How we use AI

The agents are AI. We always say so.

Who does what, what you decide, what we still have to write down.

Sumipost works with a team of AI agents: one writes, one critiques, one checks the finished email. Here is what each one does.

Plainly

The team of agents

Who writes. The agent that builds the email in MJML with its own tools, from your brief and the materials you give it, inside the brand's rules.
Who critiques. A second agent, on a different model, reviews the email against the brief and the brand's rules and asks for fixes before it reaches you.
Who checks. The visual check photographs the finished email and looks for serious defects: overlapping text, cropped photos, broken images. If the email rebuilds a screenshot, it compares it with the original.
Who retouches images. An assistant that, when you ask, changes a photo in the image editor.

Colours, typefaces and button shapes are measured by the program on the client's page, without AI. A model only steps in for two things: deriving the brand's voice from the site's copy and, when there are several candidates, choosing the logo by looking at the page.

The brand rules: what is locked and what is free. The AI only writes in the free zones.
Every change, in words or by hand in the editor.
When an email is ready: the client sees it first, through review, and approves it.

We give them a fact sheet for the brand and ask them not to invent news, prices or addresses. They can still get things wrong: that's why every email goes through you and the client before it leaves.

The site's photos are AI-generated, and we say so next to each one: the portraits are roles, not real people.
The photos of the sample brands and of Brezza are AI-generated too. The screens, on the other hand, come from the real app.
How and when we made them is in the Colophon.

Which model providers we use, where data is processed and how we mark generated content: we'll write it here and in the privacy notice, which we're preparing with our lawyer, before launch.