I have not filed an invoice myself in a year

The trigger was a weekly review last winter. I went looking for an invoice from the month before, could not find it, and discovered while searching that my Downloads folder held over a thousand files. Invoices, quotes, screenshots, three versions of the same PDF. The review I had blocked ninety minutes for turned into a clean-up session.

This is not a time problem. Two minutes per invoice are two minutes. The problem is what those two minutes do to the system I am supposed to trust.

Why digital clutter is a GTD problem, not an efficiency problem

Any task I do by hand is a task I can also forget. And any filing habit I sometimes let slide becomes a filing system I no longer trust. If I am not sure whether the invoice ended up in the right folder, next time I search in three places instead of one. That costs more than time, it costs trust. And a system I do not trust is one I have to carry in my head — exactly what GTD is supposed to end.

The two-minute rule works for one invoice. For forty a month it is a bad idea. Forty times two minutes is over an hour, and more importantly it is forty interruptions. Filing is not the work. Filing is the condition under which the work stays findable.

Tasks like that do not belong on a list. They belong gone.

What changed: the programmer now sits in your browser

I have been building software for over thirty years, professionally since 1998. I can write scripts. I still did not do it for years, because the maths did not work: two evenings of work for a script that saves me ten minutes a week pays for itself somewhere inside a year in which I no longer need the script.

That calculation no longer holds. I describe to an AI what should happen, it writes me the code for Google Apps Script, I test it and adjust two small things. Two evenings become forty minutes. At forty minutes, almost anything is worth it.

If you have never written a line of code you can get there too, but you will need patience while testing. That is the honest part most articles on this leave out. More on it below.

Six automations running here

1. Invoice filing with no clicks at all

Trigger: an email carrying the label Invoice, or a PDF landing in a particular cloud folder.

What happens: the script pulls the attachment, sends it to an AI model for extraction, gets back supplier, invoice date, number and gross amount, renames the file to the pattern 2026-05-15_Supplier_InvoiceNumber.pdf and files it under year and month. In parallel it appends a row to a Google Sheet.

Result: a filing system with consistent names that I can actually search, plus a summary table I can export for my accountant. The effort is zero, as long as nothing breaks.

Extraction is the part that did not work before. OCR on its own gives you text mush. A model that understands which of the four numbers on the document is the gross total turns that mush into structured data. That is the actual leap.

2. The project folder appears when the card does

Trigger: a new card in a particular Trello list.

What happens: the automation creates a Drive folder with the same name and writes the link back onto the card.

Result: I never again have to decide where the files for this project go. Sounds trivial. It is, however, the decision I used to postpone every single time, which is why project files ended up in Downloads. See above.

3. The weekly review gets its data served

This is the automation that helps me most in methodical terms.

Trigger: Friday morning, on a schedule.

What happens: a script collects every Trello card that moved to Done this week, every calendar entry from the past two weeks and the next two, every project without a next action, and every waiting-for entry older than ten days. All of it lands as a single document in my vault.

Result: I start the review by thinking rather than by gathering. The thirty minutes that used to go into scraping the facts together now go into the question of whether these are even the right projects.

If your review keeps slipping, start here. The most common reason reviews get skipped is not a lack of time, it is the hurdle of the first fifteen minutes.

4. A waiting-for list that keeps itself

Trigger: I apply the label Waiting for when I send a mail.

What happens: once a day a script scans every thread marked that way, checks whether a reply came in, and maintains a list with recipient, subject and the number of days since sending. Threads that got an answer disappear on their own.

Result: the list is always current without me tending it. And I can see at a glance what has gone unanswered for three weeks. That is the point at which a waiting-for turns into a phone call.

5. Meeting notes to next actions

Trigger: a document in a folder called Meeting notes.

What happens: a model reads the note and pulls out everything that sounds like a commitment, a task or a date, each with an owner. The result lands as a list in my inbox, not straight onto a project list.

That last point matters. I do not let the AI decide what a next action is. It delivers raw material into the inbox; the clarifying is mine. Hand that step to the machine and within two weeks you have lists full of things you never agreed to.

6. Downloads tidies itself

Trigger: daily at six in the morning.

What happens: anything older than seven days moves into a folder Archive/Year-Month. Images, installers and zip files older than thirty days are deleted.

Result: the folder is what it is meant to be again, a place things pass through rather than a warehouse. This is the simplest automation of the lot and it cost me the least. Ten minutes.

Where it goes wrong

Four things I learned, some of them the hard way.

Automated mess is still mess. If your folder structure did not work before, a script just breaks it faster. Sort out the process first, then automate. That is the same mistake as buying a new tool instead of repairing your own system.

The code runs until it doesn’t. AI-generated scripts typically have no proper error handling. When Google has a moment or a PDF is corrupt, the thing fails silently. I have it email me on every error. That is the one line of code you should not skip.

Watch out for idempotence. A script that runs over the same data twice must not create the same folder three times. Ask the AI for it explicitly; it rarely offers on its own.

Not everything should be automated. Clarifying and deciding are the heart of the matter, and no machine can take those off me. What can be automated is transport: moving files, renaming them, gathering data, keeping lists current. The moment the question becomes what something means to me and what the next action is, I am up.

How to start

Take the most annoying recurring task of the past week. Not the biggest — the most annoying. Then describe it to an AI like this:

Write me a Google Apps Script. Every day at 6 a.m. it should move every file in my Drive folder “Downloads” that is older than 7 days into a subfolder “Archive/YYYY-MM”, matching the file’s modification date. Create the target folder if it does not exist. If an error occurs, email me the error message. Explain step by step where I paste the script and how I set up the trigger.

That last sentence is the important one if you have never worked with Apps Script. Test with a handful of copies first, not with your real data. And never let a script delete anything permanently until you have watched it for two weeks.

My first automation was the Downloads one. Not the most elegant, but the one that made me want to carry on.

Common questions

Do I need programming skills to automate routine tasks?

No. For scripts in Google Apps Script, or automations between Trello and Drive, describing the desired sequence precisely is enough. What you need is patience while testing and a willingness to feed error messages back to the AI instead of giving up.

Which tasks are worth automating?

Anything that comes up regularly, follows clear rules and requires no judgement: renaming and filing documents, creating folders, extracting data from documents, keeping lists current, gathering information for the weekly review. What is not suitable is anything that requires a decision about meaning and priority.

How much time does this actually save?

For me it is around three hours a month on invoice filing and about two hours on preparing the review. The bigger effect lies elsewhere: the filing is reliable, and I no longer have to think about it.

What is the most common mistake when automating?

Automating a broken process. If your filing structure did not work before, a script only speeds up the chaos. Settle the structure first, then automate.

If the same thing happens by hand every day at your place: automation and AI is exactly that — analysis and implementation in production, not a prototype. Tell me what your recurring handwork is.