Local AI: client documents never leave your office
We set up a language model on your own computer. It reads documents, extracts data, translates and drafts text — with no internet connection and no data passed to third parties.
Price
from€1,200
Timeline
14–21 days
Who it is for
Who this package is for
- You handle personal client data and cannot send it to ChatGPT — because of GDPR or professional confidentiality
- You work with documents in several languages: Finnish, Estonian, Russian, English
- You spend hours pulling data out of decisions, certificates and contracts by hand
- You tried cloud AI and your lawyer or compliance officer said no
- You handle GDPR special-category data: health, criminal records, religion, political views, sexual orientation
- You need processing with no per-document or per-token bill — the volume you run does not change what you pay
This is not for you if
- Your documents contain nothing confidential. Then a cloud subscription at around €20 per seat is simply cheaper, and we will tell you so during the audit
- You are counting on saving money. For an office of five to eight people, a local setup costs more over three years than cloud seats — hardware and our work are paid up front. You are buying permissibility, not savings, and that is a legitimate reason to say no
- You need AI for writing code — local models trail cloud ones by roughly 25 points on independent benchmarks, and no amount of configuration fixes that. You also expect AI to replace an employee: it speeds the work up, but a person checks every result
- Your machines have less than 32 GB of memory and there is no budget for hardware. Automated mailing or social posting with no human involved is also something we do not build — the reasons are below
What is included
What the engagement covers
A. Choosing the model for your hardware — measured by hand, not read off a spec sheet
- The model is chosen by measurement on your machine, not by leaderboard position
- Why this matters: in our own testing, the model with the best Russian-language score turned out to be four times slower than a smaller one — an architecture effect, invisible on paper
- Another model invented a Finnish authority called “Migraatio” that does not exist. Both failures show up only when you measure
B. Installation and configuration on your equipment
- Model, engine and a browser interface — running with no internet connection
- Parameter tuning for your tasks (one setting, in our testing, took the same job from 2.5 minutes to 7 seconds)
- A system prompt for your field — removes needless hedging and improves accuracy
- Automatic conversion of PDF and Word files into text
- Scan recognition with Finnish, Estonian and Russian support
C. Testing on your documents — real ones, not demo samples
- We run real documents from your practice through the setup
- We measure speed and quality
- We find and write down where the model gets things wrong on your document types
D. Documented limitations — the main thing you are buying
- A list of the tasks where the model is reliable
- A list of where it fails, with examples
- What must always be checked by hand: names, dates, reference numbers, sums, citations of law
- Ready-made prompts for your tasks — copy and paste
E. Training your team
- A two-hour session: how to use it, and how to recognise when the model is wrong
- A briefing and a dated one-page policy — this satisfies the AI-literacy requirement of AI Act Article 4
- A written reference sheet
F. The legal side
- A template AI usage policy
- Wording for your privacy policy
- A walkthrough of what changes under GDPR with local processing (no data processing agreement is needed — there is no transfer to a third party)
How we work
How the work runs
Audit, day 1
What we do
A 90-minute conversation: which documents, which tasks, where the time actually goes
What we need from you
An hour and a half
Audit, day 2
What we do
Hardware check: is there enough memory, what would need buying
What we need from you
Access to the machines, or their specifications
Audit, days 3–4
What we do
Testing models on your documents — two or three candidates, measured for speed and quality
What we need from you
Five to ten typical documents; anonymised is fine
Audit, day 5
What we do
The report: what will realistically work, what will not, the costing and the implementation plan. You can take the report and go to any supplier with it. Some clients we will tell plainly that they do not need this
What we need from you
Read it and ask questions
Implementation, week 1
What we do
Installation, configuration and measurements on your hardware
What we need from you
Access to the machine, two to three hours
Implementation, week 2
What we do
Testing on your documents, tuning prompts for your tasks, hunting for the failure modes
What we need from you
Fifteen to twenty documents, and feedback
Implementation, week 3
What we do
Documentation, staff training, legal templates
What we need from you
Two hours for the session, everyone who will use it
Ongoing support, monthly
What we do
Models are replaced every six to ten weeks. Without support, a setup goes stale within a year
What we need from you
Nothing routine — you get in touch when something comes up
What you get
What stays with you
A working setup on your own equipment
The document “what it does well, where it fails, what to check” — written for your practice
A set of ready-made prompts for your tasks
A dated AI usage policy — satisfies the AI Act requirement
Wording for your privacy policy
A reference sheet for your team
A measurement report: how many seconds each routine task takes
All of it stays with you. If you leave, nothing switches off — the software is open source and the model sits on your own drive.
Results
Results — measured, not promised
7–15 sec
Extracting data from a document, against 10–15 minutes by hand. Measured, not estimated.
14 sec
Translating two or three pages from Finnish or Estonian. A working translation for understanding — not a certified one.
15–30 sec
Summarising a twenty-page document.
30–50 sec
A draft letter to a client, roughly 70 per cent of the way there.
Important note
We do not promise that AI will replace an employee, or that there will be no mistakes. The model does get things wrong — in our own testing it confused Finnish authorities and invented names of institutions. What we are accountable for is that the setup works, that you know its limits, and that your people know how to check it. Responsibility for the final document stays with you, exactly as it would if an assistant had drafted it.
Not included
What we do not build — and why
- Automated sending of email without human approval. The model is wrong confidently — a correct answer and an invented one look identical. There has to be a person between “the model wrote it” and “the client received it”, because a sent letter with an invented date cannot be recalled. Legally it also protects you: mandatory human review removes the AI-disclosure duty under AI Act Article 50(4), keeps the system out of the high-risk category, and breaks the chain of causation in any claim
- Publishing to social networks. LinkedIn explicitly prohibits automated post creation — section 8.2 of its user agreement — and it is the client's account that gets blocked. On top of that, automated posting needs cloud APIs, which defeats the entire point of a local installation
- An autonomous agent that decides for itself what to do. Measured: models in this class complete a single tool call with 86 per cent accuracy, and a multi-step chain with 30 per cent. Cloud models manage 68 per cent. That is the difference between “sometimes gets there” and “usually gets there”. You cannot sell autonomy on those numbers
- Replacing a lawyer, a doctor or an accountant. The tool prepares a draft. The specialist makes the decision and carries the responsibility for it
All of the above is technically possible. We decided against it deliberately — this section matters more than the list of features.
Price
Pricing
- Minimum term for support is 3 months, then month by month, cancelled in writing with 30 days' notice
- Hardware requirement: at least 32 GB of memory. A model of the required quality will not run on 16 GB. If the hardware is not there, we cost it during the audit — usually a MacBook Pro or a mini PC from €2,500
- Being straight about the support price: we costed it from real load — 23 to 35 hours per client per year for model updates, troubleshooting and advice. At the basic rate that breaks even or a little better. We will not go lower: a subscription that does not pay for itself ends with the supplier no longer answering email
- The audit is credited in full against implementation. If the audit says no, you have paid for the audit and nothing else
Straight about cost: a local setup is not the cheaper option. A ChatGPT Team seat is about €25 a month, and over three years cloud seats for a small office come to less than hardware plus our work. What you pay the difference for is that client data never leaves the office, that no supplier can switch you off, and that volume never appears on an invoice. If your documents are not confidential, take the cloud subscription — it is cheaper and we will say so.
Guarantee
What we guarantee
We sell a setup service, not a software product. The guarantees therefore cover our work, your right to walk away, and honesty about what the model can and cannot do.
If the audit shows it will not work, we say so plainly. The report stays with you and you pay for the audit only
If it does not work as agreed in the first month, we return the cost of implementation
Everything is yours. The software is open source, the model is on your drive, the documentation is in your hands. You leave with no losses
No traps. Support is month by month after the first three, cancelled by email with 30 days' notice
We write the limitations down honestly. The document setting out where the model fails is part of the delivery, not something you discover later
Questions
Common questions
You can check it in thirty seconds: switch off the Wi-Fi and ask a question — the model answers. Software that needs the internet does not work without it. The model is a file on your drive and the computation happens inside your computer. This is a test you can run in front of your own client if they ask.
In three ways. The data never leaves your computer, so you can process client documents that a data processing agreement would not cover. There is no per-use charge — as many documents as you like, at no extra cost. And the model is weaker than a cloud one: for legal texts and document work the gap is small, for writing code it is substantial. What this is not is the cheaper option — on a three-year view a cloud subscription for a small office costs less.
You can — the installation takes two clicks. What you are paying for is not the installation. It is choosing the model for your hardware with real measurements, tuning the parameters, and the document that says where it fails on your documents. From our own testing: the model with the best benchmark score was four times slower than the one we needed, and another invented names of government authorities. Neither is visible until you measure.
It will — the only question is when. That is exactly why we document where it happens and train your people to check. Always checked by hand: names, dates, case numbers, sums, citations of law. The tool produces a draft; responsibility for the outcome stays with you. Automated sending without human review is something we will not build, on principle.
With local processing there is no transfer to a third party — legally it is much like opening the document in Word. Separate consent is usually not required and no data processing agreement is needed. Adding a couple of sentences to your privacy policy is still sensible practice, and we supply the wording. On professional confidentiality in your own jurisdiction we recommend a one-off consultation with a local lawyer — that is the single area where we do not offer an opinion.
Yes, it reads and translates both. One caveat from practice: Finnish surnames in inflected forms are sometimes mangled in translation — “Petroville” can come out as “Petrovich”. In data extraction the name carries across correctly. This is in the documentation, and names in translations are checked by hand.
Models are replaced every six to ten weeks, and newer ones are usually better at the same size. With support we handle the updates and test them on your documents before switching. Without support the setup keeps working, but it gradually falls behind. What cannot happen is the thing that happens to cloud services: in December 2024 the accounting provider Bench Accounting shut down overnight and its clients lost access to their own books the same day. A local model is a file on your drive — nobody can switch it off, and no price change or discontinued plan reaches you.
Start with the audit
In five days we check your hardware, test models on your own documents and tell you straight whether this will work or not. The report stays with you either way.