Skip to main content

Local AI Setup

Your documents never leave your office

We configure AI on your own hardware — a setup service for law firms, clinics, accounting practices, and any business that handles data it cannot send to the cloud. Estonia and Finland.

Order an audit

The problem

Tools like ChatGPT are useful, but every document you paste into them travels to a server you do not control. For client files, patient records, and payroll data, that is often not an option — by contract, by professional rules, or simply by common sense. The result is that the firms with the most paperwork are the ones least able to use AI on it. There is a practical alternative: run the AI model on a computer inside your own office, so the documents stay where they already are.

How it works

The document stays on your machine at every step

01

Your document

A contract, a medical record, an invoice — whatever you work with. It stays in your office, on your storage.

02

Your computer

A language model (an LLM — software that reads and drafts text) runs on hardware you own. We select and configure it; the software is open and widely used, we do not sell it.

03

Your result

A summary, extracted fields, a draft — produced on the same machine. Reviewed by your staff before it is used.

Notice what the diagram does not have: an arrow leaving the building. Nothing is uploaded, and the setup keeps working with the network cable unplugged.

Why us

Installing a local model takes an afternoon. Knowing where it breaks takes months.

01

Leaderboards do not predict your hardware

A model that topped the public benchmarks for our working language ran four times slower than acceptable on the specific machine we tested it on. Nothing in the rankings tells you this — only a measurement on the actual hardware does. That measurement is part of every setup we do.

02

One setting: 7 seconds instead of 2.5 minutes

On the same task, on the same machine, a single configuration parameter made the difference between two and a half minutes and seven seconds. Defaults are rarely right for a given machine, and finding this out is exactly the kind of work you are paying for.

03

Some free tools only pretend to work

A popular free program we evaluated displayed convincing file edits on screen while writing nothing to disk. It looked finished; it wasn't. We test what actually landed in the files, not what the interface reports.

04

Finnish name forms trip models up

Finnish inflects names, and models misread the inflected forms — turning one person's name into a different name entirely. In a legal document that is not a cosmetic error. We test for it and document a verification step for names and dates.

We run this setup ourselves, on our own paperwork. A recent example: a decision letter from the migration authority, parsed on a local machine in 7 seconds — every relevant field extracted, including the appeal deadline computed from the dates in the document. Then checked by a person, which is how we teach every client to use it.

Honest limitations

What local AI does not do well

  • 01

    It is slower than cloud services. Cloud providers run data centres; you run one machine. For document work the difference is workable, but it is real.

  • 02

    It is noticeably weaker for programming — roughly 25 points behind leading cloud models on independent agentic-coding benchmarks. If code is your main use case, local is the wrong tool.

  • 03

    Names, dates, and figures must be checked by a person before use. We document where errors occur and train your staff to verify — we do not promise accuracy no model can guarantee.

  • 04

    It needs suitable hardware. Some offices already have it; some need a purchase of roughly €1,500–3,000. The audit answers this before you spend anything.

Frequently asked questions

Start with the audit. If local AI is not right for you, the report will say so.

Order an audit — €1,200