running on ai · story 01

We built an AI email assistant for a bus hire company. It saves 2,000 hours a year.

july 2026 · 6 min read · by lathe labs

Rentabus and Busify hire out buses. Between them, about a hundred enquiry emails arrive every day, and for years each one cost roughly five minutes to answer: check the calendar, check the rates, remember what this customer booked last time, write the thing.

Five minutes sounds like nothing. Multiply it out. A hundred replies a day, two hundred and fifty working days a year, and you land north of 2,000 hours. That is a full time job that produces no buses and no bookings, just sentences the business has already written thousands of times.

Worse, the clock was always running. In quoting businesses the fastest answer usually wins, whether it is yours or not. An enquiry that lands at 9pm and gets answered at 9am has often already booked with someone else. Nobody ever agreed to compete on reply speed. Everyone does anyway.

What we built

The assistant lives inside the Gmail accounts the businesses already ran on. That constraint came before any AI did: no new software, nothing for the team to learn, because tools that need a training course die quietly in small companies. If you can read a draft, you can use this.

Before it wrote anything, it read. All 39,188 email conversations the two companies had ever had were sitting right there in the sent folder, which is the part people forget: a small business’s best training data is its own outbox. The tone. The standard answers. The pricing logic. Which questions you ask a wedding party (how many, how late, how loud) and which you ask a school trip.

Now when an enquiry lands, a draft is already waiting. Personal, in the business’s own voice, with the details of any previous booking pulled up without anyone digging through old threads. It also scores each new deal against the hundreds that came before it, so the owner can see at a glance which enquiries look ready to book and which have gone quiet. The evening goes to the right ones.

The rules it lives by

This is the section we would want to read first if someone else had written this.

Every reply waits for a human sign-off. Nothing sends on its own. It never sets or changes prices. Complaints go straight to a person, untouched. It ignores instructions hidden inside incoming emails. There is an off switch, and no lock-in.

AI drafts. People approve. Five minutes of writing becomes twenty seconds of reading, a hundred times a day.

The numbers

2,000h+a year back, across ~100 emails a day at five minutes each
12 minfrom enquiry to a considered draft, at any hour. It used to take four hours.
+15%more enquiries won, because the fast reply usually wins the booking
measured in daily use, not projected · 100 × 5 min × 250 working days ≈ 2,083 hours

The hours get the headline, but the 12 minutes is the money. Speed was never about efficiency for its own sake. It is the difference between quoting first and quoting second, and the 15% more enquiries won is what quoting first turned out to be worth.

What we deliberately didn’t automate

The send button was the argument we kept having. Letting the assistant send would save another few seconds per email, and plenty of vendors happily do it. We didn’t, and we won’t. One wrong price or one badly handled complaint costs more trust than a year of saved seconds ever buys back.

So the assistant took the part that was eating the evenings, the writing, and the judgement stayed where it belongs. Is that the future of small business email? Honestly, we don’t know. It is the present of two bus companies, and the maths worked.

Could your inbox do the same?

If your week has a hundred small replies in it, the arithmetic above probably applies to you. The free audit takes twenty minutes and puts a number on it.

Get a free audit →
running something clever on ai yourself? we want that story · hello@lathelabs.ai