News

What happens when you let an AI look into your office for one day?

Most people expect a project: workshops, a concept, three months until the first result. We do it the other way round. One day, one mailbox, one folder. And the machine may only read.

In the morning you choose what it is allowed to see: a single mailbox and one folder of documents. By the afternoon both flow into an archive that understands questions. Then you ask your own questions — not our examples. «What do we owe this company?» «When does the contract expire?» «Who replied last to the offer?» The answer comes with the sources it was taken from.

By the evening you know two things nobody knew before: how well your own records can be questioned, and where the admin work really sits in your company. The report from that day is also the offer — which tier, which scope, which price.

Nothing is changed on that day. No mail is moved, no document renamed, nothing leaves the house.

One day shows what is in one mailbox and one folder. What becomes of it over 90 days is shown by the pilot: then agents reach into the inbox, sort, file and prepare replies — as drafts as well, sent only after your yes.

The audit day costs CHF 2,000. What it covers is described on the service page.

Which question would you ask your own archive first?

What happens at your company when a server reboots at night? In most companies: you hope everything comes back up. Usually it does. And if it does not, someone notices in the morning — often the customer who calls.

An AI-run back office also works at night. A reboot must not stop it. Security updates therefore run at night, and afterwards every server comes up in a fixed order: first the encrypted link between the servers, then the database, then the services that need it.

A server does not know this order by itself. You have to tell it — and tell it what to do when something is missing: wait until the link is up, and try again by itself after an error. A few lines of configuration, set up once.

The lesson goes beyond technology: most outages are not defects. They are orders nobody wrote down. The routine that only works because the same person always arrives first belongs in the same category. Writing it down and setting it up is part of operations: your back office, run by AI.

Which routine at your company only works because, by chance, the same person is always there first?

Do you know how many Swiss companies have fewer than ten people? Nine out of ten.

More precisely: 561,952 enterprises with one to nine employees, around 90 percent of all market-economy enterprises, as of 2023 according to the Federal Statistical Office. Switzerland is not a country of corporations. It is a country of micro-businesses: fiduciaries, law firms, property managers, trades, practices.

And these are exactly the companies that hear the sentence most often: “You are too small for that.” Too small for IT of your own. Too small for automation. Too small to keep your data in-house. So they rent everything from the big players, on the big players’ terms.

It is the other way round. The smaller the company, the larger the share of the week that goes into admin, because there is nobody to take it off your hands. A company of four feels every hour of administration more than a corporation of four hundred.

We built our back office for exactly this size: one company, one mandate, a handful of people. For the four from the start, not as a small edition of something big.

We are one of these companies ourselves. What we build for micro businesses runs with us first — and has to stay simple for four people. This is how your back office, run by AI works.

Have you ever been told “too small” — and believed it?

Do you know where the spam in your inbox got its name? From a tin of luncheon meat.

SPAM has been a brand of tinned meat since the 1930s; the name is widely taken to be short for “spiced ham”. In 1970 a British comedy troupe made fun of it: in a sketch, every dish in a café comes with Spam, and a choir of Vikings sings the word until no conversation is possible. Spam, Spam, Spam, Spam.

That is exactly how the first internet users felt when, in the 1980s, someone flooded their discussion rooms with the same text over and over. They called it spam. In 1994 two lawyers from Arizona posted the same advert to thousands of newsgroups at once — the first big wave. Ever since, anything unsolicited and sent in bulk has carried the name.

The funny part: today the problem is not the spam but the filter. It decides what you never get to see. The quote, the confirmation, the bank statement — if it sorts them out, you will not hear about it.

That is why an AI-run back office keeps a list: every mail you are waiting for goes on it. If it lands in spam, a machine pulls it back into the inbox and pings the phone. If it has not arrived by the deadline, the machine pings as well — “did not arrive” is information, not silence.

The list almost keeps itself: whoever waits for a mail says so once — with sender and deadline. The machine does the rest. The filter that decides for you becomes an inbox that looks out for you. What that looks like day to day: your back office, run by AI.

How many mails did your spam filter decide for you this week?

How many hours a week go into admin in your company that nobody ever ordered? Sorting mail, filing receipts, reconciling appointments, hunting for that one figure in last year’s contract. Nobody ever wrote it down as a task. It is simply there, every morning, before the real work begins.

Do the sum with your own hourly rate: half a day a week, over a month, at CHF 100 to 120 an hour comes to around CHF 2,000. Every month. Not as an invoice, but as time you cannot bill to anyone — because it is usually the owners doing it themselves, in the evening.

We asked what it would take for all of that to be done by 8 a.m. Not faster — done: mail sorted, documents in the right place, replies drafted, any question to our own archive answered in seconds. It has been running like that for months, for our own company and for mandates. On servers in Switzerland. The agent drafts, a human sends.

There is now a page for it. And a first step that takes exactly one day: one mailbox, one folder, read-only. The same afternoon the archive answers your real questions, and you see on your own data what is left of your admin.

The CHF 2,000 is our own sum with an assumed hourly rate, not a measurement from your company. How much it is for you shows on the day with your data. That is why it starts there, not with a quote.

What part of your admin would you hand over first?

How many mailboxes do you open every morning? Company, mandates, board seats, projects — each with its own login, each with its own mess.

An AI-run back office turns that around in two steps.

First: everything flows into one inbox. Your own mailboxes move, other people’s are only read. Your clients’ mailboxes stay where they are — and you still have one place instead of five.

Second: an agent reaches in and works. It sorts the mail, files documents and prepares replies as drafts. The agent drafts, a human sends.

The result: you only read what matters, and you decide instead of sorting. This is how your back office, run by AI works.

Which mailbox would you hand over first?

In how many calendars is your next appointment? If you have to think for a moment, you are not alone. The company calendar. The private one. The calendar of a mandate. A client’s shared calendar. Each sensible on its own — and together the reason appointments get double-booked or forgotten.

The usual fix is discipline: check everywhere, copy everything. That lasts exactly until things get busy.

The better fix turns it around. Calendars and contacts flow automatically into one single calendar on your own server. Every 15 minutes, in one direction only, so nothing ever breaks at the source. There is exactly one place to look.

What matters is not the technology but a rule: the source stays where it is. Nobody has to move, no client has to change anything. You only read together what is there anyway.

For an AI-run back office this is the precondition. An agent that prepares appointments and flags overlaps needs one calendar — not five. This is how your back office, run by AI works.

How often did you miss an appointment this year that sat in another calendar?

Do you know where the @ in your email address comes from? From the wine trade.

In 1536 the Florentine merchant Francesco Lapi wrote a letter from Seville to Rome. It says an amphora of wine cost 70 to 80 ducats — and instead of «amphora» he wrote an a with a loop around it. The @ was shorthand for a measure, as ordinary as «kg» is today.

More than 400 years later, in 1971, an engineer needed a character that appears in no name, to separate person and computer. He took the @ from the keyboard. Since then it has been in every email address.

So the character has already changed jobs twice: from wine measure to address. And the email address itself is often a company’s most valuable asset today. It is the customer relationship, more than any card file.

That is exactly why one question is worth asking that hardly anyone asks: who owns the address behind your @? Whoever owns the domain can move the mailbox at any time. Whoever only rents a mailbox stays with the provider.

Honestly, as far as we are concerned: our addresses also sat with big providers for a long time. We only moved this summer — and not all of them, because mailboxes at clients and partners stay with them.

Who owns the address behind your @?

Sources: Smithsonian Magazine, «The Accidental History of the @ Symbol»; Wikipedia, «At sign».

Before client data goes into an AI: five questions. None of them needs an IT specialist, and a provider can answer all of them in writing — if they want to.

1. Where is it computed? On your servers, in Switzerland, or wherever the provider finds it cheapest right now? The answer belongs in the contract, not in the sales pitch.

2. Who can read along — even just «for troubleshooting»? Many providers reserve the right to look at requests for quality assurance or abuse monitoring. For a law firm, that is exactly the decisive point.

3. Does the model learn from your data? And can that be excluded by contract, not just by a setting that may change with the next update?

4. Does each person only see what they may see? A tool that shows everyone everything moves the problem from the provider into your own team.

5. Can you get out again? With your data, complete, in a format you can open without the provider.

Whoever can answer these five has usually already answered «may we use AI?».

Honestly, as far as we are concerned: our answer to question 1 is «on our own hardware in our own building». That has a price: the model there is weaker than the big ones in the cloud. Where that limit lies is what the Sunday edition works out.

The five pages are meant to be printed — for your next conversation with a provider.

Can a law firm use AI at all? The usual answer: not really. Client data must not go to a provider that does not guarantee what happens to it. Anonymising destroys the workflow. And local models are said to be too weak. So the firm stays out while everyone else gets faster.

I think that is a false shortcut. Professional secrecy does not forbid AI. It forbids giving up control. That is a difference, and it can be built.

This is how it looks at our company: our archive — 144,000 mails and 13,500 filed documents — can be asked questions in plain language. «When did we last talk to X about Y?» The answer comes with the source. The search runs on our own servers, and the answer is written by a model running on a graphics card in our own building. No document leaves the building. And every person only sees the collections released to them.

Honestly, as far as we are concerned: the model in our own building is not as strong as the big cloud models. We would not use it for «think for me». But for the question law firms ask most often — «where does it already say that?» — you do not need a top model. You need a reliable one.

On Sunday the full calculation: which tasks need which model, and what the limit costs.

Which question do you ask your archive most often?

Do you know what a computer originally was? A person.

The English word «computer» first appears in 1613, in the English writer Richard Brathwait, and it means someone who calculates. For more than 300 years it was a job title: people who worked through columns of figures in observatories, insurance companies and engineering offices. The machine only got the name at the end of the 19th century. And by 1943 most human computers were women — calculating trajectories and tables until machines took over that work.

The job did not disappear because people calculated badly. It disappeared because the work itself became a machine. People then asked different questions.

Exactly that is happening again. Not with calculating, but with searching, summarising, filing and following up. The question for management is therefore not «will AI replace my people?», but: which work at our company is already a machine that is still being carried out by people?

Honestly, as far as we are concerned: we made this list for our own company, and it was longer than expected. Collecting receipts, watching for awaited mails, recording time — all work a machine does today while we ask the questions it cannot ask.

Which activity at your company would be a museum piece in 50 years?

Sources: etymonline.com («computer»), Wikipedia «Computer» (etymology).

Do you know what a protocol originally was? A glued-on sheet.

The Greek word «prōtókollon» literally means «the first thing glued on». It was the first sheet stuck to the front of an official papyrus roll. It stated who had written the document and when. Whoever opened the roll saw first: who stands behind this?

More than two thousand years later the same question arises with every machine that works in a company. Which machine did this? With what? Does the result still match what it delivered back then?

Today we glue that sheet on digitally. Every graphic and every video we publish carries a digital seal: who made it, with which tool, and a fingerprint of the file. More than 240 works are sealed this way, and each one can be checked publicly. Change a file by a single pixel and the fingerprint no longer matches.

For a law firm or a fiduciary this is not a gimmick but the core question of every automation: can you show afterwards who did what, and when — person or machine?

Honestly, as far as we are concerned: we built the seal for our graphics first, not for files. And there were weeks when works ran before they were in the register — we catch up after every publish since.

Which work in your company has no glued-on sheet today?

Source: DWDS, Etymologisches Wörterbuch des Deutschen, «Protokoll».

The best purchase is sometimes the one you postpone. That sounds like saving money. It is not. It is a decision with its own value.

Anyone who buys hardware knows the situation: the old machine still runs, but not forever. The dealer has a solution that fits today. And in a few months the next generation arrives, which does exactly what you actually need.

In August a new platform for our AI machine was on the table, around 3,700 francs. The alternative would have been a cheap intermediate step on a socket the manufacturer no longer develops. Both would have worked.

We bought neither. A new platform generation had been announced, and everything we would have bought then would have been thrown out at the next step. The old machine keeps running. Its ailing disk is mirrored: if it fails, the second one takes over, and we replace a disk instead of a platform.

Waiting is not an excuse when three things are true:

1. You know what you are waiting for — not «better times», but a specific product or a specific price.
2. You know what carries you until then, and you have secured it.
3. You have written down when you reopen the decision — for us: as soon as the new generation is available in Switzerland.

Honestly, as far as we are concerned: waiting has a price. The old machine is not as fast as we would like, and we plan around its limits. That is a deliberate choice — but it is not free.

Which purchase is coming up for you that would already be replaced a year from now?

Can a traffic jam consist of a single car? Yes. And it can cost you hours without anybody noticing.

Picture a traffic light that only turns green when the junction is empty. A sensible rule. Now the sensor counts the car waiting at the light. The junction is never empty, and the car waits forever — for itself.

We had exactly that jam. A background job was meant to step back whenever somebody used the system. A good rule: whoever works with the system should not wait for a machine that is analysing documents on the side. Only the job counted its own work as «somebody». So it stepped back. And again. And again.

The tally of a single run: 2,201 times stepped back, 23 hours and 32 minutes spent waiting. During that time nobody had used the system for more than 30 hours. After the fix the job no longer counts its own load: zero delays.

The treacherous part: from the outside this does not look like an error. The job was running, the system was quiet, every indicator was green. It was just many times slower than necessary. No alarm finds errors like this, because nothing is broken.

Honestly, as far as we are concerned: we wrote that rule ourselves, and it ran that way until somebody measured. We found it because we wanted to know why the analysis took so long — not because anything raised an alarm.

If something at your company «just takes a long time»: has anybody ever checked what it is actually waiting for?

Before you pay for more performance: five questions for your database. Question 1 gave us a factor of 15, and question 3 cancelled a bill of 138 francs a month.

1. Is it running on factory defaults? They are chosen so the software starts anywhere, not so it runs well on your machine. Our database was allowed to remember 128 megabytes; its indexes are 8,247 megabytes. So it re-read from disk on every request. One line changed: 2,025 milliseconds became 90 to 136.

2. Is an index missing? A textbook without a keyword index gets read from the front. That is exactly what a database without an index does, on every request. The signs: a list that used to appear instantly now takes noticeably longer. A report that runs at night because it blocks everything during the day. A search everybody on the team avoids.

3. What generates the load, the big part or the small one? We have 27 gigabytes in four databases queried a few times a day. Next to them, 414 megabytes in two databases carrying 112 million transactions. Planning by storage space means planning past the problem.

4. What happens to open connections when the server changes? Without a broker the answer is: a maintenance window on Sunday evening. With one, every connection stays open, requests are collected and then sent on to the new location.

5. When did somebody last measure? The one question that makes all the others answerable. For us, the distance between «too slow» and «fast enough» was a single line.

Honestly, as far as we are concerned: that setting had been like this since installation and nobody had measured. We found it because we looked, not because a system warned us.

The five questions are meant to be printed — for your next conversation with your IT partner.

«Your archive is too big», says IT. But is it really the archive slowing you down?

Almost never. What takes up space and what does the work are two different things — and only one of them appears in the quote.

At our company 27 gigabytes sit in four databases that somebody queries a few times a day. Next to them sit 414 megabytes in two databases that carry 112 million transactions a year. The smallest item does the most work, by more than a thousandfold.

That is not a special case, that is the rule. The archive is big and quiet. The calendar is tiny and constantly moving. Twenty years of invoice archives fill the disk; the day’s bookings fill the queue.

Why it matters: whoever plans by storage space plans past the problem. Backups, servers, caches and contingency plans follow what works — not what sits.

Honestly, as far as we are concerned: our first thought was the bigger machine too. Only measuring showed that the smallest item was the bottleneck.

The test for you: name the largest amount of data in the company. Then the one without which nobody can work tomorrow. Are they the same?

Another quote for a bigger server?

It usually comes with the same justification: too little memory, too little performance, the system is at its limit. We had exactly that kind of offer on the table, 138 francs a month. We paid zero.

The reflex when performance runs short is always the same: buy bigger. It is not even wrong — just expensive, and often the second-best answer.

We first worked out what actually generates how much load. The result was unexpected: six databases were sharing one machine. Four of them are 27 gigabytes together and are rarely queried. Two are 414 megabytes together — and carry 112 million transactions.

The big ones and the small ones kept pushing each other out of the cache. Not because the machine was too small, but because two opposite working patterns sat in the same pot.

So we did not buy, we separated: the big, rarely queried ones went to a machine that runs anyway and has memory to spare. The small, constantly working ones stay where they belong — close to the programs that use them.

Honestly, as far as we are concerned: our reflex was the same as everywhere. The bigger machine was already picked, the price known. The only difference was that we measured before ordering.

Cost: zero instead of 138 francs a month. Reversible at every stage. What was the last purchase at your company that was really a tidying-up job?

Everybody knows this mail: «The system will be down on Sunday between 6 and 10 pm.»

You read it, sigh, and postpone whatever you still wanted to get done on Sunday evening. Nobody asks, because everybody assumes it comes with the territory.

It does not. A server move without a noticeable interruption is no magic, and the trick is old.

Between the programs and the database goes a broker. From now on all programs talk to it instead of directly to the database. During the move it gets two commands: Pause — it keeps every running connection open and collects the requests. Resume — it sends them on to the new location. As far as the programs are concerned nothing happened, except that one request took a moment longer.

The nice part: this broker is not only good for moving day. It pools connections, and we run a service that opens a new connection per call — 112 million transactions. That is exactly what it is built for.

We put it in place while everything was still running at the old location. So the risky part was done before the move even began.

Honestly, as far as we are concerned: the broker has only been in use since this move. We did not add it out of foresight, but because we had to move. Every maintenance window is a decision, not a necessity.

How long does everything stand still at your company when a server is replaced?

Know the feeling when Outlook cannot find a mail that is definitely there?

You know who it was from. Roughly when. You type three terms and nothing comes up. Then you scroll by hand, and there it is.

Behind it is the same principle that makes every piece of software slower the more you work with it: an index is missing.

Picture a textbook without a keyword index. You look for a term and leaf through from the front. At fifty pages that works. At a thousand pages you stop searching.

That is exactly what a database without an index does: it reads every record, every time. With a hundred customers nobody notices. With a hundred thousand everybody waits.

An index is the keyword list. It costs space and a little time on every write, and it saves the search across everything else. At our company the indexes take up eight gigabytes — for data you could hardly search without them.

How you can tell one is missing at your company: a list that used to appear instantly now takes noticeably longer. A report that runs at night because it blocks everything during the day. A search that everybody avoids.

Honestly, as far as we are concerned: our indexes were there. What was missing was the memory to use them — and we only noticed once we measured the slowest query instead of guessing.

That is rarely a reason to buy more hardware. It is a reason to take a look. Which report in your system deliberately runs at night?

How many times a day do you watch a spinning circle?

Opening a list. Searching the archive. Loading the customer file. One or two seconds each time, each time too short to complain about — and each time a person is waiting.

For us it was exactly two seconds per search. After one single changed setting it is under a tenth of a second.

The reason was banal. The database was running on the factory defaults from the installation package. The memory it may use to remember frequent lookups was set to 128 megabytes. The indexes it needs for that are a good eight gigabytes. So on every request it re-read from disk what it could long have had in its head.

One line changed, service restarted: 2,025 milliseconds became 90 to 136. A factor of 15, without new hardware, without one line of code.

Honestly, as far as we are concerned: that setting had been like this since installation. Nobody had measured. We found it because we looked — not because any system warned us.

The lesson is not «our database was configured wrong». The lesson is: factory defaults are made for nobody. They are chosen so that the software starts on any machine — not so that it runs well on yours. When did somebody at your company last measure instead of guess?

The most expensive software is not the one with the highest price. It is the one you cannot get out of.

Five questions belong before every signature. None of them is technical, and a vendor can answer all of them — if they want to.

1. Can a program operate this, or only a human with a mouse? Whatever only works through the screen cannot be taken over by automation later. This question decides whether an agent will ever work for you.

2. Do I get my data out, complete and readable? The promise does not count, the test does: request an export and actually open the file. What cannot be opened does not count.

3. Does the login go through our own door? Every piece of software with its own login is another password somebody has to manage, reset and block when an employee leaves. The test is simple: what happens on the day somebody leaves the company?

4. Who can switch it off? Not whether somebody could, but who, under what conditions, and how fast you would notice. Licence servers that phone home during operation, cloud accounts cancellable mid-term, remote maintenance without a log — that belongs in the contract, not in the sales pitch.

5. How many days would we need to leave? The most important question appears in no offer. Ask it before signing and you negotiate. Ask it afterwards and you pay. An evasive answer is an answer too.

Honestly, as far as we are concerned: tools we had used for years failed this check. They had no interface, and there is no trick that changes that after the fact.

The five questions are meant to be printed — for your next round of vendors.

Does your AI agent know your passwords? Ours doesn’t. And that is the whole trick.

The question comes up in every conversation about automation, and it is justified: if a machine logs in to portals on my behalf — where do the credentials live? In the script? In a text file? In the history of a chat?

At our company, in none of those. The credentials live in a vault that only one service opens. The agent tells that service «log me in to this portal», and the service fills in the login window itself. The agent never sees the password, it appears in no log, and it shows up in no chat history.

What that is good for became clear at the half-year close: 34 receipts collected from three portals, all named and filed the same way, in 90 minutes. Three logins, and not a single password left the vault.

Honestly, as far as we are concerned: this rule is younger than our automation. We only wrote it down in the summer, once the vault was in place — before that, the clean way simply did not exist.

The rule behind it is simple: an agent may be allowed to act without having to know. Where do the passwords live that your automations log in with?

Are you waiting right now for an e-mail that may have arrived long ago?

The case that made us build this: a promised bank statement arrived, was filed automatically as a machine notification, and nobody passed it on. Days of asking «has it come yet?» while the document sat filed and searchable in the archive. The problem was not the filing. It was the silence.

Since the beginning of September you can tell the agent: «I am waiting for X.» It puts that on a list and compares every new mail against it, every three minutes, in the inbox and in the junk folder. When the mail arrives, it notifies your phone. If it lands in junk, the agent pulls it out. And if the deadline passes without anything arriving, it reports that too — because that silence is the expensive case.

This is not a small thing. Checking whether a certain mail has arrived yet is interrupted work, every single time, and it happens exactly when something important is pending: a confirmation, a quote, a statement, a deadline.

Honestly, as far as we are concerned: the list has to be kept. Whoever forgets to register the wait, waits the old way. In return the agent now also reports when a deadline passes with nothing received — the notification nobody used to get.

The difference is small and changes the day: no more checking, just being notified. Which mail are you waiting for this week — and how many times have you already looked?

What has to be right in your company before AI is of any use at all?

Not the model. Not the tool. The question is decided beforehand, on five points, and none of them is a software question.

1. Your programs need an interface. Whatever can only be operated by mouse and screen, an agent cannot take over. Tools we had used for years failed this check at our own company.
2. Your filing has to be findable. Not pretty, not perfect — just so that a machine can tell what belongs together.
3. It takes one door for all programs, not twenty passwords. Otherwise every automation ends at a login window.
4. Your data has to belong to you and be exportable. What you cannot get out, no agent can read either.
5. You have to measure instead of guess. Without numbers you do not know what pays off.

This is not an IT list, it is a leadership list: every point is a decision, not a product. We worked through it at our own company, in weeks rather than years — and the point that brought the most was the second one. It needs no AI at all.

Honestly, as far as we are concerned: point 1 hurt. Tools we had used for years had no interface — and there is no trick that changes that after the fact.

Whoever has these five points needs no strategy to start, just a task. Whoever does not have them ends up buying a tool nobody can use. Which of the five is most likely to be your sticking point?

How many hours went into your last project? Guess first. Then look it up.

Most companies have two answers to that question, and both are wrong: the boss’s gut feeling and a timesheet that gets filled in from memory on Friday afternoon. What ends up on the invoice is a compromise between the two.

Time tracking rarely fails because of the software. It fails because it is work. So it gets backfilled, estimated, rounded — and the numbers are good enough for the invoice, but not for decisions.

Since July an agent keeps our clock. Not as a stopwatch somebody presses, but as a by-product of the work: it sees which project is being worked on and books the time there. Plus the cost of the machines that did the computing. At the end of the week there is one number per project that nobody estimated.

The uncomfortable part came immediately: two projects that felt small were not. And a task that had two hours reserved every week had long needed twenty minutes.

Honestly, as far as we are concerned: the clock only sees what happens at the computer. The phone call and the drive to the client it does not see.

The benefit is not control. It is the foundation: you can only decide what to automate once you know what costs you what. What did you spend most of your time on this week — and how do you know?

Your contacts on LinkedIn — do they belong to you?

The honest answer is no, and that is not a criticism of LinkedIn. Every platform works this way: it creates the relationship and it keeps it. If an account is suspended — accidentally, automatically, mistakenly — it is not only the posts that are gone. The list is gone. And nobody sends it to you beforehand.

For most companies this is the largest unsecured position in the whole business, and it appears on no risk list. You notice it only once you try to reach the same people without the platform.

The distinction is simple: rented is everything somebody else can switch off — followers, groups, ad accounts. Owned is what sits on your own address: a mailing list, a website, a feed anyone can subscribe to.

Honestly, about us: LinkedIn is still our largest channel, and that is exactly why it is our largest rented position. We write every piece on our own site first and link from there — and the short notes additionally go out over open networks that nobody can switch off.

The first step costs one afternoon: export your contacts while you still can. Then start a list that belongs to you and write the first piece into it. Even if it is only twenty addresses — twenty that nobody can take away from you.

Do you know which personal data your company collects? Most do not.

Not out of negligence. It is because the data spreads out: a form on the website, a newsletter tool, a CRM, three spreadsheets on different laptops, the bookkeeping. Each harmless on its own. Together they add up to a picture nobody in the company has ever seen in one piece.

The test takes an hour and needs no lawyer. Sit down and write out where personal data sits. Not what is permitted — only where. Most people stop halfway through because the list grows longer than expected. That is exactly the result.

Because only this list makes the usual questions answerable. Who may access it? What happens when somebody requests disclosure? What do you delete when a customer leaves — and in how many places? Without the list those are guesses; with it they are tasks.

Here, every place that holds personal data carries an entry with a purpose and a retention period. Not because it looks tidy, but because an agent working for us could otherwise not know what it is allowed to touch.

Data protection does not start with the statement on your website. It starts with the question of what you actually hold. Begin with one line: the form on your home page.

If your most important provider shuts down tomorrow — can you still read your data?

Most people say yes and mean: “It is all backed up.” Usually it is. Only the backup sits inside the provider’s system, in a format that only that system opens. That is a copy in the same sense that a spare key locked inside the car is a spare key.

The difference is the format, not the location. An invoice as a PDF will still open in ten years. A database dump only the provider’s software understands is worth exactly as long as that software exists. For contracts, bookkeeping and customer data that is the difference between an archive and a promise.

Here, the mail archive and the documents sit in open formats on our own servers, and the search across them runs on our own hardware — the same search that kept working on 4 September, when the large AI services went down.

Honestly, about us: we rent metal from a hosting provider and we buy cloud services. The difference is that for each of them it is decided in advance what the copy looks like and how long a move would take.

The test takes ten minutes and you can do it today: take one system, export everything, open the file on a different computer. If it does not open, you do not have a copy — you have a subscription.

Who owns the login your company starts every morning with?

Most people have to think about it. Access hangs on an account at a large provider, the account hangs on a mail address, and the mail address hangs on the same provider. As long as everything works, nobody notices. It becomes visible on the day an account is suspended — and then it is no longer an IT question but an operating question.

An example that concerns many people in Switzerland right now: in the canton of Zurich the deadline to register with justitia.swiss — the platform for electronic legal correspondence — ran out on 1 September. Registration goes through AGOV. The convenient path is an app built on Google services. Anyone who does not want that uses a plug-in security key; they start at around 26 francs and some are built in Switzerland. One path is convenient, the other depends on nobody. Both work.

That is the whole point: sovereignty is not a question of where something sits. It is the question of whether your access survives losing the foreign account.

Honestly, about us: our domains sit with a US registrar, and our visibility on LinkedIn is rented like everyone else’s. The difference is not that we avoid it — we know, for every single point, how we would get out and how long it would take.

The first step takes half an hour: write down who has which access — humans and machines in the same list. Next to each line, note which foreign account that access hangs on. The lines with no answer are the good ones.

How many days does your company stand still if your most important provider switches off tomorrow?

Hardly anyone knows that number. It is in no contract, it appears in no budget, and it usually gets asked on the day it is too late to ask. The 4th of September was such a day: ChatGPT, Claude and Grok went down worldwide at the same time.

On a day like that the only thing that counts at management level is how quickly you are working again without that provider. One day is annoying. Two weeks is a business risk — and the difference between the two is not a technical question but a decision someone has to have made beforehand.

What kept running here that Thursday is what computes in our own house: sorting the mail, searching the company archive, dictation. Since July without interruption. Not because we avoid cloud services — we buy them ourselves — but because for each one it is decided in advance how long we can manage without it.

The test takes ten minutes and needs no consultant: write down your five most important services. Next to each, one number — how many days until you are able to work again without it. The line with the largest number is your next project.

Which number is at the top of your list? Send us two lines and we will tell you what can be halved within four weeks.

What holds up a half-year closing? Not the bookkeeping.

It is the receipts that are still missing. Credit-card receipts, subscription invoices, expenses — each in a different portal, behind a different password. Your fiduciary is not waiting for numbers, they are waiting for paper. And collecting it costs you half a day you do not have.

This Monday that list was ours. An agent worked through it in a single run: logged in to three portals — without ever seeing a password, those stay in the vault. Searched our own mail archive for invoices, gathered 34 receipts, named them all the same way (date, vendor, amount), filed them the way the fiduciary wants them. And reported which ten items are still missing — instead of someone noticing in October.

34 receipts, 90 minutes, zero passwords seen. The only human touch: reviewing the reply mail and saying GO. Your Saturday belongs to you again.

What is on your closing to-do list? Send us two lines — we will tell you what you could hand off.

The five pages, for browsing:

Who guards your passwords from your AI?

You sign in ten times a day. Password, Face ID, fingerprint — nothing runs without it. Naturally.

And then you let an AI agent work for you. It needs the same accounts: mail, bookkeeping, customer data. Most people simply hand it the passwords.

At our place the agent never sees one. It asks, a human approves, a vault fills in. The agent gets the result, not the key. And the vault sits on our own servers, in Switzerland.

That is not security decoration. It is the precondition for letting an agent touch real systems at all. How it works: AI without your keys.

Why do you have to make three phone calls to learn a price? You don’t.

The usual routine: fill in a form. Two days later somebody calls and asks questions. Then a meeting. Then the quote. You wanted a number, you got a sales process.

We asked ourselves why we should do it that way — and found no good answer. So the price is on the website. Eight questions, 60 seconds, an honest range, right on your screen. No callback, no sales pitch.

A price you only learn after two meetings is not a price. It is a negotiating position. Try it: web analytics calculator.

How fast can you find a contract from 2019? Honestly.

First the folder on the shared drive. Then the mail attachments. Then you ask a colleague whether she remembers where it was filed. Twenty minutes later you have a file — and you are not sure it is the latest version.

Those twenty minutes cost nothing, says the bookkeeping. They only cost every day, for everyone, and nobody bills them.

At our place you type the question into the company chat and get the answer with its source. One minute. Paper too: putting it on the scanner glass is enough — a minute later it can be queried.

Start with a single folder, not the whole archive. The rest follows by itself the first time somebody finds something. How we approach old files: legacy cleanup.

Who really visits your website? Your tool doesn’t know.

500 visitors this month, says the dashboard. Sounds good. Except: part of them are bots nobody filters out. Another part clicked the cookie banner away and never shows up at all. And the data sits with an ad corporation in the US.

What remains is a number you cannot base any decision on. It still appears in every monthly report.

There is another way: count anonymously, evaluate honestly, hosted in Switzerland. Without a banner, because there is nothing to consent to. That is exactly how we measure our own website — Sunday’s two-month review was built on it.

How many of your marketing numbers would you sign off on if you knew how they were made? Calculator and package: web analytics.

How many hours does a single blog post cost you? With image, audio and distribution to every channel.

Writing is the least of it. Then comes the image. Then the audio version. Then the same thing four times into four channels, each cut differently. Half a day — and the text was finished after an hour.

For us it is one job. The post, its image, the audio version, the announcements in the channels: the same machine, one run. This very entry was scheduled by it — weeks ahead, on a single planning day.

We sell nothing that does not run in our own daily business. What that looks like for your marketing: channels you own.

What costs the most in a dispute? Not the lawyer. The sorting.

One binder. 2,400 pages. Four years of history. Contracts, amendments, invoices, mail threads, a few phone photos. Before anyone argues the merits, someone has to read all of it and put it in order. That work is billed by the hour — and it is the dullest part of the whole case.

Here is what the binder looks like after a machine has read every page:

• 12 contracts and amendments, 3 of them unsigned
• 87 invoices and payment records, 4 of them booked twice
• 1,930 emails in the thread, 61 of them relevant
• 220 scanned pages, some handwritten, all read

It gets interesting where documents turn into a sequence. Day 0, the other side confirms delivery in writing. Day 14, they qualify it in an email — but never withdraw it. Day 31, they pay a first instalment, without reservation. Day 96, they send a reminder based on an amendment only one party signed. Day 214, the claim arrives, resting on exactly the promise from day 0.

What you end up with is not a binder but an index. Numbered, citable, every line with its source. No. 9 is the one-sided amendment. No. 12 is the payment record that contradicts it. Your lawyer starts where they would otherwise be after two weeks.

The case above is invented. Names, dates and amounts match no real file — the method does. And the part that matters most: the file never leaves the house. No contracts, no client data, no payroll goes to a provider abroad.

The same thing works outside the law: checking a set of books, matching receipts to payments, working out who promised what and when.

Do you have a binder like that? Tell us what is in it — we will tell you what can be made of it.

Flip through — the five pages from the case:

How many software subscriptions does your company pay for? More than it uses.

Per seat. Per month. Per module. Plus the tool only one person really needs, and the other one that has quietly kept running since the project ended in 2023.

The bill grows silently, because cleaning up is nobody’s job. And because every single line item is too small to catch anyone’s eye.

We build companies their tools ourselves. Once, tailored, no subscription per click. On their own servers, in Switzerland. That is not worth it for every tool — but for the three or four that carry the daily business, it almost always is.

The first step costs one hour: list all subscriptions, sorted by “is used” and “is paid”. The difference is your budget for the rebuild. How we approach it: legacy cleanup.

Tomorrow LinkedIn locks your account. Now what?

All gone. The followers you built over years were never yours — you rented them. You do not know their addresses.

A newsletter is yours. Your own website is yours. The rest is a shop window, and shop windows can be cancelled.

How we solved this for ourselves — hub and spokes, what we own and what we merely borrow — is in the post Channels you own. The test question for your company: if your most important channel changes the rules tomorrow — what is left for you?

Where does your company start with AI? Small.

Not with the strategy that takes half a year. Not with the tool that can do everything and that nobody opens any more after three months.

Take the one process that annoys you today. Inbox. Quotes. Filing. Four weeks, one measurable result — then the next one.

The reason is unromantic: a small process can be understood completely. A large project cannot. And what you do not understand you cannot automate — you can only buy it.

That is how a company becomes AI-native: step by step, not in one go. Which process annoys you most this week? That one is yours.

The credit card statement arrives, and the receipts for it are everywhere: in the inbox between newsletters, as a scan, somewhere in the filing. In every company someone then spends an hour searching.

At our place a graphics card in the office does it. It reads amount, date and sender from the mail and files the receipt where it belongs — 1.3 seconds per receipt. No receipt leaves the building.

That is not a promise but a measurement — on our own bookkeeping, every answer checked by hand: reading out fields 97 % exact, sorting 100 real mails 92 % right, transcribing paper receipts 92.9 %. In 4 out of 5 disciplines the graphics card beat Claude Fable 5, the top cloud model. All numbers and all caveats are in yesterday’s LAB report.

Want this for your company? I am happy to walk you to this result.

The most-discussed piece on our channels by far was not a product announcement but a design decision: this website has no cookie banner. Yesterday the full post went live — how it works technically and legally, and what you deliberately give up.

The most frequent question since, publicly and privately: “Would that work for us?” The answer, in most cases: yes. The banner is only mandatory where visitors are tracked — if you count anonymously instead, there is nothing to ask. Everything better content and measurable marketing require, cookie-free measurement delivers in full.

So it is now a service: web analytics without a cookie banner — self-hosted analytics (Umami) on Swiss hosting or your own infrastructure, migration from Google Analytics included, and at the end the consent dialog disappears. Fixed price, set up exactly the way we run it ourselves — this website is the reference.

Our dashboard reported visitors from the US — at two in the morning. On closer inspection: Chrome on «Windows 7», zero seconds of dwell time, and every hit coming from the same Azure datacenter in Virginia.

So for one week we audited every single session. The result: 107 «visits» by raw count — 29 humans. Two thirds of our traffic were crawlers that execute JavaScript and therefore pass every common bot filter. In the AI era that is the normal case: machines reading for machines, dutifully counted as an audience by the statistics.

Our consequence: we count engagement-based now. A session only counts as human after ten seconds of dwell time or a real interaction — and the weekly report lists bots and humans separately, on purpose. Smaller numbers, but true ones.

This audit was only possible because we run our measurement ourselves — cookieless, banner-free, every session inspectable. With a black-box statistics product we could never have cross-examined those nocturnal Windows 7 visitors. Why that is also the more relaxed path legally is this week’s story on the blog.

Next to the monitoring dot in our taskbar sits a number: today’s working hours, counted live. Nobody started a stopwatch for it — and nobody will have to reconstruct a timesheet on Friday.

The principle behind it is beautifully simple: work always happens in the directory of the project at hand — so every working session knows by itself where it belongs. A small service mines the work logs hourly and keeps the books: hours per project, rate, and — because an AI agent works alongside us — its costs too, attributed token-exact to the right project. No timer, no back-filling, no month-end guessing.

A dashboard delivers the weekly picture: which mandate got how many hours, what the AI cost along the way, what is billable. The honest answer to «what does an AI hour really cost?» is in there at any moment — not as gut feeling, but as a ledger entry.

For clients this means billing transparency that does not rely on memory. What it looks like day to day is shown at the station «The computer keeps watch» on our new page How we work — the hour count is already visible in the taskbar image there.

For fifteen years the standard doorman for Linux servers was fail2ban: it reads your own logs, counts failed attempts and bans whoever rattles the door too often. Solid — but it only knows the attackers who knock on your door.

This month we replaced it across our fleet with CrowdSec. The principle is the same, with one decisive addition: detected attacks are shared with a worldwide community — as plain IP reports, never logs or content — and in return the server pulls the blocklist built from everyone’s reports. Anyone who misbehaved somewhere in the world is turned away here before their first attempt even lands. Neighbourhood watch instead of lone fighter.

The migration was uneventful — except for two snags worth passing on: the local API wanted the same port as a Docker service (fix: move the port, then install the firewall bouncer). And current Debian no longer writes a classic auth.log — the setup wizard detects this and correctly reads the system journal instead.

Why no antivirus, no heavyweight intrusion-detection system, and what a one-person fleet actually needs — the bigger security story is coming as a blog post of its own.

Since this week the notebook’s taskbar shows a small dot: green means everything on our server is fine. Red means something is broken — and a computer voice additionally announces it out loud, so it doesn’t starve in the corner of the eye.

The check service looks after the services, the disks and the nightly backup every two minutes. Ninety seconds after going live, the first alert came in — and it was real: last night’s backup had failed. Silently and unnoticed — the backup itself was there, but the cleanup step was stuck on an orphaned lock from the previous evening.

Without the dot we would only have noticed when we actually needed the backup. Which is to say, at the worst possible moment. As it was, it took five minutes to repair and harden against repeats.

Self-hosted means self-responsible — but not alone: the machine keeps watch with us. How the whole chain works, from the check script through the VPN to the speech output, we’ll tell in more detail on the blog soon.

It has existed since early July, and we never announced it: our newsletter. Time to catch up.

Subscribers get new blog posts and the Sunday edition as e-mail — nothing more. No “offers”, no third party reading along: the list runs on our own server (Listmonk, open source), the mails leave our own house, and the subscription ends with one click on the unsubscribe link. Double opt-in, it goes without saying — nobody lands on the list unasked.

The signup field sits below every blog post.

And because a newsletter should do more than redeliver what already exists, we are working on a small head start: subscribers will soon read new posts three days before everyone else. As a thank-you — not as a paywall.

Since this week, every note here carries its own dot-matrix image. What emerged almost in passing: a small series of works. Deterministically generated from a seed — same seed, same pixels, provably reproducible — and each work declared an edition of 1/1.

At “digital artworks, edition, authenticity” many people immediately think of NFTs. We asked ourselves the question seriously — and decided against. Because what an NFT technically proves is exactly three things: who issued a work, when, and who owns it. All three work without a blockchain:

Who: Every work has a provenance statement (title, edition, seed, SHA-256 checksum), signed with an Ed25519 key. The public key sits on this website — verification is a one-liner with standard tooling (ssh-keygen -Y verify).

When: The checksums live in our Git history, in web archives — and from now on in every entry of our RSS feed. Thousands of feed readers and archives thereby become independent witnesses to the timestamp. It is the old notarial principle of “notice in the newspaper”, just machine-readable.

Whose: Ownership changes hands via a signed transfer statement — the way the art market has traded photography and video art for decades: certificate of authenticity instead of token. Scarcity is guaranteed not by mathematics but by reputation. For a brand built on traceability, that is no disadvantage.

The complete catalogue — every work with seed, checksum and signature — is now available at analytikdata.ch/works.

And should tokenization ever become worthwhile: the signed statements are exactly the metadata a mint would need. We lose nothing by deciding later.

A short announcement in our own cause: these notes — and the blog posts — now reach you outside the big platforms as well.

Bluesky: @analytikdata.ch — our domain is the username there. The identity lives in the protocol, not in a purchased checkmark.

Fediverse/Mastodon: @aktuell@social.anadat.ch — not an account on someone else’s platform, but an instance of our own: a node of the network that we run ourselves.

RSS: /en/news/rss.xml for these notes, /blog/rss.xml for the blog — for everyone who never gave up on feeds. Rightly so.

There is nothing exclusive there and no conversion sermon. It is the same content, distributed by a small service on our own server — nobody types it out twice, and no algorithm reorders it.

Why we bet on channels that belong to us — and what that means in practice for a small company — will become its own blog post at the end of August.

The screensaver on this machine now shows our own brand mark: the dot-matrix A that tips over into a D column by column and back again, accompanied by a scan line — the same animation as the logo on our website, just in the terminal instead of the browser.

In between, the pictograms of our blog posts appear, each in its accent color:

The nicest detail: the screensaver doesn’t know these marks at all. On startup it reads them from the same system that generates our blog covers. When a new post appears, its mark runs across the screen from the next idle moment on — without anyone maintaining anything.

Effort: a good 250 lines of Python, built in one afternoon together with the AI agent. The video above isn’t filmed off a screen either — it is rendered from the same scene logic as the screensaver itself, so the two cannot drift apart.

Why the effort for an idle-screen image? Because the workplace is the most seen surface of a small company — in the café, in a meeting, on every screenshot. How far you can push this principle when the entire visual identity exists as code — more on that soon on the blog.

Perhaps you noticed — more likely you didn’t, and that is exactly the point: this website has no cookie banner.

Not an oversight and not a legal loophole, but a design decision. The banner is only mandatory where consent is required — and consent is only required where visitors are tracked. We track nobody: no advertising cookies, no fingerprinting, no third-party scripts.

What we still know: which pages get read, where visitors come from, which campaign works. That is measured by a self-hosted, cookie-less analytics solution — counted anonymously instead of tracked personally. For better content, that is entirely enough.

The pleasant side effect: the first impression of this website is its content, not a consent dialog.

How exactly this works technically and legally — and what you deliberately give up along the way — will appear soon as a full post on the blog.

By the way, we set up this same stack for clients: web analytics without a cookie banner.

Our ticker — the automatic trail of our own posts, security patches and press picks — got a page of its own: with day separators, filters (“patches only” works straight from a link) and a marker for what is new since your last visit.

And because news belongs wherever you want it, the feed now comes in every form: as RSS for the feed reader, as a newsletter in four flavors (per post, per note, daily, weekly — signup at the bottom of every page), via Fediverse and Bluesky — and as an app: open analytikdata.ch/ticker on your phone and you can install it on the home screen. Its own icon, its own window, works briefly offline too.

The part that surprised even us: the app was the smallest item of the evening. No app store, no approval procedure, no cut for a platform — the open web has shipped the entire app infrastructure for a long time; most companies just never use it. One manifest, one small script for the offline moment, done. The longer story follows on the blog.

Our visitor numbers are, frankly, still small. And that is exactly when something starts to matter: whoever builds their own website all day is its most loyal visitor. Every deploy check, every glance at a new post — all of it lands in the statistics and distorts them.

So: exclude ourselves. With classic analytics services there would be an opt-out cookie for that. Except — we deliberately track cookie-less (Umami, self-hosted, no personal profiles). No cookie banner for you, but also: no cookie to write an opt-out into.

The solution is almost disappointingly simple. The Umami tracker respects a flag in the browser’s localStorage:

localStorage.setItem('umami.disabled', 1);

For that there is now a small page, linked from nowhere: visit it once per device, done — the browser no longer counts. A button undoes it. And since the whole thing can only disable tracking, it is no secret either: analytikdata.ch/opt-out is open to anyone who would rather not be counted while reading. Fits the attitude.

Why we measure at all — and everything you can measure without surveillance (and what we deliberately don’t) — more on that soon on the blog.

While planning our AI cockpit, the uncomfortable question came up: what happens if the model behind it is discontinued — or if in a year we would rather have a different one? Maybe a Chinese one, maybe ChatGPT.

Our answer has two parts.

Part 1: The knowledge belongs to us, not to the vendor. Everything our agents need to know — decisions, runbooks, inventories, even the agent’s long-term memory — lives as Markdown in Git. No proprietary format, no vendor cloud. If today’s agent goes away, tomorrow’s reads the same files and carries on. Structure beats intelligence — at parting, too.

Part 2: The vendor becomes a config line. Anthropic’s API format has become a de facto standard: several providers (GLM, Kimi, DeepSeek) offer compatible endpoints, and translators like LiteLLM cover the rest (ChatGPT, Gemini). Our tools will therefore no longer point at the vendor but at a gateway of our own:

# Which model answers is decided by our config — not by the vendor
ANTHROPIC_BASE_URL=https://our-own-gateway

Honestly: there is never zero lock-in — the agent tools themselves come from one vendor, and every switch costs habituation. But it is the difference between moving house and expropriation. Own your stack also means: own your exit.

On this machine, almost every work session starts the same way: open a terminal, start the AI agent. If the first command of the day is always the same, it should cost exactly one keystroke.

That is why every new terminal now greets you with a small menu:

Type a digit, hit Enter, go. Technically this is unspectacular — and that is exactly the point. It is Bash aliases (yes, 1 is a valid alias name) plus a small function that shows the menu only in fresh terminals, not in nested shells (abridged here):

alias 1='claude --dangerously-skip-permissions'
alias 2='claude --continue --dangerously-skip-permissions'
alias 3='claude --resume --dangerously-skip-permissions'

quick_menu() {
  echo "Quick launch  (type ? to show again)"
  echo "  1  claude (skip permissions)"
  echo "  2  claude --continue (resume last session)"
  echo "  3  claude --resume (pick a session)"
}
alias ?='quick_menu'

[[ $SHLVL -le 1 ]] && quick_menu

About twenty lines of Bash. The agent wrote them itself, tested them (including the question of what shell nesting depth a fresh terminal under Hyprland actually has) and documented them in the inventory — effort: a few minutes of shouting instructions into the chat.

An honest note belongs here: key 1 starts the agent in autonomous mode, that is, without asking back on individual actions. On a machine of your own, encrypted, where everything important lives in Git, that is a deliberate trade of control for speed. You should make it deliberately — which is why it is stated here.

Hand-sized jobs like this come up daily for us. Individually they are too small for a blog post — hence this new section: short notes from the ongoing rebuild, unfiltered and in real time. The long stories will keep appearing on the blog.