The verdict in three sentences
A private AI document search tool (RAG architecture, retrieval-augmented generation) lets a law firm find a clause, an easement or a precedent wording across 40,000 documents in about ten seconds, with a reference to the source document. Setup costs between 15,000 and 45,000 EUR excl. VAT (2026 market estimate), plus 50 to 400 EUR per month in compute. The non-negotiable condition: a model hosted in the European Union, with no reuse of data for training, compatible with legal professional privilege and the GDPR.
Why a law firm loses time searching
In a 25-person firm in Dublin, solicitors, trainees and paralegals search for precedents every day: a non-compete clause in a share purchase agreement, a right-of-way wording in a conveyancing file, a limitation of liability clause in a commercial contract. Documents sit in the practice management system, in scanned PDFs or in shared folders.
Classic full-text search often fails: words do not match, scans are not OCR'd, templates vary over the years. A paralegal easily spends 20 to 40 minutes a day searching, i.e. 80 to 160 hours a year per person.
A RAG engine understands the question in natural language, retrieves the relevant passages and writes a short answer with links to the cited documents. The lawyer always checks the source: the tool speeds up research, it does not replace legal analysis.
Options and costs in 2026
| Option | Setup | Monthly cost | Data hosting | Limits |
|---|---|---|---|---|
| Consumer AI assistant | 0 EUR | 20 to 30 EUR/user | Outside the firm's control | Incompatible with privilege |
| Legal-tech vendor solution | 3,000 to 10,000 EUR excl. VAT | 40 to 120 EUR/user | Ireland or EU | Corpus limited to the vendor's software |
| Custom RAG, EU-hosted API model | 15,000 to 30,000 EUR excl. VAT | 50 to 250 EUR compute + hosting | EU, no-training contract | Depends on the provider's terms |
| Custom RAG, self-hosted open-source model | 25,000 to 45,000 EUR excl. VAT | 200 to 400 EUR (shared GPU server) or more | Dedicated server in Ireland | Heavier technical maintenance |
| Maintenance and upgrades | Included in contract | 300 to 900 EUR excl. VAT/month | Depends on option | Corpus and model updates |
These figures are 2026 market estimates in Ireland and France. They assume 40,000 documents, part of which need OCR, and 25 users.
What makes up a RAG project budget
| Work package | Budget share | Content | Duration |
|---|---|---|---|
| Audit and scoping | 10 % | Source inventory, access rights, GDPR impact assessment | 1 to 2 weeks |
| Corpus migration and OCR | 15 to 25 % | OCR, cleanup, splitting documents into passages | 2 to 3 weeks |
| Vector indexing | 15 % | Vector database, metadata (date, document type, partner) | 1 week |
| Answer engine | 20 to 25 % | Model choice, mandatory citations, guardrails | 2 weeks |
| Interface and permissions | 15 % | Internal web search, authentication, matter-level isolation | 2 weeks |
| Testing and training | 10 to 15 % | 200 test questions, accuracy measurement, staff training | 1 to 2 weeks |
The critical point is permission isolation: a trainee must not access sensitive matters, and the AI's answer must respect the same rights as file access, including ethical walls. A serious quality benchmark targets at least 85 % correctly sourced answers on the test set before go-live.
On compliance, the project must appear in the record of processing activities, undergo a data protection impact assessment (DPIA) if sensitive data is processed, and rely on a processor agreement that forbids training the model on your data. The EU AI Act classifies this kind of internal support tool as limited risk, with a transparency obligation towards users.
Mini case study
*Illustrative example.* Siobhan, managing partner of a 25-person firm in Dublin, estimates that her 14 solicitors and paralegals lose 30 minutes a day looking for precedents.
Need a professional website?
Kolonell builds websites that attract clients, optimized for the Sénégalese market. Free quote in 2 minutes.
- Time lost: 14 x 0.5 h x 210 days = 1,470 hours per year.
- Value at a 45 EUR loaded hourly cost: 66,150 EUR per year.
- Cautious assumption: the AI halves this time, i.e. 33,075 EUR saved.
- Costs: custom RAG with an EU-hosted model estimated at 28,000 EUR excl. VAT, plus 150 EUR compute and 500 EUR maintenance per month, i.e. 35,800 EUR in year one.
Break-even comes around month 13, after which net savings exceed 25,000 EUR per year. The qualitative gain matters too: junior staff access the firm's best drafting from day one.
FAQ
Is legal professional privilege compatible with AI?
Yes, if documents stay in a controlled, EU-hosted infrastructure with a contract excluding any training on your data. Consumer assistants at 20 EUR a month do not meet these conditions.
Can the AI invent a clause?
A well-configured RAG only shows answers citing real passages, and declines to answer when no source is found. On 200 test questions, the target is fewer than 5 % unsourced answers.
Do all old documents need OCR?
Not necessarily: firms often start with the last 10 years, which covers 60 to 70 % of searches. OCR generally costs 0.01 to 0.05 EUR per page in 2026.
How long does the project take?
Count 8 to 12 weeks, including 2 to 3 weeks for corpus migration. A pilot with 5 users can run from week 6.
Can compute costs spiral?
With 25 users asking 20 questions a day each, compute generally stays between 50 and 250 EUR per month. A monthly cap and usage monitoring prevent surprises.
Let's scope your project. Describe your corpus, software and hosting constraints: we will scope the architecture and the timeline of a pilot. Detailed quote within 48 h via our form. WhatsApp +221 77 596 93 33.
Mohamed Bah
Fondateur, Kolonell
Passionate about digital and entrepreneurship in Africa, Mohamed has been helping Sénégalese businesses with their digital transformation since 2020. Founder of Kolonell, he believes every SME deserves a professional and accessible online présence.
