Tell me which task repeats every week.

I design the system that takes it over. The plan and the price are free, and you see both before anything is built. I use a language model where the work needs judgement, and ordinary code where it does not.

I run Routine Bureau on my own, from the Netherlands. You speak to the person who also builds it.

A robot sitting behind a desk, with a stack of paper beside it

The kind of work a model is actually good at

The test I use: is someone reading something, and then deciding what to do about it? A rule cannot do that. A model can, and it can hand back the cases it is unsure about instead of guessing.

01
Someone reads every message that comes in and decides where it goes
A model reads it, routes it, and puts the ones it is unsure about in front of you
02
Someone retypes what is already sitting on a PDF, a form or a photo
It reads the document, pulls the fields out and flags anything that disagrees with your records
03
Someone writes the same kind of reply forty times a week
A draft that answers in your own earlier words, ready for you to send or change
04
Someone writes up a call, a meeting or a site visit afterwards
The recording goes in; the summary and the follow-up actions come out where the team can see them
05
Nobody can find what was agreed a year ago
An archive you can ask a question, that answers with the source attached
06
Someone checks one list against another, line by line
A match that explains itself, and stops to ask when two lines are close but not the same

Your version is probably not on this list, which is fine. The same thing runs through all of them: a person has to read something before they can act on it.

Why I started selling this

My father has worked in IT for years. He came round, saw the server and saw what I do with it. He said I should sell this to businesses and that they would pay well for it.

I had the same idea. I kept putting it off, because I can build but I cannot make cold calls.

So this is the agreement I made with him: if someone comes to me with a specific problem, I'll take it on. That is what I set up Routine Bureau for. I have only just started and work comes through introductions. Building systems is not new to me.

What I run at my own desk

Six systems that are running right now, laid out as they would be on a desk. Two of them are businesses that run unattended: the website platform below, and a housing-alert service that watches Dutch rental sites and emails its subscribers the moment something matches.

The desk, seen from above A desk drawn from above with the robot at the top edge and six marked objects on the surface: A an invoice tray, B a ring binder, C a laptop, D a chart readout, E a rack of small modules, and F a gate with a stop mark. A B C D E F
ROUTINE BUREAU DESK NL 2026
A

Invoice Fetcher

A browser extension that collects the invoices from every supplier portal you use and hands them back as a single zip. It works inside sessions you are already signed into, so no passwords are stored anywhere. A portal it has not seen before, you teach it by walking through the download once.

B

Concord

A private memory. What I am told goes in raw and stays raw. A model then distils it into short notes, links them to each other and records where every fact came from. Near-duplicates are held back for review rather than quietly merged, because a memory that edits itself is worse than no memory.

C

PageWren

A website platform that runs itself. One model writes the site from the business's real data, a second looks at the rendered screenshots and files defects against them, a bounded repair pass fixes what it found, and then it ships. Client sites run on it, and this page came out of it. pagewren.com

D

Search and AI visibility

An engine that crawls a site, joins it to Search Console, market data and live search results, and reports what actually changed. It also tracks the newer question: whether AI assistants mention a business at all when someone asks them for one. What was measured and what was inferred stay in separate columns.

E

Twenty-five specialist agents

A library of task-shaped agents I reuse instead of starting from nothing: secrets handling, DNS and domain moves, mailbox provisioning, server provisioning, market research, compliance documents. Most new work starts as one of these plus a brief.

F

ExecGuard

The rail underneath an automated decision. Every order is checked against limits fixed in advance, and anything outside them is stopped before it reaches the market. When a system is allowed to act on its own, this is the part that says what it may never do.

How it goes when you email me

  1. STEP 01You get in touch and tell me the task. One paragraph in plain language is enough.
  2. STEP 02I work out what it takes and write you a plan with one price for the whole job. This part is free.
  3. STEP 03You approve it. Nothing gets built before you do.
  4. STEP 04I build it and test it.
  5. STEP 05You get it delivered and running.

What does it cost?

The plan and the price cost you nothing. You send me the task, I work out what it takes and what I would charge for the whole job, and you get that in writing. If I think it is not worth automating, I will tell you that instead, and it still costs you nothing.

I don't quote before I understand the job. An integration between two systems is not the same shape as a model that has to read your post, and a number invented up front is a number that turns out to be wrong.

Contact

Write down what comes back every week. I'll read it, ask what I need to ask, and send you a plan and a price. That costs nothing.

You can also email me directly: [email protected] Based in the Netherlands. I work remotely, and visit when a job needs it.

This form opens your own email app with the text already filled in. If that does not work, email me directly at the address beside it.