My AI Assistant Has No App. It Lives in My Inbox
The address that received it is not a person. It is my AI assistant. No app, no chat window, just an email address.
The first line of every reply can only be one of three things: done, not done, or you need to decide.
Most of the time I have spent on it went not into teaching it what to do, but into deciding what it may not do. It can draft; the send is mine.
I used to leave the house with a laptop. Now the mail app on my phone is enough.
Saturday morning, on the DLR.
Three minutes earlier, an email had landed on my phone: the three new people in the client’s Operations department were on the Team page, as I had asked. I opened the screenshot. Operations was now six people, and the whole row had been squeezed onto two lines.
I replied in the same thread: “See how to fix it, send me screenshots.” Phone back in my pocket.
Two stops later, the reply came. First line: “Operations is now a single column. The change is on a branch, not merged, not released.” Two screenshots attached, before and after. Last line: “Reply ‘merge and release’ to put it live.”
I replied: merge.
The address that received it is not a person. It is my AI assistant. No app, no chat window, just an email address.
I build AI systems for companies in London. This year I added an alias to my Gmail and wired it to the AI I already subscribe to. Every email I send to that address is a task.
Here are some of the other emails I have forwarded to it.
A partner sent over a Google Slides deck with five pages left blank for us to fill. It filled the five pages, matching the original template exactly. I said make it Chinese; twenty minutes later the second version came back, two of the pages re-laid out to match the ones she had already filled.
Checking my insurance policy. I asked whether a physiotherapy clinic would be covered. The reply: you can go, it is £70 a session, the insurance pays at most £20, you pay £50.
An email from a client. I forwarded it straight to the assistant and told it to get it done.
I asked for a list of target clients. It built a Google Sheet: 15 UK-based clients, ranked by priority, each with a note on where to start.
Scheduled: before a meeting, it reminds me with the script and the key points.
In an inbox, one thread is one piece of work. Every message and reply stays together, and a month later you can still see how it started. The attachments are the deliverable: screenshots, reports, files, videos. On the train, in the middle of a meeting, awake at midnight, I do the same thing every time: write a few lines and send.
Anything I can log into from a computer, it can use. Mail, Sheets, Docs, Drive, Slides and other tools: it reads and writes them all. It also edits code, runs tests, and releases to clients’ production systems.
In the last six days it has replied to 346 of my emails.
The first line of every reply can only be one of three things: done, not done, or you need to decide.
Of those 346, 221 began with “done” and 96 with “you need to decide”.
Most of the time I have spent on it went not into teaching it what to do, but into deciding what it may not do.
Code changes stay on its own branch; merging into the main branch needs a word from me. So does a release.
Emailing a client unprompted needs a word from me. It can draft; the send is mine.
I also opened a second address for my colleagues. It is online around the clock and answers their questions about project progress.
It has made mistakes too.
Last night we were discussing a clock-in process, and it gave an example of what would happen if a site manager forgot to clock out. In its next email it wrote on its own: “That example in my last email was an extreme case I made up. I should not have let you take it for the normal flow.”
But I don’t use it because it never gets things wrong. I use it because every time it tells me how far it got, and when it is wrong it says so. That is no different from working with a new colleague.
I used to leave the house with a laptop. Now the mail app on my phone is enough.
Getting there takes one step: connect the AI you already pay for to your inbox.
The train pulls in. Its reply is on my phone: done.