Multichannel AI assistant in production
The assistant reads email, calendar and documents for you. It suggests, you approve, it acts. In service since 2025.
- Around the clock: it keeps working when you're away from the screen
- Nothing sensitive happens without your approval
Before and after
The starting point
Useful information is rarely where you work. A request comes in by message. The detail is in an email. The date is in the calendar. The follow-up ends up in a spreadsheet.
None of this is hard. But someone spends an hour a day moving that information around.
Three things came back every week:
- Everything had to be explained again. No tool remembered the previous day’s conversation.
- Tasks slipped. A follow-up to send, something to monitor, and nobody to think of it at the right time.
- Failures went unnoticed. When a login expired, the tool answered "nothing found" instead of saying it was broken.
What was put in place
One single conversation
First, one single conversation. The assistant is reachable from the everyday messaging app. It keeps the thread. Three days later you pick up where you left off.
It actually acts
Then, it actually acts. It reads email, checks the calendar, opens documents, pushes code, drives a browser. Every sensitive action stops at an approval step. It suggests, you confirm, it acts.
It watches itself
Finally, it watches itself. It checks on its own that its access works. If it goes down, a watchdog restarts it. It logs its decisions in a journal and rereads it so it does not make the same mistake twice.
You stay in controlEvery sensitive action stops at an approval step. It suggests, you confirm, it acts.
What it taught us
The hard part was not the quality of the answers. That comes quickly. The real work is elsewhere: restarting after a failure, saying "I am broken" rather than "I found nothing", and never acting alone on what matters.
How it works
