AI for law firms in Europe is spreading fast, and unevenly. In 2025, 37.9 percent of EU firms in legal, accounting and consultancy work used at least one AI technology, against 20.0 percent of all enterprises.1 The rules on client data have not converged. Italy now requires professionals to tell clients which AI they use. Germany's federal bar sees no general duty to do so.

Summary

  1. Adoption splits north and east. In 2025, 64.9 percent of Swedish legal, accounting and consultancy firms used AI, against 15.2 percent in Romania.
  2. Language AI drives the growth. Text mining in these firms rose from 6.3 to 23.7 percent in two years.
  3. Client rules differ by country. Italy requires disclosure by law, the Dutch bar recommends consent, the German bar sees no general duty. The CCBE calls on-premises AI the most secure option.

How many European law firms use AI?

Eurostat does not count law firms on their own. Its closest group combines legal and accounting firms with head offices, management consultancies, architects and engineers (NACE M69 to M71). The survey covers firms with ten or more staff, so most sole practitioners fall outside it.1

The EU average hides a wide spread. Exhibit 1 ranks member states with a published 2025 figure.

The north uses AI at four times the rate of the east

In Sweden, 64.9 percent of professional firms used AI in 2025. In Romania, 15.2 percent did.

Swedish professional firms use AI at more than four times the Romanian rate.

Legal, accounting and consultancy firms using at least one AI technology, 2025, % of enterprises

EU27 average Member State
0 10 20 30 40 50 60 70 Sweden 64.9% Denmark 60.7% Netherlands 52.8% Austria 50.6% Germany 46.1% EU27¹ 37.9% Spain 34.7% Czechia 32.1% France 29.8% Ireland 27.9% Poland 20.5% Romania 15.2%

¹ All 27 member states. Selected member states shown; Italy, Belgium, Finland and Lithuania have no published 2025 value for this sector. Scope: NACE M69 to M71 (legal and accounting activities, head offices, management consultancy, architecture and engineering, technical testing), enterprises with 10 or more persons employed. Eurostat does not publish legal services separately.

Source: Eurostat, Artificial intelligence by NACE Rev. 2 activity (isoc_eb_ain2), EU survey on ICT usage in enterprises, updated 15 June 2026; accessed 29 September 2026

Nordic and Dutch firms lead. The Netherlands rose from 38.0 to 52.8 percent in one year. France nearly doubled, from 15.3 to 29.8 percent. Germany grew more slowly, from 40.8 to 46.1 percent. Estonia and Luxembourg also pass half.1 For a firm with cross-border clients, the benchmark is the client's home market. The legal and professional services page sets out why the rest hold back, led by data protection concerns.

What do law firms use AI for?

Professional firms read, write and transcribe for a living. The AI technologies they adopted fastest do the same.

Text mining nearly quadrupled in two years

AI that analyses written language rose from 6.3 to 23.7 percent of legal, accounting and consultancy firms between 2023 and 2025.

Language technologies lead AI use in EU legal, accounting and consultancy firms.

Firms using each AI technology, EU27, % of enterprises

2023 2025
0 5 10 15 20 25 6.3 23.7 3.9 15.1 7.3 15.0 15.0 7.1 12.8 6.3 11.3 3.6 4.8 Text mining Language generation¹ Speech recognition Image, video or sound² Workflow automation³ Machine learning Image recognition

¹ Generation of written or spoken language or programming code. ² Generation of pictures, video or sound; first surveyed in 2025. ³ AI-based robotic process automation: automating workflows or assisting decisions. Scope: EU27, NACE M69 to M71, enterprises with 10 or more persons employed. Multiple answers possible.

Source: Eurostat, Artificial intelligence by NACE Rev. 2 activity (isoc_eb_ain2), EU survey on ICT usage in enterprises, updated 15 June 2026; accessed 29 September 2026

Every language technology at least doubled. Generation of text, speech or code rose from 3.9 to 15.1 percent, speech recognition from 7.3 to 15.0 percent. Image recognition barely moved. By purpose, 13.7 percent of these firms use AI for accounting, controlling or finance, three times the all-enterprise rate of 4.6 percent.1

Users also stack tools. In 2025, 25.0 percent of these firms used two or more AI technologies and 15.8 percent used three or more. Across all enterprises the figures were 13.0 and 8.3 percent.1 Each extra tool is one more place where client data can end up.

In a law firm this comes down to four kinds of work:

  • Document review and due diligence: a first pass over contracts and data rooms, confirmed by a lawyer.
  • Knowledge search: the firm's own opinions, templates and precedents, with every source shown.
  • Drafting and translation: first drafts of memos and letters, reviewed before they leave the firm.
  • Transcription: client meetings, hearings and interviews turned into searchable text.

Accountants follow the same pattern. In June 2026 the Dutch professional bodies NBA and NOREA published guidance with dozens of practice examples, from journal entries and contract review to fraud detection and ESG reporting. Validation and human judgement, they write, must remain the basis of public trust.10

Must a law firm tell clients it uses AI?

It depends on the country. European bars start from the same duties of secrecy, competence and independence. They reach different answers. Table 1 compares the main texts.

Table 1. AI rules for lawyers in five European jurisdictions

Jurisdiction Source Date Tell the client? Core rule
EU (CCBE) CCBE guide Oct 2025 If client would object Safeguards before client data
Italy Law 132/2025 Oct 2025 Yes, always AI for support work only
Netherlands NOvA advice Nov 2025 Ask consent first No client data in public AI
France CNB guide Mar 2026 Model fee clause Consent to train on client data
Germany BRAK guide Dec 2024 No general duty Anonymise; prefer EU servers

Statute Professional guidance

Sources: CCBE, 2 October 2025; Legge 132/2025, Art. 13; NOvA, November 2025; CNB, 17 March 2026, as summarised by Village de la Justice; BRAK, December 2024. Accessed 29 September 2026.

Italy is the only one of the five with a statute. Article 13 of Law 132/2025, in force since 10 October 2025, limits AI to support work, and the professional's own intellectual work must prevail. Clients must be told which AI systems are used, in clear, simple and complete language.4 The law covers all intellectual professions, so Italian accountants and consultants are bound too.

The Dutch bar recommends consent, which may go further than its conduct rules strictly require.5,6 The German federal bar found no general duty to inform in December 2024. It stressed that secrecy is protected by criminal law, and that a provider's mere ability to read client data counts as access.7 France's CNB proposes a model clause for fee agreements. It requires consent before client data trains a firm's own model, and disclosure when a chatbot answers clients.8,9

A firm with offices in several countries is safest applying the strictest rule firm-wide.

Is AI used by law firms high-risk under the EU AI Act?

Mostly not. Several vendor pages claim legal AI became high-risk in August 2026. The text says otherwise. Research, drafting, document review and knowledge search do not appear in Annex III.11

Two uses do. AI used by or for a judicial authority, or in a similar way in alternative dispute resolution, is high-risk under Annex III, point 8(a). So is AI for recruiting, promoting or evaluating staff under point 4, which reaches a firm's own HR. After the Digital Omnibus, these duties apply from 2 December 2027.12

Some duties already apply. Since 2 February 2025, firms must support the AI literacy of staff who use AI. Since 2 August 2026, a chatbot must tell people they are dealing with an AI. A firm that builds its own client chatbot carries that duty as provider. The AI Act timeline lists every date.

What happens when a lawyer relies on unchecked AI output?

Courts now act on it. In June 2025 the High Court of England and Wales heard two cases together. In one, a barrister and solicitors cited five fake cases. In the other, 18 of 45 citations put before the court did not exist.13 The court warned the whole profession and referred lawyers to their regulators.

The CCBE lists professional misconduct, contempt of court and malpractice claims as risks of unverified output.2 The BRAK calls final checking by the lawyer mandatory.7 The Dutch bar asks lawyers to check citations, case law and facts by hand, and to prefer tools that show their sources.5 A tool that links every answer to a document in the firm's own files makes that check fast.

Where can AI for law firms run safely?

The CCBE's technical guide of March 2026 sets out four ways to run a model.3 Control and cost both fall from top to bottom.

Table 2. Four ways a law firm can run AI, per the CCBE

Option Runs the model Audit rights Cost to the firm
On-premises Firm Full Highest
Own servers, colocation Firm Full High
Own model, rented cloud Firm Certificates only Lower
Vendor SaaS or API Vendor Certificates only Lowest

Firm runs the model Vendor runs the model

Source: CCBE, Technical guide on the use of AI tools and models by lawyers, 27 March 2026, section on deployment options; accessed 29 September 2026. Cost ranking as described by the CCBE, not measured.

The CCBE calls on-premises deployment the most secure option for confidentiality. Data never leaves the firm, and in some countries a lawyer's office has extra protection against searches. That protection may not reach servers in a colocation data centre.3

The CCBE sees renting as a middle route. A firm runs its own model on rented cloud servers, picks the data centre region and can keep the encryption keys. It also avoids lock-in to one AI vendor and costs less than owning the hardware. The catch is audit: most providers only show certificates.3 An EU-owned provider also keeps the provider itself under European jurisdiction.

Data held by a provider can be demanded from the provider. Since 18 August 2026, the EU e-Evidence Regulation lets authorities in one member state send production orders directly to service providers in another.14 Firms may want a provider to tell them at once about such measures. Smaller providers may agree; large ones often will not, the CCBE notes.3

Own hardware costs less than many firms assume. At September 2025 prices, the CCBE puts a dedicated machine for 20 to 40 billion parameter models at about €2,000. A 96 GB RTX PRO 6000 at about €8,000 runs gpt-oss-120b. Around €20,000 lets several users share that model. The largest open-weight models need 1 to 2 TB of memory, and more than €150,000 for GPU-only inference. Memory prices doubled during 2025, so these figures date quickly.3 The hardware guide sizes models to GPUs, and the European GPU price index covers renting. Lindstead's AI infrastructure cost analysis compares both for a given firm.

What should a law firm do first?

Lindstead recommends this order:

  1. List every AI system in use, including AI features inside translation, PDF and office tools.
  2. Sort the firm's work into confidentiality tiers, from public research to sensitive litigation and deals.
  3. Check each vendor against the CCBE safeguards: confidentiality or zero retention, a data processing agreement, EU or EEA processing.
  4. Decide what clients are told, using the strictest rule of any country the firm works in.
  5. Move the most sensitive tier to a model the firm runs itself.

Lindstead maps these duties per system through AI governance and compliance work. Model selection tests open-weight models on the firm's own documents and languages. AI deployment builds a first private use case on infrastructure the firm controls. The break-even analysis shows when self-hosting beats an API.

Frequently asked questions

  • Not without safeguards. The CCBE advises lawyers to keep client data out of generative AI unless the provider is bound to confidentiality or zero retention, a data processing agreement limits use, or the model runs in an environment the firm controls. The Dutch bar advises against client data in public AI models. The German federal bar says removing names is often not enough, because the context can identify the matter.

  • It depends on the country. In Italy, Law 132/2025 has required professionals to tell clients which AI systems they use since 10 October 2025. The Dutch bar recommends asking for consent, while the German federal bar sees no general duty to inform. The CCBE expects transparency wherever an informed client would object or set conditions.

  • Legal research, drafting, document review and knowledge search are not high-risk under Annex III. AI that assists a judicial authority or alternative dispute resolution is, and so is AI for recruiting or evaluating staff. After the Digital Omnibus, those high-risk duties apply from 2 December 2027.

  • The CCBE's technical guide of March 2026 puts a dedicated machine for 20 to 40 billion parameter models at about €2,000, and a 96 GB GPU able to run gpt-oss-120b at about €8,000. Around €20,000 lets several users share that model. The largest open-weight models need more than €150,000 for GPU-only inference.

  • Eurostat does not count law firms separately. In 2025, 37.9 percent of EU firms in legal, accounting, consultancy and related activities with ten or more staff used AI, against 20.0 percent of all enterprises. Use ranged from 64.9 percent in Sweden to 15.2 percent in Romania.

Sources

  1. Eurostat, Artificial intelligence by NACE Rev. 2 activity (isoc_eb_ain2), updated 15 June 2026. ec.europa.eu/eurostat/databrowser/view/isoc_eb_ain2/default/table?lang=en
  2. CCBE, Guide on the use of generative AI by lawyers, 2 October 2025, sections 3 and 4. www.ccbe.eu/fileadmin/speciality_distribution/public/documents/IT_LAW/ITL_Guides_recommendations/EN_ITL_20251002_CCBE-guide-on-the-use-of-the-use-of-generative-AI-for-lawyers.pdf
  3. CCBE, Technical guide on the use of AI tools and models by lawyers, edition 2026, 27 March 2026. www.ccbe.eu/fileadmin/speciality_distribution/public/documents/IT_LAW/ITL_Guides_recommendations/EN_ITL_20260327_CCBE-technical-guide-on-the-use-of-AI-tools-and-models-by-lawyers.pdf
  4. Legge 23 settembre 2025, n. 132, Article 13, Gazzetta Ufficiale n. 223 of 25 September 2025, in force 10 October 2025. www.gazzettaufficiale.it/eli/id/2025/09/25/25G00143/sg
  5. Nederlandse orde van advocaten, Aanbevelingen AI in de advocatuur, November 2025. www.advocatenorde.nl/voor-advocaten/digitalisering-ai/aanbevelingen-ai-de-advocatuur
  6. Mr. Online, NOvA publiceert langverwachte AI-aanbevelingen met opvallend advies, 9 December 2025. www.mr-online.nl/advocatenorde-publiceert-langverwachte-ai-richtlijnen-met-opvallend-advies/
  7. Bundesrechtsanwaltskammer, Hinweise zum Einsatz von künstlicher Intelligenz, December 2024, sections 3 and 4. www.brak.de/fileadmin/service/publikationen/Handlungshinweise/BRAK_Leitfaden_mit_Hinweisen_zum_KI-Einsatz_Stand_12_2024.pdf
  8. Conseil national des barreaux, Le CNB adopte un guide sur la déontologie et l'intelligence artificielle, 17 March 2026. cnb.avocat.fr/actualite/le-cnb-adopte-un-guide-sur-la-deontologie-et-l-intelligence-artificielle
  9. Village de la Justice, IA et déontologie des avocats: notre lecture du guide du CNB, 23 April 2026. www.village-justice.com/articles/deontologie-des-avocats-guide-cnb,56753.html
  10. NBA and NOREA, Leidraad 2: AI toegepast, June 2026. www.nba.nl/tools-en-ondersteuning/publicaties/2026/leidraad-2-ai-toegepast/
  11. Regulation (EU) 2024/1689 (Artificial Intelligence Act), Articles 4 and 50 and Annex III. eur-lex.europa.eu/eli/reg/2024/1689/oj
  12. Regulation (EU) 2026/1744 (Digital Omnibus on AI), Official Journal, 24 July 2026. eur-lex.europa.eu/legal-content/EN/TXT/?uri=OJ:L_202601744
  13. High Court of England and Wales, Ayinde v London Borough of Haringey and Al-Haroun v Qatar National Bank, [2025] EWHC 1383 (Admin), 6 June 2025. www.judiciary.uk/judgments/ayinde-v-london-borough-of-haringey-and-al-haroun-v-qatar-national-bank/
  14. Regulation (EU) 2023/1543 on European Production Orders and European Preservation Orders for electronic evidence, applicable from 18 August 2026. eur-lex.europa.eu/eli/reg/2023/1543/oj

Method: Eurostat values read from the isoc_eb_ain2 API and rounded to one decimal. Legal texts and professional guidance read in the original language. All sources accessed 29 September 2026. This page is general information, not legal advice. Corrections: contact@lindstead.com.