LEX SUPREMA · On-premises AI platform
AI on your case files.
Without those files ever leaving your network.
A fully sandboxed AI platform for law firms, municipalities, healthcare institutions and privacy practices, delivered and implemented complete. Model weights, vector index, documents and logs all sit on equipment under your own control. Outbound traffic is refused at firewall level — there is no technical route to an external inference API. Whatever area of law or regulation you need, we build it in.
The question everyone asks
“Am I even allowed to put this case file into an AI service?”
In practice the answer is almost always no — and so your organisation misses the productivity gain the rest of the market is taking. Not because AI is unsuited to legal work, but because the question of where the data sits goes unanswered.
The file is too large for the deadline
Two thousand to twenty thousand pages, a hearing date, and a team that has to work through it by hand. The inventory alone costs days that are not spent on substance.
The duty of confidentiality rules out the cloud
For many organisations, putting client data, patient records, police data or Wjsg material into a public AI service is not a trade-off but a prohibition — contractual, disciplinary, or written into a tender.
Your own knowledge stays unfindable
Years of advice, pleadings and precedents sit on network drives where nobody can find them again. Precisely the knowledge you do not want to send to an outside service.
How this differs from a cloud subscription
Three properties a subscription cannot deliver
Data residency by design
Model weights, vector index, documents, prompts, answers and logs all sit on your own equipment. The inference and data zones have no outbound route; a connection attempt is blocked and alarmed. That is a network rule, not a setting inside an application.
Need-to-know down to the document
A built-in IAM system with relationship-based authorisation filters retrieval before anything is fetched, not after. What an employee may not see does not appear in their search results and never enters the model’s context. Separation between matters and between teams is enforced in the retrieval chain, not just in the interface.
One platform, several projects
Case analysis today, contract review tomorrow, an internal knowledge assistant the week after. New models, sources, knowledge layers and workspaces are added as configuration — no reinstallation, no downtime for other workspaces, and no mixing of indexes.
Zoning
There is no network route from the case files to the outside
Traffic runs one way: from the user to the gateway, which decides per request what the inference zone gets to see. The processing zone only reads from the source systems and writes to the data zone.
Application
What it takes over in practice
The system produces drafts; the qualified professional reviews and settles them. It takes over the preparatory analysis so that lawyers get to the judgement work.
Case file inventory
A delivered file of 2,000–20,000 pages is indexed, given document numbers and reconstructed chronologically. Days of manual work become hours of checking.
Evidence matrix per element
Per element of the offence: the evidence, its source, the prosecution’s position, the defence’s position, reliability, probative weight and the gaps — side by side.
Procedural audit
Systematic checks on the caution, consultation and interrogation assistance, powers, warrants and time limits, distinguishing a possible defect from an established one.
Contradiction detection
Times, locations, shifting statements and digital traces are compared across the entire file and classified by legal weight.
GDPR subject access request
Compiling, assessing and redacting a disclosure file, including determining which regime applies: GDPR, the Wet politiegegevens or the Wjsg.
Internal knowledge retrieval
All earlier advice, pleadings and precedents searchable with source references — without that knowledge leaving the firm.
Customisation and extensibility
Criminal law is where we started, not where it stops
The knowledge layers are version-controlled instruction layers, not hard-wired software. A new area of law, a sectoral regulation, an internal guideline or your own practice’s working method is added as an extra layer — no reinstallation, no downtime for existing workspaces, and no mixing of indexes. Any body of legislation you can supply as text and case law can be added.
Legislation that can be added as a knowledge layer
Not exhaustive, and none of these layers is included as standard: you choose the ones you need and pay only for those. Anything that can be supplied as statute, decree, guideline and case law can be indexed and built into the working method.
Codes
Grondwet · Burgerlijk Wetboek, books 1–10 · Wetboek van Strafrecht · Wetboek van Strafvordering, old and new · Wetboek van Burgerlijke Rechtsvordering · Algemene wet bestuursrecht · Algemene wet inzake rijksbelastingen · Wetboek van Koophandel · Faillissementswet
Criminal and sanctions law
Opiumwet · Wet wapens en munitie · Wegenverkeerswet 1994 · Wet op de economische delicten · Penitentiaire beginselenwet · Wet forensische zorg · Wet DNA-onderzoek bij veroordeelden · Wet internationale misdrijven · WETS and WOTS · Sanctiewet 1977
Data protection
GDPR and UAVG · Directive 2016/680 · Wet politiegegevens · Wet justitiële en strafvorderlijke gegevens · Telecommunicatiewet · Wiv 2017 · Archiefwet · Wet open overheid · Wet digitale overheid · Wabvpz
Government and public administration
Gemeentewet · Provinciewet · Omgevingswet and the framework beneath it · Participatiewet · Aanbestedingswet 2012 · municipal and provincial by-laws · mandate and delegation decisions · policy rules
Healthcare and social domain
Wmo 2015 · Jeugdwet · Wvggz · Wet zorg en dwang · Wet BIG · Wkkgz · Wgbo · Zorgverzekeringswet · Wlz · employment law from BW 7:10, WAB and the Arbowet
Financial services and supervision
Wet op het financieel toezicht · Wwft · Pensioenwet · DORA · MiCA · DNB and AFM policy rules · Wet toezicht trustkantoren
European law
ECHR and the Charter · AI Act · NIS2 and the Cyberbeveiligingswet · CER and the Wwke · Data Act · Data Governance Act · DSA and DMA · e-Privacy · EUR-Lex sources
Professional conduct and disciplinary law
Advocatenwet · Verordening op de advocatuur and the conduct rules · Wet op het notarisambt · Gerechtsdeurwaarderswet · disciplinary case law · the Public Prosecution Service’s Aanwijzingen and Richtlijnen voor strafvordering · LOVS sentencing guidelines
Case law and standards
Supreme Court, courts of appeal and district courts by ECLI · Raad van State · CRvB and CBb · ECtHR and CJEU · decisions of the Dutch DPA · parliamentary history · Staatsblad and Staatscourant · BIO, NEN 7510, ISO 27001 · collective labour agreements and control frameworks
On top of that, your own material as a layer: internal guidelines, templates and decision trees, house conventions for court documents, tender conditions. And beyond law: contract review, procurement and due diligence, an internal knowledge assistant or research data, each in its own workspace with its own authorisation group, index namespace and retention period.
Flexible and fast: weeks, not years
The platform has four extension points — the model registry, the source connectors, the knowledge layers and the workspaces — and an addition at one point does not touch the other three. That is why a new area of law is not a new project and a new source system is not a migration. If your integration, area of law or working method is not listed here, that is a question for the feasibility study, not a reason to walk away.
| New workspace for a separate project | days |
| Source connector on a known system type | 1–2 weeks |
| Your own guideline or template set as a layer | 1–2 weeks |
| New area of law including source import | 2–6 weeks |
| New model in production, tested and released | 1–3 weeks |
Typical lead times once the platform is in place, with sources supplied. Substantive validation runs in parallel and determines the finish date.
What customisation does not solve
The quality of a new layer stands or falls with verified sources and an evaluation set that a specialist in that field compiles and checks. For every new area of law that is real effort, and it is the reason we do not deliver a layer on the basis of a statutory text alone. We budget that validation separately and make no commitment about it that we cannot measure.
Quality mechanisms
Every assertion with a source you can click through to
Hallucination is the greatest substantive risk in legal AI. Here it is limited, not eliminated — which is why five overlapping measures run alongside every answer.
What cannot be verified is flagged or removed. What is missing is named as a gap instead of being filled in.
- Mandatory sourcingEvery factual assertion refers to a document number, page, date and passage. Without a source: labelled unconfirmed.
- Citation verification after the factA separate checking step tests whether every quotation and every ECLI actually appears in the indexed source.
- Counter-analysisAn independent red-team step asks a fixed set of questions about missing facts, unproven assumptions and over-confident conclusions — before any final advice.
- Refusing over guessingWhere source material is insufficient, the system names the uncertainty, its cause and a concrete verification step.
- Continuous evaluationAn in-house, growing evaluation set of Dutch legal questions runs at every model change, so that quality regression becomes visible.
What we stand behind without reservation
Seven commitments, demonstrable on request
These points can be demonstrated and tested at handover. Anything we cannot prove is not listed here.
- 01 Model inference takes place on equipment under the client’s control.
- 02 There is no network route from the inference environment to external AI services.
- 03 Documents are not used for training unless you explicitly instruct us to.
- 04 Every search is filtered against the authorisations of the logged-in user.
- 05 Every question, the sources used and the answer land in an immutable audit trail.
- 06 The platform runs on open standards and remains usable without the vendor.
- 07 Data is stored encrypted and transported encrypted.
Honest about the limits
And what it expressly is not
These limitations belong in the sales conversation, not in the small print. You will meet them within three months.
Who it is for
Five people decide on this, each on a different argument
Your case files do not leave your network — and for every file you can show who saw what
Evidence supplied: an audit example and the DPIA package.
Standard components, open standards, your own IAM, full logging and an exit without the vendor
Evidence supplied: network architecture, zoning and patch policy.
Privacy by design here is not a claim but a firewall rule; you get the DPIA building blocks with it
Evidence supplied: input for the record of processing activities and the retention periods.
Every sentence with a source reference you can click through to; what cannot be verified, the system does not say
Evidence supplied: a demonstration on a wholly fictitious but realistic case file.
Full digital sovereignty, aligned with your obligations under the Cyberbeveiligingswet
Evidence supplied: a compliance overview per obligation.
Typical buyers: criminal defence firms (5–60 FTE) · general practice and civil-law notaries (20–250 FTE) · privacy and DPO practices (1–30 FTE) · municipalities and safety regions · healthcare institutions · financial institutions and insurers · government and defence suppliers.
Three sizes
Delivered complete: hardware, installation and implementation included
One order, one supplier, one price. The machines are delivered, installed, hardened and configured, the source systems are connected, the knowledge layers are activated and your people are trained — right through to handover to your own administrator. You provide the room, the power, the network ports and a point of contact. The step between sizes is not about compute but about GPU memory and about moving from one server to a cluster; growing on is possible without data migration or rebuilding authorisations.
A single 2U server with two 96 GB GPUs. Runs a 70B model at 128k context plus embedding, reranker and OCR. For a criminal defence firm or privacy practice with a few hundred thousand to two million pages.
No high availability: if the server fails the system is down until a part is replaced.
indicative € 60,000 – € 95,000
All in, over five years, that works out at € 300 – € 400 per user per month at 15 users. Here you are buying data residency, not a price advantage.
A small cluster with 768 GB of GPU memory, around 60 TB of usable storage and redundant nodes. Runs a 70B main model with ample concurrency alongside a workhorse model, OCR and a second AI project.
Losing one GPU server halves capacity but brings nothing down.
indicative € 270,000 – € 450,000
All in, over five years: roughly € 297 per user per month at 60 users, roughly € 149 at 120. A fixed amount, however intensively you use it.
An HA cluster with an NVLink-coupled GPU platform, 300–500 TB of distributed storage, a redundant spine-leaf network and a second site with asynchronous replication. For several AI projects running side by side.
Automatic failover; fallback on loss of a site. Requires a genuine data centre environment.
indicative from € 900,000
All in, over five years: under € 150 per user per month at 300 users, and falling further with every user added.
What is included. Hardware, delivery, installation, hardening, platform configuration, source integrations, knowledge layers, evaluation set, training, documentation package and handover. Not included: your server room or colocation, power and cooling, and ongoing administration after handover — that is either self-managed or a separate support contract. Amounts are indicative orders of magnitude; a binding figure follows from a quotation on your configuration. Component lists and the full cost model are in the specification.
How this compares with a subscription
The comparison that matters is not the investment against zero, but your five-year cost against five years of subscription fees — for the same number of lawyers, with the files held by a third party.
| Option | Pricing model | Per user per month | Files are held by |
|---|---|---|---|
| Legal AI assistant, Dutch market | Per-seat subscription | around € 59 | the provider |
| Legal AI platform, Dutch market | Per-seat subscription | € 175 – € 200; market-wide € 100 – € 300 | the provider |
| International platforms | Per seat, minimum 10–25 seats | estimates from $ 250 to $ 1,200 and above | the provider |
| LEX SUPREMA, Tier 2 | Owned outright, fixed cost | around € 297 at 60 users · around € 149 at 120 | you |
Your bill does not rise with your usage
The market is moving to billing on actual consumption: in June 2026 Legora announced that shift for its most advanced product, and observers expect other large providers to follow. Firms that want to keep using the best models are budgeting for AI costs two to three times higher. On your own hardware your bill does not double when your usage doubles — you buy capacity, not tokens.
Where it turns
Somewhere between forty and eighty intensive users, ownership becomes cheaper than a platform-level subscription. Below that you are buying this for data residency and demonstrability, not for the price — and we would rather say so now than in year two.
Trial installation on site
If you want to see it working on your own material first: we place a Tier 1 machine in your rack, connect one source system and index one archive. Six weeks at a fixed price, set off against the project price if you proceed. If you do not proceed, the machine comes back with us and your data stays where it was.
Market figures: rate indications per user per month from research by ICTRecht, as reported by Mr. Online, and the € 100 – € 300 range plus the announced move to consumption billing from reporting by the FD, 29 June 2026. Figures for international providers are third-party estimates; those vendors do not publish their rates. Our own per-user figures are derived from the cost model in the specification and change with your user count and your choices.
Implementation
Six to twelve weeks from order to production
The critical path is almost never the technology, but GPU lead times and getting access to the server room. That is why every project starts with a site survey — before the quotation.
Feasibility study
Data sources, user numbers, room, power, and choice of size.
Design
Architecture, authorisation model, DPIA input and acceptance criteria.
Delivery and hardening
Hardware ordered, installed, connected and hardened.
Platform and first pipeline
Orchestration, IAM, storage, monitoring, backup; one source fully indexed including ACLs.
Substantive tuning and pilot
Knowledge layers active, evaluation set built, 5–10 users measuring and adjusting.
Rollout and handover
Remaining sources, all users, training, documentation and transfer of administration.
Six acceptance criteria, agreed in advance
You do not sign off on a demonstration but on a test.
- In no scenario does a user see a document they have no access to in the source system — demonstrable with a test script.
- Every factual assertion in an analysis carries a source reference that leads to the right page.
- An outbound connection attempt from the inference zone is blocked and alarmed — demonstrable with a test.
- The five audit questions can be answered within five minutes.
- Restore from backup has actually been performed once and the recovery time measured.
- The evaluation set reaches the threshold agreed with you.
Compliance · position as at 14 October 2026
Aligned with obligations that already apply
On-premises simplifies compliance considerably but does not remove it: you remain the controller, and the obligations around legal basis, DPIA, security and transparency apply in full.
GDPR
No processor for the inference itself, no transfer outside the EEA. Data minimisation is enforceable through retrieval scoping; you receive technical descriptions per pipeline and a DPIA template with the technical sections filled in.
Wpg and Wjsg
Police data and judicial data have their own regime. The platform determines which regime applies per data collection and actively prevents two common mistakes: applying GDPR access rules to Wpg data, and mixing both regimes in one index.
EU AI Act
The Digital Omnibus, Regulation (EU) 2026/1744, entered into force on 27 July 2026 and moved high-risk Annex III to 2 December 2027 and Annex I to 2 August 2028. Prohibited practices, AI literacy and transparency already apply. The deferral affects the obligations, not the duty to classify under Article 6.
Cyberbeveiligingswet
The Cbw entered into force on 15 August 2026, with no general transition period. Registration, duty of care and the notification duty have applied in full since then — with an early warning within 24 hours. Through supply-chain responsibility this reaches your suppliers as well.
Documentation package at handover
System and network architecture with zoning and data flows · technical and organisational measures mapped to Article 32 GDPR · input for the record of processing activities per pipeline · DPIA template · authorisation matrix and role descriptions · logging specification and retention periods · a model card per deployed model covering provenance, licence, version and known limitations · exit and data portability plan.
Self-selection
When you are better off not buying this
On-premises rarely wins on pure cost at small user numbers and almost always wins at large ones. The buying argument is not “cheaper” but manageable risk at predictable cost. The turning point sits roughly between forty and eighty intensive users.
- Below twenty users with irregular usage, an EU-hosted service with a sound processing agreement is usually cheaper and operational sooner.
- Where volumes swing sharply and are peak-driven, fixed capacity is expensive.
- With no internal IT capacity at all, administration becomes the largest hidden cost — reckon on 0.3 to 1.0 FTE, or a support contract.
- Anyone who always wants the very best model is by definition accepting a gap behind the newest closed models.
Common objections
The questions that always come
“Open models are less good though, aren’t they?”
On the very hardest reasoning tasks, that is true. In case analysis with retrieval the difference is far smaller than expected, and the human stays in the loop: the system delivers drafts, not finished products.
“In a year the hardware will be obsolete.”
GPUs are usable for four to five years; plan replacement in year three or four. The platform is model-independent, so new models run on the same machine.
“The cloud is secure too, isn’t it?”
It may well be. The difference is not security but control and demonstrability: who actually holds access to the data.
“We have no IT department.”
Then a support contract belongs in the quotation. Remote administration runs over a secured, logged connection; administrative access is technically separated from substantive access, and case data is subject to the four-eyes principle.
“Our data is a mess.”
Familiar and normal. A search system makes existing access problems visible and worse, so cleaning up authorisations is an explicit project phase — before indexing, not a precondition beforehand.
“What if you go out of business?”
Open standards, hardware and data in your own possession, documentation and an exit plan at handover. No key held by the vendor.
“Our professional rules do not allow this.”
The professional rules do not prohibit AI; they set requirements for confidentiality and diligence. This design is aimed precisely at those. The lawyer’s own assessment remains.
“We will wait until the rules are clear.”
The high-risk obligations under the AI Act have been deferred, but classification, AI literacy and transparency apply now, and the Cyberbeveiligingswet has applied since 15 August 2026 with no transition period.
Next step
Start with the feasibility study, not with the quotation
One to two weeks: an inventory of your data sources, user numbers, server room and power, a first picture of authorisations, and the choice of size. Only then does a substantiated quotation follow — including an honest answer to whether on-premises is the right choice in your situation.
Privacy Solid
Daniel de Rijk — CDPO, Data Protection Officer and cybersecurity consultant. Privacy law and Linux infrastructure in one pair of hands: the DPIA and the firewall rule come from the same desk.
[email protected]
privacysolid.com
The Netherlands