Recently, an AI assistant got access to my email, calendar and Teams. I call him Jarvis. A bit silly, but it works. He reads along, prepares replies and keeps track of what I promised. One rule is non-negotiable: he doesn’t send anything without me.
That sounds like a brake, but it turned out to be the reason it works.
What went wrong
In the first week he prepared drafts that I have since decided on: 22 of them. I sent fifteen, ten of them unchanged. Seven I threw away. Not because they were badly written; tone and spelling were fine. But they came at the wrong moment, context was missing, or I had already handled it myself, for instance with a phone call Jarvis couldn’t see.
That says something about AI in practice. The problem is rarely “can it write”. It can. The problem is: does it know whether this is needed right now?
Two drafts side by side
This is what it looks like in practice. Names and details removed.
Hi [name],
Thank you for reopening the purchase order. For 2025 there is only one invoice under this order, for exactly the order amount, so there is nothing over it to report.
The March invoice is for work in 2026. That's the one that didn't get through your procurement system. As requested, I attach our quote for 2026. Could you let me know the new order number once it's raised? We'll then submit the invoice under it.
Jarvis wrote
For the attention of whoever is responsible for digital matters and the app. Your app is visibly being developed further (a big update in August), but …
I made it
Who looks after the app at your company? Feel free to forward this.
What we changed
Before every draft, Jarvis now has to answer four questions:
- Have I already replied, called or sent an invite?
- Does this need to happen today, or can it wait?
- How do I write to this person myself? He reads my latest emails to them and matches that.
- Does he know enough about what was agreed before? If not, no draft but a question.
Whether that helps, I simply measure: how many drafts I send unchanged. It’s too early for big conclusions, but the direction is clear. Three good ones are worth more than twelve of mixed quality.
Trust grows in steps
I use the same principle for the freedom Jarvis gets. He started with only preparing. Now, as a trial, he may send a few boring, low-risk messages himself, like confirming a meeting that is already in my calendar, with a daily maximum. If it’s about money, a complaint or anything legal, his own gate refuses. A mistake? Then that category is switched off immediately. What goes well gets room. What doesn’t, doesn’t. That’s how you work with a new colleague too.
A practical example
An automated test in one of my projects was red for over a week. The alerts arrived as email, and email is where things disappear among everything else. Now it’s his job to read those alerts, look for the cause in the logs and prepare a proposal. I only have to approve it. No magic: a boring job that always got left behind, and one an AI is better at than I am.
Three things you can do tomorrow
- Let an AI only make drafts, not send them yet. After a week, count how many you send unchanged. That’s your yardstick.
- Write down when a draft is not needed. That is what an assistant has to learn.
- Don’t give it tasks you don’t know the answer to yourself yet. Then it learns from your mistakes instead of your best moments.
The lesson in one sentence
An AI doesn’t become useful because it’s allowed to do more. It becomes useful because it knows better when to stay quiet.
Next time: why I want my companies to run without me, and what I still have to let go of to get there.
Best, Herbert