The problem I saw in Slovenia, and had been seeing in Formentera for two years

In March 2026 I went to Celje, Slovenia, for my Erasmus internship. The company is called CETIS Flex and it makes commercial packaging. My brief was clear: a web app that worked out quotes for orders.

They gave me one number on the first day. Quoting an order took three to four days.

I thought what you're thinking. The maths must be complicated.

It wasn't.

The three days had nothing to do with maths

What took three days wasn't the calculation. It was finding the person who knew how to do it.

The price of a package depends on a lot of things: the material, the ink, the quantity, the speed of the machine that runs it. None of that is hard on its own. The hard part was that the rules for putting it together lived in the heads of two people with twenty years there.

When an unusual order came in, the sales rep didn't open a document. He asked. Then he waited.

If that person was on holiday, the quote waited until they came back.

I built the backend in Python with FastAPI and the frontend in React and Next.js. I added a 3D view with Three.js so the client could see the box before confirming. That was the visible part. The hard part was something else: sitting with those people and getting the rules out of their heads so I could write them down.

Three or four days turned into seconds. But speed wasn't what got fixed. What got fixed is that the quote stopped depending on someone being available.

I came home and started seeing it everywhere

When the internship ended I went back to Formentera and looked over the work I'd done before. That's when I realised I'd been seeing the same thing for two years without giving it a name.

At Tisa, in Ibiza. I worked on three projects as a developer. One was Ibiza Circular, the Consell's official platform for handling traffic permits and fines with automatic plate reading. Every time I needed to understand a business rule, the answer started the same way: "ask so-and-so, he handled that one".

At Love Formentera, an estate agency that also rents boats. I modernised the website, ran the social accounts, shot photos of properties and boats. Which boat fits which mooring, which property takes which kind of booking, who to warn in each case: none of it was written down. It lived in two heads.

At Love Bici, the electric bike shop. I digitised the rental contracts and rebuilt the online store in WooCommerce, mobile-first. I had to reconstruct the group booking process by asking, because the only complete version of how it worked was in the owner's memory.

At LucasRe, the island's only luxury watch shop. I ran partnerships with high-end restaurants. Which restaurant suits which type of client wasn't written anywhere. It came from years of dealing with people.

Four businesses of different sizes, different sectors, different countries. The same bottleneck. Only the name changed.

It isn't an AI problem. It's a memory problem

This is the line I keep coming back to.

The companies I work with don't have an artificial intelligence problem. They have a memory problem. The knowledge that makes the business run isn't in any system. It's spread across a handful of people and passed on by talking.

Look at it this way:

• Meetings are shared memory, forced.
• Documentation is written memory, when someone finds a spare hour.
• Holidays are memory that walks out the door.
• Onboarding is memory rebuilt from scratch.
• Processes are memory you can run.

And AI, set up properly, is memory you can ask questions.

That's the gap. Nobody is short of a better model. They're short of something to feed it.

What it costs, and why it never shows up on an invoice

This cost is hard to see because it never arrives as a bill.

When an answer takes three days, the order isn't cancelled. It's delayed. When the senior goes on holiday, the team doesn't stop. It slows down. When someone leaves after eight years, no line on the balance sheet says what they took with them.

A client told to wait three days for a price doesn't always complain. Sometimes they ask for that price somewhere else, and then they really are gone. That one shows up in no report, because it never became an order.

Nobody measures what never happened.

What I did with that

I started building Savia while I was still working freelance.

The idea is the one you've already read between the lines. Take the knowledge that today lives in your best employees' heads and make it available around the clock to the whole team. With a chat that always cites where each fact comes from, instead of inventing it. With the knowledge split by department, so each person only sees their own.

The first thing I did was use it myself. My information was scattered across four different places and I was my own bottleneck.

I wrote about how it works inside somewhere else, because it fills a whole article on its own.

What I took away from those three days

I went to Slovenia thinking I was going to automate a spreadsheet.

I came back with something else. With a hunch that almost every small and mid-sized company I'd work with had the same crack. And that almost none of them called it by its name.

That factory had been running well for years. They had clients, they had product, they had good people. Even so, an order could sit for three days because the person with the answer was away.

He's still the one who knows most. He's also entitled to take a holiday.