Legal AI: My Takeaways From a Paris Law Firm Engagement

In short: Legal AI dramatically speeds up contract analysis, case law research, document drafting, and regulatory monitoring, but final responsibility always stays with the practitioner. Two non-negotiable rules apply: every AI-generated ruling or legislative reference must be verified against primary sources like Légifrance, because hallucinated case law is well documented, and client data covered by professional privilege must only run on enterprise solutions that contractually guarantee no use for training, or on-premise deployments.

In 2025 I supported a Paris law firm (45 staff, specialised in business law) through their transition to AI tools. The engagement ran for six months and put me up against some particularly delicate ethical trade-offs. Here is what I took away from it — a subject where caution pays off far more than enthusiasm.

The legal sector is one of the most transformed by AI, and one of the most cautious in adopting it — for good reasons. AI can dramatically speed up research and document analysis, but the final responsibility always stays with the practitioner. Here are the categories of tools that actually matter.

Contract analysis and review

This is the use case where AI delivers the most immediate time savings:

• Luminance, Kira: analysis of document corpora, clause extraction, M&A due diligence review
• Ironclad, Docusign CLM: contract lifecycle management with AI
• Claude: excellent for analysing long contracts thanks to its extended context window — usable in enterprise with the right guarantees

Legal research

• Lexis+ AI, Westlaw Precision (English-speaking markets): AI-powered case law
• Doctrine.fr: French case law database with AI features
• Perplexity AI: fast research on general points of law, to be checked systematically against primary sources

AI research does not replace verification against the official texts (Légifrance, the Official Journal of the EU). It speeds up the initial clearing of the ground.

Drafting documents and pleadings

• ChatGPT Enterprise: first drafts of pleadings, legal memos
• Claude for Business: high-quality legal structure and reasoning
• Harvey.ai: an LLM specialised for law firms, trained on legal content

Critical point: never publish a document or pleading produced by AI without a complete review by a qualified practitioner. Hallucinated rulings or references to legislation that does not exist are well documented.

Compliance and regulatory monitoring

• ComplyAdvantage, LexisNexis Risk: AML/KYC compliance with AI
• Perplexity: sourced daily regulatory monitoring
• ChatGPT: summaries of new regulations, impact analysis

The risks specific to the legal sector

• Professional privilege: documents sent to a cloud LLM can step outside the perimeter of privilege. Favour enterprise solutions with contractual guarantees, or on-premise deployments.
• Legal hallucination: LLMs can invent rulings, articles of law, references that simply do not exist. Systematic verification is mandatory.
• Client confidentiality: a firm's client data is covered by professional privilege. Processing it through a cloud AI has to be covered by an appropriate agreement.

On all of these points, our GDPR and AI compliance guide sets out the questions to ask every vendor.

My takeaways from the Claude Enterprise deployment at the firm

After the audit, we chose Claude Enterprise over ChatGPT Enterprise for the main deployment. The reason: the 200,000-token context window made it possible to analyse a complex contract in one pass, without splitting it up. For an M&A lawyer reviewing a 200-page share purchase agreement, that is a major qualitative win — the consistency of the analysis is preserved.

The ROI showed up from the first quarter: on a typical M&A due diligence, document analysis time dropped from 80 hours to roughly 30 hours, with better coverage quality (the AI spots unusual clauses that juniors can miss out of fatigue).

The legal hallucination trap — a case I witnessed

At the start of the engagement, I watched a junior produce a draft legal memo that cited two Court of Cassation rulings entirely invented by ChatGPT. The references looked plausible (credible date, chamber, appeal number), but the rulings did not exist. The reputational and ethical risk of publishing an error like that would have been enormous.

The absolute rule that came out of it at the firm: any case law or legislative reference generated by AI must be verified on Légifrance or Doctrine.fr before publication. This discipline is not optional, it is a condition of using AI in law.

Harvey.ai vs generic Claude: my call

The firm tested Harvey.ai alongside Claude Enterprise. The verdict after four months: Harvey is more precise on technical Anglo-American legal terminology but remains expensive for French-speaking firms. Claude is generalist, but its overall quality in legal French is superior for the French market.

For an international firm with heavy Anglo-American law activity, Harvey can be justified. For a French-speaking firm, Claude Enterprise remains the best value-for-quality compromise.

Case law research: where Doctrine.fr wins

For pure legal research, Doctrine.fr with its AI features remains the most reliable tool on the French market. The case law database is exhaustive, expert commentary clarifies the scope of decisions, and the semantic search AI surfaces relevant rulings that a keyword search would have missed.

Perplexity and ChatGPT can serve to quickly clear the ground on a general question, never for substantive case law research. The boundary has to be crystal clear for the juniors who use these tools.

Client confidentiality: the red line

A lawyer's professional privilege is uncompromising. We put three inviolable rules in place:

1. No identifiable client document goes onto a open access version of an LLM
2. The Enterprise versions used are strictly those that contractually guarantee data will not be used for training
3. The most sensitive matters go through an on-premise solution (a local Mistral deployment was assessed for 2026)

This discipline was validated by the Bar during an informal audit. Without it, the use of AI would have been legally questionable.

What I learned about change management

The hardest part was not technical but cultural. Many experienced partners initially saw AI as a threat to quality, or to the profitability of the billable-hour model. Six months later, those who adopted it saw their capacity to handle complex matters increase, and their added value shift towards strategic advice. Those who refused now find themselves competing against firms that became faster at the same access conditions.

Our reading for Trust-Vault

Legal tools are among the most heavily scrutinised in our evaluation. The Trust Score gives high weighting to the Security pillar and the Privacy pillar for these tools. Our methodology details the criteria we apply.

Further reading

For a complementary implementation angle, read Prompt Engineering: The Techniques I Actually Use Daily.

For a complementary implementation angle, read AI Tools for SMEs: The Stack I Actually Deploy in 2026.