Open rates live and die in the inbox preview. A customer sees a sender name, forty-odd characters of subject line, and maybe a snippet of preview text — and decides, in about half a second, whether your email is worth a tap. Everything we automate downstream, the segment, the offer, the send time, only matters if that half-second goes our way.
So when we set out to let the autopilot write subject lines on its own, I was skeptical. Not that a model could string words together — that part is easy — but that it could write something a specific shop’s customers would actually open. Those are different problems, and we spent the first few weeks confusing them.
Our first attempt was the obvious one. We asked the model for “a catchy subject line for this campaign,” gave it the product and the offer, and let it go. What came back was exactly what you’d expect: exclamation marks, the occasional ALL CAPS word, an emoji or two, and a tone that sounded like every other promotional email in the folder. It was fluent. It was also generic, and generic doesn’t get opened.
The realisation that turned the project around was simple. The model wasn’t bad at language — it was bad at your customers. It had no idea what your shop sounds like, or which of last quarter’s subject lines people actually tapped. We were asking it to guess in a vacuum. The fix was to stop asking for “catchy” and start handing it the two things it was missing: your voice, and your open data.
What we changed
Every Reloadify shop already has a history of sends, and we already know which subject lines were opened and which were skipped. That’s the signal. Instead of prompting the model to be clever, we let it study what has worked for this shop specifically — the length people respond to, whether questions beat statements, how the audience reacts to a discount stated plainly versus dressed up.
We also fed it the shop’s tone of voice, the same profile the autopilot uses everywhere else. A homeware brand that writes in calm, full sentences should not suddenly start shouting in the inbox, and once the model could see the voice, it stopped trying to.
Now, for any campaign, the model drafts a handful of candidates and scores each one against everything it has learned about the shop. You see the options and the scores, not a single take-it-or-leave-it line.
Candidates for De Wasmand · "Spring linen, back in stock", ranked by predicted open rate for this audience.
That last row is the interesting one. We never wrote a rule that said “no exclamation marks.” We didn’t have to. When every candidate is scored against real open data, the line with three exclamation marks consistently lost — so the model learned, on its own, to stop writing them. Restraint won because restraint gets opened. It turns out your customers were telling us that all along; we just hadn’t been listening to the numbers.
Making it prove its work
A score on its own is easy to distrust, and rightly so. So before any line is allowed to ship, the model has to do two things: predict an open rate, and explain the reasoning behind its pick in plain language. If the explanation doesn’t hold up, you overrule it with one click, and that correction feeds back into what it learns.
This is the part that made the team comfortable handing it the keys. You’re not approving a black box — you’re approving an argument you can read.
This audience opens specific over superlative. "The colours that sold out first" names a real reason to look now, without urgency language. Removed "Don't miss out" — flagged as spam-adjacent and below this shop's recent open rate. Confidence: high.
If the model isn’t confident — a brand-new shop with no history, say, or a campaign unlike anything it’s seen — it says so, and leans on the tone profile instead of pretending it has data it doesn’t. We’d rather it tell you “I’m guessing here” than dress a guess up as a forecast.
I’ll be honest about the limits. The model is only as good as the open history behind it, so the first few campaigns for a new shop are more art than science. And a subject line can’t save a bad offer — it gets the email opened, not the product sold. We’re careful not to oversell what those forty characters can do.
It was never about teaching a model to be clever. It was about teaching it to listen to your customers, who’d been answering the question all along.
What’s different about doing this inside Reloadify is that the loop never leaves one place. The draft, the score, the reasoning, your edit, and the result the next morning all live together — so every campaign you approve makes the next one a little sharper. You’re not training a model in the abstract. You’re training the one that writes for your shop, on the only data that matters: whether your customers tapped.
We’re rolling subject-line drafting out to every plan over the next few weeks. If you want it sooner, it’s already live in beta — turn it on from the autopilot settings and let it learn from your last few sends.