Sonntagsausgabe · Part 10
How do you get out again?
On Thursday, 4 September, ChatGPT, Claude and Grok went down at the same time. Worldwide, for hours. For many companies that was a day without warning — for us it was a test we had already written down.
We run an AI agent that takes part in a good chunk of operations: shipping software, planning posts, time tracking, feeding the books. It computes in the cloud of one of those three providers. That afternoon it stood still.
What kept running: mail sorting, the search across the company archive, dictation. Everything that computes on the graphics card in our own office. Since July without interruption, that Thursday included.
Three weeks ago I wrote here that sovereignty is not a question of origin but a question of exit: not “where does it sit” but “what does it cost me to leave”. That was the theory. 4 September was the exam. This Sunday edition is the record: the exit test, run on our own company, with numbers.
The test
The rule is simple. Take your biggest vendors, one line each. In every line one number: how many days until we work again without this vendor. Not “how many days until the replacement is perfect”, but until we write invoices again, answer mail, serve customers.
The number may be an estimate. It may not be missing. Because a line without a number is not an arithmetic error — it is a decision nobody has made yet.
Here is our list. Seven vendors, seven numbers, as of today.
| Vendor | Level | Days until we work without it | Status |
|---|---|---|---|
| Password manager (Dashlane until July) | Data | 14 | left, July 2026 |
| Mail server (Exchange until July) | Data, identity | 14 per domain | left, July 2026 |
| Services at the hosting provider | Operations | 3 | moved, August 2026 |
| Domain registrar (abroad) | Identity | 5 (transfer window) | move has a date |
| Relationship | 0 for the content | rented, on purpose | |
| Cloud AI for the agent | Operations | 0 for document work | covered locally |
| Local AI on our own graphics card | Operations | 0 | running since July |
Two of those lines surprised me. The rest confirmed what we knew. But the two are the reason for this text.
First surprise: the cheapest door was the one I hesitated at longest
The password manager. Not switched for years, because everything is in there and you do not touch that. When we finally did, the exit was exactly what it says on the box: export, import, run both in parallel for two weeks, cancel the old one. Fourteen days, no data lost, no drama. The dependency was never technical. It was habit.
Same pattern with the mail server, one size up. Two weeks of parallel operation per domain, the first domain as a pilot, then the next. The mail server is ours now, on a machine whose configuration lives as text in version control. If it dies tomorrow, it stands again the day after.
The lesson: the exit number is usually smaller than the feeling says. But only if you work it out once.
Second surprise: the biggest number is not at the biggest vendor
I would have bet on the hosting provider. Whole servers, everything on them. It is the line we have under the best control, because it is the one we have practised most: every machine is described as text, every backup sits encrypted in a second place. In August we moved half our services from one machine to another — mirrored on Wednesday, everything running in its new home by Sunday, restore from backup proven. Three days.
The biggest number sits at the smallest line item: the domain registrar. Twelve francs a year, and everything else hangs beneath that line — mail, logins, the website, our identity on the open networks. Whoever holds the domain holds the company. Technically the exit is cheap: release the lock, fetch the code, five-day transfer window. It only gets expensive if you do it for the first time in an emergency. Which is exactly why the move to a Swiss registrar now has a date in the calendar instead of a place on a list.
The lesson: do not sort the list by invoice amount. Sort it by what hangs beneath the line.
The line that 4 September answered
And then the cloud AI — the line that needed no estimate, because it was tested live that Thursday.
For document work the replacement has been there for a while, and it is measurably better: in August, the model on our own graphics card beat the cloud in four out of five disciplines on our accounting data, at zero francs per run. That is why mail sorting, archive search and dictation simply kept going on 4 September. Zero days, with nobody having to switch anything over.
For the agent that ships software and co-runs processes, the answer looks different: no local model replaces the cloud there today, so every one of its tasks comes with a written procedure for how a human does it by hand. On 4 September that meant an afternoon without automation. No standstill, no lost order, no improvising — the work ran slower until the service was back.
That is the whole difference between dependency and decision: we buy cloud capability because today it is the better one. And for every task we know what an outage costs — in hours, not in guesses.
What we decided after the test
A test that only produces a table is busywork. Three decisions came out of it, each one line long.
- Whatever reads documents runs on our own graphics card first. The cloud is the exception from now on, not the default — it is faster to obtain, but it is not faster.
- The registrar move has a date. With a test run on an unimportant domain first, so the five days are rehearsed before they count.
- One line per new purchase. Before we buy a service, the exit number goes into the decision. Whoever cannot name it does not buy. That holds for every vendor, and it holds for us first.
Your version of the test
Ten minutes, a sheet of paper, no consultant. Write your five most important services one below the other: accounting, mail, filing, the industry software, the platform where your customers find you. Next to each, one number in days.
Under three days: sovereign. You stay because you want to.
Over thirty days: that is no longer a dial. That is a switch, and someone else has their hand on it.
A line without a number: that is your next project — and the only line that will genuinely surprise you when it matters.
For us the test produced seven lines and seven numbers. The biggest one sits at twelve francs a year. That is exactly what you run it for.
What number sits at your biggest vendor?