Every time we change our software, it gets saved with a date. Over eight months that added up to 12,381 changes. Line them up and you can see where our time went.
It isn't the plan we would have drawn up. Nobody decided to work on customer service in May. Then the first real customer ordered something, and suddenly it couldn't wait. You don't pick the order once. The business keeps telling you what's next, if you pay attention.
Getting that order right mattered more than any single tool we built. Get it wrong and you polish parts of the company nobody uses yet. Get it right and every finished piece points you to the next one.
The map
Each row is a part of the business. Each column is a month. The darker the square, the more of our time that part got that month.
Eight chapters
Look before you build.
The first thing we built wasn't a store. It was a tool that watches which ads other brands keep paying for. If a brand keeps paying for an ad, it's selling something. That's the most honest sign of what people actually want.
Strip it down, then split up.
We cut everything back to that one tool. Then the three of us split up: one on research, one on the store, one on customer service. On April 2 we put the pieces together.
The first sale came in while we were down.
We had just set the software up to run on its own around the clock, and to pick up where it left off after a crash. On April 23 our first real order arrived right in the middle of a six-hour crash. The customer paid, but our side never noticed. Once we had things running again, we went back and picked the order up. That morning decided what we built next.
The first order drags in everything else.
Customer service, emails, handling orders, payments and bookkeeping all got the most work in the same month. That wasn't the plan. A real customer waiting for a real parcel was.
Let it practise before it acts.
The AI answered customer messages quietly in the background, without sending anything, so we could see how often it would have got them right. Then we let it answer for real, one kind of question at a time. That same month we worked out the steps for opening a second store.
Go deep where the money is.
Our ads used to be templates we filled in. Now the software studies ads that are already working and makes its own versions. It was our busiest month. July 27 alone had 270 changes.
Make it work twice.
Everything had been built for one store. We reworked it so one setup can run several. A new store now starts in its own language, with its own ad accounts, from day one. Then the new brands needed ads, and September became our biggest month ever for buying them.
The first order, again.
Our newest brands started getting orders, and just like in May, orders, email and customer service got busy again. Then on October 9 the platform we sell on froze one brand's payments for an identity check. Its own systems said everything was fine, so the ads kept running for hours while nobody could pay. That same day we taught the software to open each checkout every hour, the way a customer would, and to pause the ads by itself if payment is blocked. Within six hours that brand was taking payments another way. By October 19 its customers were checking out on a page we built ourselves.
One part, up close
No job goes from people to software in one jump. Customer service shows the steps best.
It writes, we send.
The AI writes the reply. A person reads it and hits send.
It sends the easy ones.
Only simple ones, like confirming an order, go out on their own. We tried giving it more too early and took it back within days.
It practises quietly.
It answers every message but sends nothing. We just count how often it would have been right.
It earns it, one type at a time.
Each kind of question gets switched on only once it has a good enough track record. Three kinds stay with people.
It stays careful when things break.
If one of its safety checks stops working, it plays it safe instead of guessing. Every reply it sends on its own gets written down.
Where it stands today: it doesn't send everything on its own yet. About 70% of its replies pass every check. Letting it handle the rest is a call we haven't made, and we'd rather tell you that than round it up.
In numbers
If we did it again
Six things the route taught us.
Start with what people already buy.
Before building anything, find out what people already pay for. We did that first, and spent less time on it as soon as the answer was clear.
Build what earns money first.
The store and the ads bring money in. Everything else only matters once those work.
Let the first real order set the plan.
One customer did more for our priorities than a month of planning. Get to a real order fast, then fix whatever breaks.
Hand over in small steps.
First it writes and you send, then it handles the easy cases, then it practises quietly, then it takes over one kind of task at a time. Each step has to prove itself first.
Once it works once, make it work twice.
If it can only run one store, it's still an experiment. We spent August and September making all of it repeatable.
Check what the customer sees.
On October 9 the platform we sell on said everything was fine, but nobody could pay. Now the software opens each checkout every hour, the way a customer would. Checks that only look at your own side miss the moment a customer gets stuck.
You can't plan the order ahead of time. You can only do the piece in front of you well enough that it shows you the next one.
Next · Field study 07 · Memory