You have read the recommendations from the NOvA, the Dutch bar association. You know that 68 percent of your fellow lawyers work with generative AI every week. You see firms around you testing tools, running pilots and talking about "AI policy" more and more often. And still you are walking around with the same question we hear in almost every first conversation: where do I start, and how do I make sure it is safe and worthwhile?
That question makes sense. The legal profession is one of the few sectors where the wrong AI choice is not just inefficient but could breach professional privilege. That sets the bar high, and rightly so. But it makes it all the more important to choose the right approach instead of doing nothing at all.
At Green Creatives we build AI systems for knowledge intensive organisations, including law firms working in criminal law, administrative law, property law and employment law. In this article we share what we run into in practice: where firms get stuck, which applications really work, and how you implement AI without compromising on confidentiality.
Why AI for law firms is urgent now
The pressure on law firms comes from several directions at once. Clients expect shorter turnaround times and more transparency about costs. The job market for lawyers stays tight. And meanwhile the volume of data per case keeps growing: digital correspondence, bodycam footage and ever larger court documents.
At the same time something structural is shifting. In house counsel now ask how their law firm uses AI. They want to know whether there is transparency about that use, and whether their data is being used to train models. That expectation is not going away, it is becoming the norm.
At the end of 2025 the NOvA published official recommendations on the use of AI in the legal profession, structured around the five core values: independence, partisanship, expertise, integrity and confidentiality. That is not optional advice. It is the framework every firm now has to shape its AI strategy within.
Lawyers who still read and search everything by hand are falling behind. Not next year, but now. And the expectation is that a government AI will eventually start scanning and processing legal content too. All of this is happening before we even need to talk about artificial general intelligence (AGI): the baseline is already shifting right now.
Where do most firms go wrong?
The biggest mistake we come across is not that firms pick the wrong tool. It is that they think their data is safe, and will stay safe, with cloud based AI services. A lawyer who enters client files into ChatGPT, NotebookLM or a comparable tool is trusting that the provider does not use the data for training, does not share it with third parties, and does not fall under a jurisdiction that can compel access to that data.
That trust is fragile. The privacy policies of big tech providers change regularly. The American CLOUD Act potentially gives the government access to data stored with American companies, even when that data sits physically on European servers. And even with paid accounts and a data processing agreement, your control is limited. You are relying on a promise, not on an architecture.
The second common mistake is starting with the technology instead of the problem. A firm buys licences for a legal AI tool, rolls it out, and notices after three months that nobody uses it. Not because the tool is bad, but because the problem it solves was not the problem the team struggled with most.
How we approach AI implementation at law firms
We never start with technology. The first conversation is about one question: where do you run into trouble in your daily work? That sounds simple, but the answer differs per firm and per area of law. A criminal defence lawyer who searches hundreds of pages of court documents has a fundamentally different problem from an employment lawyer reviewing contracts.
From that conversation we translate the bottleneck into a concrete solution. We then offer a trial phase, usually two weeks, in which the firm can experience what the system does with its own case files. No PowerPoint presentation, no demo with made up examples, but a working system fed with your own data. Only once that convinces you do we move on to the full implementation.
The average time from trial phase to working system is six to eight weeks. That is fast, because we are not experimenting: we implement a proven architecture and configure it per firm.
From practice: a local RAG system for criminal case files
A concrete example. A criminal defence lawyer in the south of the Netherlands works with large case files every day: court documents, interview transcripts, reports and correspondence, stored as PDFs and Word documents on a NAS system. Searching those files by hand cost hours per case. Finding relevant passages in large documents was error prone. And connecting the dots between documents was close to impossible.
We built a local RAG system (Retrieval Augmented Generation) that searches those files intelligently with AI. The lawyer can ask questions in natural language, such as "What did witness X say about the incident on 12 January?", and gets an answer with a source reference: document name and page number. The system runs entirely on a Mac Mini in the office. No cloud, no external APIs, no data leaving the building.
The system also handles poor scans, a reality in criminal law, where documents from the police and the public prosecution service are regularly scanned badly. Advanced OCR processing recognises skewed, blurry or partly illegible scans, and says honestly when a passage is not legible enough instead of guessing.
Which AI applications really work for law firms?
Based on our experience, these are the applications that deliver value fastest for firms of 10 to 50 lawyers.
Searching case files in natural language
This is the most direct application. Instead of paging through folders by hand, you ask a question and get an answer with a source reference. The system combines AI search with keyword search (hybrid search) for maximum accuracy. You can filter per case, document type or period.
Automatic case summaries
When you open a new case file, the system generates a summary with a timeline, the people involved and the key points. What normally costs a morning is available in minutes. The lawyer then always reads the original document: the summary is a starting point, not an end product.
Model switching per type of question
Not every question needs the same AI model. A quick factual search can use a compact model that answers in seconds. A complex analysis of connections across a case file deserves a bigger model that reasons more accurately. We configure the system so the user can choose per search which model fits best.
Audio summaries for the road
One feature that is surprisingly popular: the system generates a spoken summary of a case file or part of it. Ideal for getting up to speed on a case on your way to court. Downloadable as an audio file, with a choice of summary depth: key points, extended or complete.
Video analysis and transcription
Case files increasingly contain video files: security footage, recorded interviews, bodycam material. The system transcribes speech automatically and makes video files searchable by timestamp. "What does the suspect say at minute 14:30?" becomes a question you can answer.
What is overhyped?
Tools that claim to give substantive legal advice without human verification. AI is strong at searching, summarising and structuring information. It is not reliable enough to draw legal conclusions on its own. The NOvA is rightly clear about this: the lawyer stays ultimately responsible for advice and legal assistance. AI supports, it does not steer.
What about privacy, GDPR and professional privilege?
This is the question that dominates every conversation, and rightly so. The confidentiality of client data is not a feature you configure later on, it is the architectural foundation of the whole system. We solve this with an approach that differs fundamentally from most AI providers on the market.
First: we run all AI models locally on the client's own hardware. That means no data goes to a cloud, no external APIs are called, and no connection is made to third party servers. Not even when switching between models.
Second: we use European language models as standard. Mistral, developed and trained in France, is our recommended default model. For firms that care about European data sovereignty, and every firm should, that is the logical choice. Open source models from Meta (Llama), Google (Gemma) and Microsoft (Phi) are available as alternatives, all running locally.
Third: we include a complete compliance package as standard. That covers an AI policy template in line with the NOvA recommendations, advice on the DPIA (Data Protection Impact Assessment), a client consent clause for the engagement letter, and full audit logging. Every use of the system is traceable: who searched for what, when, in which case file, with which model, and which sources the system returned. If you face a disciplinary procedure or an audit by the bar association, you can show exactly how AI was used.
What the NOvA recommendations mean in practice for your implementation
The recommendations from the Digitalisation & AI project group of the NOvA are structured around five core values. This is how a well designed AI system meets them in practice.
On confidentiality, the NOvA states that you should not enter confidential data into free tools, that input and output have to stay within the firm's environment, and that you carry out a DPIA when you process personal data. A local system running on your own hardware meets this architecturally: it is not a matter of policy alone, but of it being technically impossible to leak data.
On expertise, the NOvA says you may only use tools that cite their sources, so output can be verified. Every answer our system gives includes a source reference to the specific document and page numbers. The lawyer always checks the original document.
For independence the rule is: AI supports, it does not steer. The system finds relevant passages, and the legal interpretation stays with the lawyer. On integrity, the system delivers full audit logging and a firm wide AI policy.
The cost model: a one off investment versus ongoing licences
Most legal AI tools on the market use a per seat licence model. That means your costs rise linearly with the number of users. For a firm of 30 lawyers paying €50 to €100 per user per month, you are quickly talking about €18,000 to €36,000 a year, every single year.
Our model works fundamentally differently. You invest once in implementing the system, including hardware configuration, the RAG system, user management, audit logging and security. On top of that comes a monthly maintenance contract for updates, security patches, model updates and monitoring. All software is open source, so there are no licence fees. The only ongoing costs are maintenance and the electricity for the server, which comes to around five euros a month.
The difference becomes clearer as your firm grows. With a per seat model you pay more for every new user. With our model you add a user account, and that is it. The investment has already been made.
How do you get the team on board?
Lawyers are risk averse by nature, and rightly so given the consequences of mistakes in their work. That makes adopting new technology a bigger challenge than in most other sectors. Two things help.
First: the trial phase. You do not convince lawyers with a presentation but with experience. Once they upload their own case file and get an answer within seconds that is correct and points to the right page, the discussion is over. The interface is comparable to ChatGPT, and most users are productive within a single session.
Second: the onboarding. We run a two hour training session, on site or remote, plus a user manual with instructions per module and per model profile. The aim is that every user can work independently after that one session.
Something else that helps: start with the people who benefit most. In practice those are often the lawyers handling the heaviest case files. Once they use it and speak positively about it, the rest follow.
Off the shelf versus custom: when do you choose what?
There are now dozens of legal AI tools on the Dutch market: GenIA-L from Sdu, LegalMike, Saga, Legal Mind, Andri, and international players such as Harvey. They are mostly legal search tools that work on public sources: rechtspraak.nl, wetten.nl and professional literature.
What we build is something fundamentally different. Our system searches the firm's own case files, locally, with full control over the data. They are complementary solutions. A firm can use GenIA-L for case law research in public sources, and alongside it a local RAG system for analysing its own client files. The first is a legal search engine. The second is a case file system with AI.
The choice depends on the problem. If you want to search case law faster, an existing tool is often enough. If you want to search your own case files without data leaving the firm, you need a custom solution.
What changes over the next twelve months?
Three developments we expect. First: AI becomes standard in what clients expect. In house counsel already ask how their firm uses AI. A year from now that is no longer a question but a selection criterion.
Second: the government is going to use AI to scan and process legal content. That changes the playing field structurally, and firms that have their own data in order and know their way around AI tools have a head start.
Third: the models are getting better fast. What needs a large model for complex analysis today may run on a compact, faster model in six months. That is why we build systems with model switching: you can add new models as soon as they become available, without replacing the whole system.
The advice is simple: start now, start small, but start deliberately. A two week trial phase with your own case files gives you more insight than six months of meetings about tool selection.
Book a call and we will show you where the opportunities are for your firm.
More reading on AI for legal professionals
- AI compliance for law firms: 5 pitfalls to avoid: the NOvA recommendations and the AI Act in practice.
- AI for due diligence and contract analysis: how AI takes over 75% of repetitive legal reading.
- AI Agency Netherlands: our full range of services for companies that want to implement AI.
- Request a free AI intake: specifically for law firms and legal service providers.
