AI that operates, not narrates
Most 'AI in your ESP' is a chatbox that explains your dashboards. That's not interesting. Here's the bar we're holding our assistant to.
Every email tool added a chatbox in 2024. Most of them are useless. You ask a question, you get a paragraph, you scroll back to the dashboard.
We think the bar for AI in lifecycle email is higher. The bar is: the AI should be able to take the action you'd otherwise take yourself.
If you ask Mailapp's assistant 'who's about to churn', it doesn't reply with a definition of churn. It opens a segment of contacts who haven't opened in 60+ days, sorts by lifetime value, and offers to draft a re-engagement automation.
If you ask 'why is open rate down', it doesn't link to the analytics page. It runs the cohort analysis, finds the campaigns that under-performed, identifies the common segment, drafts a hypothesis, and stages a fix for your review.
How we built it
There are three components. A planner that takes a goal and turns it into a sequence of tool calls. A toolbox that exposes every part of Mailapp as a typed function. An approver that diffs the planned changes against the current state and asks for your sign-off.
The planner is a Claude model with reasoning enabled. The toolbox is auto-generated from our REST API — every endpoint becomes a tool, with strict typing and docstrings.
The approver is the most important part. It's where the AI stops being a black box. Every action it wants to take is staged as a diff: 'I'd like to create segment X with these conditions, draft campaign Y with this body, enroll Z contacts into automation A.' You see the change. You approve, edit, or discard.
What we don't do
We don't ship features that look great in a demo but bad in a year. The first version of operate-mode could send campaigns autonomously. We pulled that. Real users want a strong approver step. The right product was the one with friction in the right place.
We don't train shared models on customer prompts or content. The default assistant runs against an Anthropic endpoint with no training carve-out. Enterprise customers can run it against their own local model.
We don't pretend the AI is always right. Every action has an undo. Every undo is one click. Every approved action is logged with the input, the output, and the approver.
The next step
We're working on a multi-step autonomous mode where the AI can chain actions under a policy you set. Think: 'every Monday morning, identify the campaigns with statistical-significance-positive A/B results from last week, promote them, and report back.'
We'll ship it when it's safe. Until then, every action goes through approval. We'd rather be slow than wrong.