Our AI reads every company email and writes the weekly recap.
Once a week the six of us at La Nao get one email. It says what the company worked on last week, who moved something important forward, the two biggest problems still open, what we should focus on next, and every promise anyone made to anyone. Nobody wrote it. Nobody filled in a status form either.
A model wrote it, from every email the company sent and received that week plus our task spreadsheet. It has gone out every week since the spring. It is the most useful thing AI does for us, and it only works because we never moved our work into Slack.
What lands in the inbox
The report has six sections, always in the same order:
- Top three topics of work. The three things that got the most attention, one sentence each.
- Congratulations. One or two people who made real progress on something that matters.
- The two biggest problems we are working on. The open problems the team is actually wrestling with, not the ones on a roadmap.
- What to focus on next week. Two or three things, drawn from open tasks and unresolved threads.
- Promises and commitments. Who promised what, to whom. Next week's report checks whether it happened.
- Situation Room suggestions. Tasks that look finished and should move to done, and topics that came up in email but are not tracked anywhere.
The fifth section is the one people notice. A supplier says "we'll send the revised quote Thursday," a colleague says "I'll chase the forwarder," and a week later the report asks whether either happened. No project manager does that for a team of six. Now something does.
The sixth keeps our task spreadsheet honest. A spreadsheet only works as a project tool if people keep it current, and people are bad at that. A weekly nudge that says "this looks done" and "nobody is tracking this" is enough.
We can't show you a real one. It is full of our suppliers, our prices and our problems. Which is the point: it is only useful because it knows our business, and it knows our business because it reads what we write.
How it works
Four small Python scripts and two programs that are older than most SaaS companies:
notmuch + the task spreadsheet
│
▼
clean_emails.py fetch, filter and clean the week's threads
│
▼
generate_report.py build the context, call the model, save the report,
│ email a draft to one person for review
▼
reports/YYYY-Www.md kept on disk, last 26 loaded as context next run
│
▼
send_report a human reads it, then it goes to the team
The mail. Our mail sits on disk in Maildir, which is just one file per message. notmuch indexes it. One query pulls everything from the last seven days that went to or from our team lists, and notmuch show --format=json hands back clean, threaded JSON. No API, no OAuth, no rate limit.
The tasks. Our Situation Room is a spreadsheet with a "doing" tab and a "done" tab. The script downloads both as CSV.
The cleaning. This is most of the code, and it is boring on purpose. Strip quoted replies, signatures, legal footers and HTML. Drop newsletters and notifications. What is left is what people actually said, and it fits in a single context window with room to spare.
The memory. The last 26 reports go in too. Six months of the company's own summaries, so the model knows that the container delay is week four of a story, not news, and can check last week's promises against this week's mail.
The model. One call to the Claude command-line tool with a system prompt that defines the six sections. That is all the "AI integration" there is.
The human. generate_report.py runs from cron and emails the draft to one person. send_report is run by hand: it shows the draft, lets you fix it, converts it to HTML with pandoc and sends it through msmtp. Nothing goes to the team that a person has not read.
Why this only works on email
Our conversations with suppliers, freight forwarders, warehouses and marketplaces happen in email. So do our conversations with each other. They sit in the same archive, in the same threads, often in the same message.
So the model sees the whole picture. When a factory pushes a ship date and someone tells the team in the same thread, both halves are there. In a company that talks to the outside world by email and to itself on Slack, you have to stitch two systems together before AI can tell you anything. Most companies never do, which is why "AI that knows your business" mostly means AI that knows one app.
The other half is format. Maildir is text files. A CSV is text. Our wiki is text files on disk. Nothing has to be exported, synced, connected or paid for before a model can read it. We did not choose these tools for AI. They just turned out to be the right shape for it.
Why no MCP server
People ask. An MCP server is a good way to let a model go and fetch what it decides it needs. For a weekly report we already know exactly what it needs: last week's mail on the team lists, the task sheet, and the last six months of reports. notmuch builds that precise slice ahead of time, and the model just reads it. Fewer moving parts, the same shape every week, and an input file you can open and check when the output looks wrong.
For ad-hoc questions it is even simpler. Open a terminal, start an AI coding assistant, and ask: "use notmuch, find threads from the last ten days that mention the new carton size, and tell me where we landed." It writes the query and reads the result. No server needed.
The rules we keep
- It drafts, a person sends. Every report is read before it goes out.
- It interprets, it doesn't calculate. Numbers come from programs. The model reads what people said.
- It doesn't write our mail. An email from a person should have been written by that person.
What it cost
A few evenings to write the scripts, most of them spent on the email cleaning, and some small bugs fixed along the way. The model call costs less than a coffee a week. There is no new subscription, because the data was already ours and already readable.
That is the real argument for running a company on open standards in 2026. The savings are nice: our whole IT stack is about $52 a month for a business doing $7.5M a year. But the bigger win is that everything we write is already in a form an AI can read, on our own disk, without asking anyone's permission.
The whole stack, written down
This workflow is chapter eight of The $52 Company, our book on running La Nao on email, mailing lists, spreadsheets and plain files. The introduction and the first two chapters are free to read here.
Get the book →Out now on Kindle, $4.99. The complete scripts behind this report, commented, are in the toolkit.