The problem
The owner was spending the first ninety minutes of most days in email, and the last hour of most evenings too. Something like 120 emails a day: client questions, supplier invoices, booking changes, newsletters, automated notifications. Maybe ten or fifteen of them actually needed the owner.
The volume wasn't really the problem. The problem was that the important emails looked exactly like the unimportant ones until you opened them. So things slipped. A client had to chase a document twice. An invoice query sat for a week.
They'd already tried filters and labels. Filters work on who sent an email, not what it says, so they broke the moment a client wrote from a new address or a supplier changed their billing system.
Week one: sorting by hand, together
I didn't touch any AI for the first few days. I sat with the owner and we went through two weeks of their email together, putting each message in a pile: needs you, routine reply, file it, or ignore. For every one I asked why, and wrote the answer down in their words.
"Anything from a client that mentions a deadline comes to me." "Supplier invoices go into Xero. I only need to see them if they're over £500." By the end of the week we had about twenty rules in plain English, and a long list of real examples for each.
That document became the brief. It's also the part most off-the-shelf AI inbox tools skip. They guess what's important from patterns across millions of inboxes, not from how this particular business works.
The first version: sort and draft, never send
The assistant connects to the owner's Gmail. Every new email is read by a language model with the rules and examples in front of it, and gets one of four labels: Needs you, Draft ready, Filed or Ignore. Anything it isn't confident about goes to Needs you, with a one-line note saying why it wasn't sure.
For the routine emails it writes a draft reply. To get the tone right, it learned from a few hundred of the owner's own sent emails: short, friendly, no "I hope this email finds you well". The drafts sit in Gmail's normal drafts folder, so there was nothing new to learn. Open it, tweak it if needed, press send.
For the first two weeks we went through the labels together every morning. When it got one wrong, we fixed the rule rather than the AI. Usually the rule was ambiguous, and a new member of staff would have been confused too. By week three, the owner was changing about one draft in ten.
How it grew
Once the sorting was trusted, the next requests came on their own, one at a time.
- Chasing. "If a client hasn't sent something I asked for in three days, nudge them." It now keeps track of open questions and drafts a polite chaser on day three, which the owner approves with one click.
- The diary. Booking requests and "can we move Thursday?" emails are checked against Google Calendar, so the draft reply offers times that are actually free. Once the client confirms, it adds the event.
- Invoices. Supplier invoices go into Xero automatically. When a client asks "have you had my payment?", the draft reply already has the answer from Xero.
- The Monday briefing. Every Monday at 7:30 the owner gets one message: what needs them this week, what's overdue, what's double-booked, and what was handled last week.
Each of these went live in draft first. Only after a few weeks of getting things right did the owner let it act on its own, and only for low-stakes things like filing and calendar holds. Anything going to a client still waits for a yes.
That's when we started calling it a chief of staff. Not because it's especially clever, but because it does what a good one does: keeps track, prepares things, and only brings you what needs you.
What changed
Email went from over two hours a day to around thirty minutes, most of it approving drafts. Nothing important has slipped through since the first month. And the owner says the biggest difference isn't the time. It's no longer carrying the inbox around in their head all evening.
What I'd tell you
- Start with the rules, not the tool. The week spent writing twenty rules is what made everything else work.
- Keep it in draft until it has earned trust. It's slower for a fortnight and much faster after that.
- Don't build it all on day one. Every piece here was the owner's idea, once the last piece had worked.