Raian Pollock

The system that did the email marketer's job

Flip Education

August 2026 · Raian Pollock

An estimated 36 person-years of work, built by two people and AI agents in two months

Every company that sells software sends this kind of email. A welcome note when you sign up, a warning when the free trial is nearly gone. Normally an email marketer owns it and spends a decent part of their week deciding who hears what.

Flip Education had two people and no email marketer. So a system did that job, and the next section shows how it worked.

Company
Flip Education. Lesson plans for teachers, 18 countries, 11 languages
Team
Two people, no email marketer
Period measured
March to August 2026
Emails sent
1,130, to 452 teachers

What it knew before it sent anything

Before any email went out, the thing looked at the teacher it was about to write to. It knew how many lessons they had ever built, and how many days since the last one. It knew which teaching methods they had tried and which they had never touched. It knew how many free lessons they had left, whether they had finished setting up, and whether they had started to buy something and walked away halfway.

It knew whether they were part of a school that had bought in bulk. Those teachers never got a single email asking them to pay for anything. Their school had already paid. Asking again would have been noise.

An internal review page showing one email in the set. Above the email it lists the subject line, what causes it to send, and the sending window of 8 to 10 in the morning in the reader's own time. Below that sits the finished email as a teacher would receive it.
How I reviewed them. Every email in the set sits on one page with the thing that causes it to send printed above it, so the whole set can be read in one sitting and in any of the languages. The email below the label is what a teacher actually gets.

Three ways an email could go out

The differences between the three matter more than anything else here.

A person is doing something Uses a lesson credit the moment it happens Warning goes out 272 sent Lesson finishes building tab open, or closed? Sent, or cancelled 276 sent called off if they are still watching Nobody is waiting Every hour, who is due? 8 to 10am, their time Whatever it decides 15 kinds, 582 sent A person is doing something Uses a lesson credit the moment it happens Warning goes out 272 sent Lesson finishes building tab open, or closed? Sent, or cancelled 276 sent called off if they are still watching Nobody is waiting Every hour, who is due? 8 to 10am, their time Whatever it decides 15 kinds, 582 sent
The top two go out as part of something the teacher is doing, so a person is waiting on the other end. The bottom one runs by itself with nobody waiting. That difference is the only thing in this diagram that ended up mattering.

The first way was instant. When a teacher used up nearly all their free lessons, the warning went out inside the same moment as the lesson they had just made. Not that night, not in the next morning's batch. Right then, as part of the same action. The welcome works the same way, going out while the account is still being created.

The second way waited for something to finish, then checked whether anyone was still there. Lessons take a while to build. If the teacher had closed the tab and walked off, they got an email when it was ready. If they were still sitting there watching it finish, the email was called off before it ever went out, because the thing checked whether they were looking at the screen at the exact second the lesson completed and stood down if they were.

That one is the reason any of this was worth building. A bought email tool sits outside your product, so it only knows what your product tells it. Whether the teacher is sitting right there watching the lesson finish stays in the browser unless you build the tracking to send it.

The third way ran on a schedule, going through everyone once an hour and asking who should hear from us today. It only sent between eight and ten in the morning in the teacher's own time. Across 18 countries that meant it worked around the clock. It wouldn't send two too close together. Each kind could only ever reach a given person once, ever. When it suggested a teaching method to try, it left out the ones they had already used.

What didn't work

Two of the emails never sent once in the system's whole life. One was meant to arrive after a teacher printed a lesson and carried it into class, which is the single moment in the whole relationship when you most want to ask how it went and stand the best chance of a real answer. Nobody ever recorded the printing. It sat there ready for months with nothing to fire on.

Then the whole scheduled layer sent nothing for 62 days, and reported itself fine every hour of them.

The instant ones kept going the whole time, which is why nobody noticed. Mail kept arriving. An email that goes out as part of something a person is doing breaks in a way you hear about within the hour, because somebody is standing at the other end of it waiting. One that runs quietly on its own breaks quietly.

Which emails need the product's own data

Most of these emails didn't need any of this. A welcome. A nudge when somebody goes quiet. You can buy those off the shelf tomorrow and they will work about as well as the ones I wrote by hand.

Three or four weren't worth buying a tool for, because they depended on something only the product knew at that exact second. Whether the teacher was watching the screen. Whether they had just spent their last credit. That information never leaves the building.

So the useful question is which of your emails are which. Most are ordinary. Buy those and forget about them. A handful depend on something only your product knows, and those earn the work. My mistake was taking two months to learn which was which, and learning it from a silence.