Start with the fair part. Onyx, PrivateGPT and their siblings are serious open-source projects, and the public AI models behind them are good now. A capable engineer can build a private question and answer system over a weekend and demo it on Monday.
Most build or buy decisions get made in that demo. It is the most misleading half hour of the project.
The demo is the easy part
A demo on ten PDFs looks like the system your estimating office leans on every day. Between them sits the boring list.
- Loading your real documents. Scanned PDFs, spreadsheets with merged cells, fifteen years of names nobody agreed on. Every file that loads wrong is a question that comes back empty.
- Making search find things. Whether steel pipe 57x3.5 finds its eleven other spellings in your catalogues is work, and the work is per set of documents.
- Measuring it. Without real questions and a pass mark, you only learn how impressive it feels. Building that test is real work.
- Keeping it alive. Model updates, drivers, backups, and someone to call when answers stop at 9 a.m. on tender day.
- The person who leaves. The stack lives in one engineer’s head. Their resignation letter is your migration project.
The open-source projects give you the engine. You build the car around it.
Where the line actually sits
Our rule of thumb, and we sell the alternative, so discount it: building pays at about three AI engineers of your own. At that size you can afford the loading work, the test, someone on call and a spare head. You get a system shaped to your way of working.
Below that line the arithmetic turns against you quietly. One engineer doing it on the side means the stack competes with the work you sell. Those costs never reach an invoice. They show up as the feature that took a quarter.
Cloud AI seats run $30 to 90 per user per month, forever, on published list pricing as of mid-2026. Document AI suites for big companies start at $100k+ a year, scoped for Fortune 500 buying departments. A DIY stack escapes both. The question is whether you escape into payroll.
The appliance answer
Perimeter is for companies below that line. It runs the same public AI models a DIY stack would, on a box in your server room, set up on your documents. It never talks to the internet.
What is worth keeping from DIY, you keep.
- Ownership. The machine is bought in the EU in your name. Your documents sit on the box, inside your walls. We keep no copies.
- Public models. Anyone can look inside them. When a better one ships, we put it on your box.
- You can check it. Pull the network cable. It keeps working.
What is a second job, you hand over. Loading your formats, making search work, the test, the load check. And a pass mark agreed in writing on your documents, before any licence is invoiced. Running in 30 days, and you hire no one.
Which one is you
Three engineers to spare and a way of working no product fits? Build. You will own every screw.
Documents that cannot leave the building, answers that show their page? The engineering project is the expensive way to buy a box.
The one page version, including the rows where cloud and DIY win, is the comparison matrix.