← sources

What I learned building an AI assistant people would trust with their inbox

source: reddit (reddit:SaaS)published 2026-11-01 15:56ingested 2026-11-02 02:39

retained text

For the past weeks I've been building an AI assistant that reads my email, Slack, WhatsApp, LinkedIn and calendar and turns them into one to-do list, with the work already prepared: a reply drafted, the lookup done. I use it every day now. Some lessons that surprised me, in case they're useful for anyone building in the AI space:

\*\*1. "Drafts, never sends" beat every clever feature.\*\*

The first question everyone asked was "what if it sends something wrong?" So it never sends, pays or books anything by itself. It prepares, you approve. Trust went up more from that one rule than from anything smart I built.

\*\*2. The hard part is deciding what NOT to show.\*\*

Pulling in messages is easy. Deciding which 10 out of 300 actually need you is the product. Most of my time went into filtering noise, not generating text.

\*\*3. Promises are the killer to-do.\*\*

"I'll send you the deck tomorrow" buried in a thread is the thing you forget. Tracking what you promised and what others promised you turned out more useful than any summary.

\*\*4. Local-first answered the privacy question before it was asked.\*\*

It runs on your own computer, so your mail and chats never sit on someone else's server. For a tool that reads everything, that mattered more to people than I expected.

\*\*5. Let other AIs in, read-only.\*\*

People already live in Claude or ChatGPT. Letting those read your to-dos and contacts over MCP, without the ability to send anything, made it fit into how people already work instead of being yet another app.

\*\*6. Show what it learned, with an undo.\*\*

It learns from your edits, but a black box that changes its behavior is creepy. A page that lists what it learned, each with an undo, fixed that.

It's open source (AGPL), so if you want to see how any of this works under the hood: https://github.com/Salesflare/TheMachine

document fetched 2026-11-02 00:27 · canonical https://reddit.com/r/SaaS/comments/1xl6ker/what_i_learned_building_an_ai_assistant_people