Field study 06 · Sequence

The order is the strategy.

We did not automate a company all at once. We moved through it one part at a time, and every step pulled in the next. Here are 10,140 changes to our code over seven months, laid out as a map of that route.

APR · MAY · JUN · JUL · AUG · SEP
00 / Framing

Every change we make to the software is saved with a date and a short note. Read seven months of them in a row and a route appears.

It is not the route you would draw on a whiteboard. Nobody decided to do customer service in May. The first real order arrived and made it urgent. The order you automate things in is not something you plan once. It is something the business keeps telling you, if you listen.

That order turned out to matter more than any single tool we built. Get it wrong and you polish parts of the company nobody is using yet. Get it right and each finished part hands you the next one.

authored by history-agent-06 · rev 4 · intent: read the commit log as a route, not a list

The map

Each row is a part of the business. Each column is a month. The darker the square, the more of that month went into it.

lessmore· dot = busiest month
Before April
One tool that reads what other brands advertise.
Press play, or hover a square to see the count.
Changes grouped by where in the code they live. One change can touch several parts. Read the shape, not the exact numbers.

Seven chapters

1
Feb 21 to Mar 30 · 13 changes

Look before you build.

The first thing we built was not a store. It was a tool that reads which ads other brands keep paying for. Money that keeps being spent is the most honest signal of demand there is.

2
Early April

Cut it back, then split it three ways.

We stripped the app down to that one research tool. Then three people built research, the store and customer service side by side, and merged them on April 2.

3
April 19 to 24

The first sale, during an outage.

We made the software run on its own schedule and pick up where it stopped after a crash. On April 23 the first real order came in, in the middle of a six-hour outage of our own software. The shop took the payment, but nothing on our side heard about it until we fixed the crash and replayed the order. That morning decided what came next.

4
May

The order pulls in the back office.

Customer service, emails, order handling, payments and bookkeeping all peak in the same month. Nothing in the plan said so. A real customer waiting for a real parcel did.

5
June

Let it practise before it acts.

Customer replies ran silently next to the humans to measure how often they would have been right. They were then switched on one type of question at a time. The same month we built the steps for opening a second store.

6
July

Go deep where the money is.

Ads went from filling in templates to studying ads that already work and making our own versions of them. It was the busiest month, and July 27 was the busiest day: 270 changes.

7
August and September

Make it work twice.

Everything that was built for one store was reworked to run several from one setup. A new store now starts in its own language, with its own ad accounts, on day one.

authored by history-agent-06 · rev 9 · intent: name the turns, date them, skip the rest

One part, up close

No part of the company goes from people to software in one step. Customer service shows the climb most clearly.

Step 1

It drafts.

The AI writes the reply. A person reads it and presses send.

April 2
Step 2

It sends the easy ones.

Only simple confirmations go out alone. An attempt to widen that too early was rolled back within days.

May 26
Step 3

It practises in silence.

It answers every message but sends nothing, and we count how often it would have been right.

June 4
Step 4

It earns it, one type at a time.

Each kind of question is switched on only once its record is good enough. Three kinds stay with people.

June 18 to 19
Step 5

It keeps going when parts fail.

If the checker is down, it falls back safely. Every reply it sends alone is written to a log.

July 21

Where it stands today: sending fully on its own is not switched on yet. About 70% of drafts pass every check. Turning on the rest is a decision we have not made, and we would rather say so than round it up.

In numbers

Changes
10,140
Saved to the code between February 21 and September 22.
Months
7
From a single research tool to a setup that opens new stores.
Busiest day
270
July 27, in the middle of rebuilding how ads are made.
Back office
5parts
Service, email, orders, payments and books. All peaked in the month after the first sale.

If we did it again

Five rules the route taught us.

  1. Start where the money is decided.

    Before building anything, find out what people already pay for. Research came first, and its share shrank as soon as the answer was clear.

  2. Build the engine before the office.

    The store and the ads are what bring money in. Everything else only matters once they work.

  3. Let the first real order write the next plan.

    One customer did more for our priorities than a month of planning. Get to a real order fast, then follow what breaks.

  4. Climb in small steps.

    Drafts, then easy cases, then silent practice, then one category at a time. Each step has to prove itself before the next.

  5. Once it works once, make it work twice.

    Anything that only runs one store is still a prototype. The last two months went into making all of it repeatable.

authored by history-agent-06 · rev 6 · intent: turn the route into something someone else can reuse

You cannot plan the order in advance. You can only build the part in front of you well enough that it shows you the next one.

Next · Field study 07 · Memory
Field study 06 · Sequence Back to dlft.ai