opening Klaviyo. There’s no logging in or dragging-and-dropping blocks.
That experience was a pipedream a year ago, but it’s now the default trajectory for every major email platform. And the timeline making that pipedream a reality has moved very quickly:
Backstroke rolled out our own MCP server in June.
Salesforce made its Marketing Cloud Engagement MCP server generally available that same month.
Braze followed in September.
Then Klaviyo announced Headless: 260+ MCP tools and nearly 500 APIs with the promise that an agent can "read, write and go live… without anyone opening Klaviyo at all."
“Headless” means the interface—aka the head—aka the person—is gone. The ESP still does everything it did, but through APIs instead of a dashboard, so AI agents can build and send campaigns without anyone ever logging in.
Salesforce’s own CTO described the concept bluntly by asking, "Why should you ever log into Salesforce again?" He later clarified, "Headless doesn't mean no UI; it means decoupling capabilities from the interface."
Fair enough. But if the interface is decoupled, why would you even pay for an ESP to begin with?
The Next Battleground is Performance
Headless access is good for marketers. The ability to build, test and launch on-brand email campaigns from inside an AI assistant, without context-switching into a separate platform, is a genuine workflow improvement. However, it also changes the competitive math.
If every ESP speaks the same language of only crafting messages and scheduling sends, the differentiators between platforms will start to fade away. Performance provided by your agents will matter most. And those agents don’t care whose API they call, but whether or not each campaign lands.
The question marketers will face isn’t “which ESP should we use?” but “did my work get done and what were its results?” As SoftwareReviews said, "Opening the interface does not eliminate Klaviyo's moat. However, headless access also makes individual Klaviyo functions easier to combine with or substitute for capabilities from other vendors."
Both halves of that sentence are important.
Funny enough, Gartner just published its first Magic Quadrant for Email Marketing Platforms at this exact moment when being judged for interface starts to matter less. But that future isn’t entirely here yet. Among martech leaders already using AI agents, 45% told Gartner that vendor-offered agents fail to meet their expectations.

Today’s opportunity is to close that gap, and not with a better interface alone, but with agents that actually deliver. As the interface becomes less central, the decision-making layer on top becomes more important to evaluate.
Faster Isn’t the Same as Smarter
Agents aren't running the show yet. In fact, only 11% of marketers have autonomous agents in their workflows, with another 24% still experimenting with them. Salesforce's State of Marketing survey polled about 4,450 marketers across 26 countries to find that only about 13% of marketers are using agentic AI, in general.
Power users exist, of course. But examples such as custom dashboards and weekly performance recaps posted to Slack show the difference between automating reporting and delegating decisions. These are useful tasks, but not the same as deciding who gets what email content, when and why.
Faster processes are only half the battle though. Better outcomes are the other half. And there are still basic workflow gaps to close. For example, the 350k+ campaigns sent through Selzy this year, only about 1.65% targeted segmented audiences. This is a prime example of where agents need to do more than speed up the existing process.
After all, a faster route to an unpersonalized blast still results in an unpersonalized blast.
Connector MCPs vs. Decision MCPs
Some tools let an agent take action, while others help it choose the right action.
A Connector MCP lets an agent do what a human could do in the UI. It can create a segment, pull a report and schedule a send. This is that important plumbing we mentioned earlier, giving an agent access to your ESP.
A Decision MCP tells the agent what will work, based on outcome data the LLM doesn't have. Which subject line will your aged 60+ buyers actually click? Which hero image is more likely to resonate with lapsed customers versus loyalists?
Generating options is one task, but ranking them using historical performance is a whole other ballgame. Backstroke’s MCP server, for example, returns creative rankings and audience-level predictions based on historical campaign outcomes in its proprietary dataset.

So let’s look at where the major platforms are on both of these fronts:
Klaviyo's engineering blog describes its Personalization models candidly. It says that each one "handles a single decision.” Coordination across those models is, in Klaviyo's own words, "further out."
Salesforce says that their platform provides "context and capability" while the LLM provides "intelligence." That explains the division of labor, but leaves marketers questioning the evidence to support the creative choices it recommends.
Bloomreach is pitching Loomi Connect as more than a way for AI agents to access its software, but also as a way for them to tap into their customer, catalog and commerce knowledge. That is a broader claim than simple connectivity, though it doesn’t by itself demonstrate audience-specific creative prediction.
The industry is converging on the problem. Whoever can explain which creative is more likely to work for each audience, what evidence supports that prediction and how those predictions hold up against actual results is the victor.
What a Prediction Layer Looks Like in Practice

You say the thing. We make it predictive. Here's how that works with Backstroke's headless MCP.
A marketer can Claude (or ChatGPT, or Gemini, whatever your preference) and say, "build a campaign for this week's Friday serum sale, targeting engaged buyers, lapsed customers and cart abandoners."
Backstroke's MCP server connects the assistant straight to its prediction models. It builds the email, generates several hero options and returns per-segment predictions. Instead of delivering one winner for the whole list, it determines which hero and subject line each audience segment is predicted to click, with lift estimates and per-cohort breakdowns.
The marketer reviews and approves the email concepts.
Backstroke publishes winners directly to their ESP.
Hero 2 is the engine underneath it all, distinguishing roughly 28,350 audience segments, up from about 300 in the prior model. That's the difference between "women 25–34" and predictions granular enough to separate a 24-year-old in San Diego from a 36-year-old in Atlanta. (Because those two audiences definitely don't click the same things.)

We have the data to back it up, too. Backstroke recently ran an analysis of ~13,500 emails across 35 brands, surfacing the kind of patterns no LLM can reason its way into. Some of these unique findings were that:
Urgency language in subject lines correlates with a 10% decrease in clicks,
Subscribers over 60 click twice as often as those under 30 on the same email.
This outcome data tracks with results. ThirdLove, running on Klaviyo with Backstroke's predictions, saw a 25% lift in revenue per recipient and a 23% increase in click rate. Cozy Earth saw a 48% jump in click rate.
Is Your Stack Ready to Go Headless? 5 Questions to Ask.
Headless email is here and the login screen is increasingly optional. To evaluate your options, take these questions into your next vendor conversation:
Plumbing or prediction? When an agent calls your ESP, does it get back raw data or a ranked recommendation with expected lift?
What's the prediction trained on? Real click and purchase outcomes across thousands of brands and audiences? Or an LLM's general sense of what sounds good?
How fine-grained are the decisions? Does it predict a single creative for the whole list, or different creative predictions for each of your audience segments?
Are the decisions coordinated? Does all your information come from one system or from separate models that don't talk to each other?
Who approves? And can you prove it? Does a human sign off before anything sends? And can you run a holdout to prove lift against a control? With 61% of marketers facing C-suite pressure to show measurable AI gains, this question is critical.
Your agent needs evidence to guide your winning decisions. Request access to the Backstroke MCP to compare subject lines, hero images and email copy using predictions grounded in real campaign outcomes. Connect your audience for segment-level recommendations, then keep the final sign-off before anything goes live.
The right headless stack will help you make better decisions and execute on them faster.


