Author
Head of Growth
Stephane leads growth at HAQQ and writes the AI & Legal Tech section: model benchmarks, tool comparisons and the engineering behind legal AI that holds up in practice.
Long documents, several jurisdictions, multi-step tasks, exact citations and the judgement to decline. What a legal AI has to do on a real matter, why single-prompt tests miss it, and how HAQQ thinks about evaluating it.
Fake AI citations went from curiosity to a docket of thousands. AI client confidentiality is on the same curve. The triggers, our forecast and four controls.
Plain Claude is enough for a lot of legal work. We counted which law Claude's public legal plugins are written around, and what you would build yourself for Gulf law.
We asked four frontier models for court judgments backing an article that doesn't exist. Two declined. Two named 16 between them. Here are the five tests.
Colombia fined a lawyer for AI-invented case law. El Salvador imposed no sanction on a citizen. A Mexican court calls AI "recommendable". We read the rulings.
We put Opus 5.5, Opus 5 and Fable 5.1 through ten lawyer-grade prompts and 120 blind verdicts. They finished four verdicts apart, in a loop. Only price separated them.
We ran TypeSafe's Jev across our legal AI stack for under $0.20. It reranks statutes well. It also scored fabricated citations higher than honest ones.
We tested the legal AI market instead of reading about it. 65% of vendors are a prompt layer over somebody else's model, and 3% train on legal data.
Microsoft's AI code of conduct puts every professional duty on the Operator. In 38 pages, "lawyer" appears zero times. Your firm is the Operator.
We ran GPT-6 Astra against 41 real legal questions, alongside Claude Opus 5 and GPT-5.6 Luna Pro. Every model we tested got Saudi or Egyptian law wrong while citing the correct article, the failure no citation checker catches.
84% of chief legal officers now report to their CEO. The seat moved to where decisions get made, and what legal hands over mostly did not move with it.
Most legal AI sends your prompt straight to a general model. The real work happens before that. Here is the architecture, and four questions that test any vendor for it.
Regulation (EU) 2026/1744 entered into force on 27 July 2026. High-risk duties moved to December 2027 and August 2028, and the 2 August 2026 transparency deadline did not move. Every change, and what it means for a firm that uses AI.
An independent evaluator ran HAQQ through 61 hard legal tasks in 21 jurisdictions, twice. It passed 27.0% overall, and 5.9% on the hardest subset against 1.1% for the average legal AI. Full numbers inside.
Reliable AI legal research retrieves real statutes and cases, then cites what it found. HAQQ scores 43/50 for research, top-3 and first for Arabic and civil-law work; LexisNexis leads US common-law.
How legal AI handles an immigration file: source-language intake, fact checks, USCIS and Home Office deadlines, UAE Golden Visa evidence, and which tools cover what.
General chatbots hallucinate law and keep no secrets. Here's how a legal AI should be built for accuracy and privacy, and the checklist to judge one before you rely on it.
AI won't replace your lawyer. It removes the empty space before one. Why legal AI is the first step, why the human keeps the liability, and why it frees good lawyers instead of ending them.
Antoine Kanaan on why law firms need a legal operating system, not another AI tool, and why context, not the model, is your firm's real competitive edge.
We made a frontier model impersonate a legal database. It refused to invent fake cases, then cited a real law for the wrong thing, the hallucination that survives a lawyer's review.
How divorce and family-law attorneys actually use AI across the matter, from intake to court prep, what stays human, and how to pick a tool that does not create a privilege or malpractice problem.
A no-hype map of San Francisco's legal-AI startups in 2026: who they are, what they do, who they're for, and how much they raised, from Harvey to Ivo.
Onit and HAQQ surface in the same searches but solve different problems. An honest comparison of Enterprise Legal Management vs an AI-native legal assistant, and which one you actually need.
AI is replacing paralegal tasks, not paralegals. The data on what AI does well, what it fails at, and what paralegals should learn to stay valuable.
Lexis+ AI is now Lexis+ with Protégé. Our review: 46/50 on legal research vs 40 for CoCounsel, the Stanford hallucination figures, and what each vendor shows on price.
Harvey 37.9, CoCounsel 36.6, Legora 35.9 on one published 50-point rubric, plus price, Word add-in, research data, Arabic and deployment. Updated October 2026.
An independent CoCounsel review backed by published benchmark scores: pricing, the Westlaw lock-in reality, and the best CoCounsel alternatives for 2026.
How many lawyers use AI in 2026? About 69% use generative AI, per 8am. 38 sourced stats on adoption, hallucination cases, funding and benchmark scores.
We scored 19 legal AI tools on a published 50-point benchmark: HAQQ 46.8, Claude Fable 5 42.7, Harvey 37.9. Ranking, prices and Arabic support, updated October 2026.
Harvey AI review, updated October 2026: $15.5B valuation, no published price, 37.9/50 on our benchmark. What it does well, who can buy it, and alternatives.
October 2026: Claude Opus 5.5 and GPT-6 Astra lead our latest legal tests. Plus 3,000 graded answers from 10 models, 24% of which cite law that does not back them.
We mapped about 110 AI lawyer apps and re-read store ratings on 20 Oct 2026: Alex AI, Vikk, Rocket Lawyer, plus Arabic apps Adel, Shwra, Eyas and Baeynh.
Six popular LLM guardrails were bypassed at 12–100% rates. Governance by construction builds the unsafe action out of the agent, nothing left to evade.
Yes, in most major markets, with duties attached. We tracked 17 jurisdictions: 10 permit with guidance, 1 restricts (Qatar), 6 say nothing at all.
2,202 court decisions now involve AI-fabricated citations, per the Charlotin database (Oct 2026). Sanctions by country, the growth curve and how to avoid them.
Mike OS, CourtListener, Docassemble, LegalBench and more: licences, GitHub stars and uses as of October 2026, plus what open source legal AI still cannot do.
Spellbook review, October 2026: no published price, 46/50 on contract drafting but 32.9/50 overall on our benchmark, the gaps buyers hit, and alternatives by shape.
Legora review, updated October 2026: $5.6B valuation, 1,000+ customers, no published price, 35.9/50 on our benchmark, and how it compares feature by feature with HAQQ.
Five adversarial NDAs, five prompt-injection payloads, one zero-dependency scanner, all blocked in under 2 ms. How injection hits legal AI and what stops it.
RAG chunking destroys legal document structure. How knowledge graphs, span-level search and extractive entity linking power portfolio-scale review.
Which model should write your client alert, law review piece or firm blog? Drafting scores, cost per answer and citation behaviour from our September 2026 legal runs.
Anthropic's legal webinar drew 20,000 registrants and 51 upvoted questions. What lawyers asked about privilege, hallucinations and Claude, answered.
Intake, scheduling and first drafts are automating fast. Where legal AI workflows break, the compliance wall, the last mile, and what works instead.
Most legal tech implementations fail, and it's rarely the software. The configuration trap, six human pitfalls, and the framework winning firms use.
What human in the loop means for lawyers, what ABA Formal Opinion 512, EU AI Act Article 14 and the SRA Code require, and how to design AI review that holds up.
Same model, same 30-doc data room. A single prompt caught 3/5 planted issues; a 3-agent swarm caught 5/5. The misses are structural, here is why.
An AI-drafted motion to dismiss costs $1.67, or $4.55 if you run everything on the biggest model. Where the money goes, stage by stage, and the fix.
We built a working A* litigation planner in 517 lines, then refused to use it. The four gates any AI planner must pass before touching a real matter.
Three US attorneys asked how HAQQ differs from ChatGPT. The answer: a 20-page NDA cross-analysis in 90 seconds and an engine that says 'I don't know'.
Claude for Word is genuinely good, but data deleted in 30 days, no audit logs, and a Team/Enterprise paywall matter for law firms. The honest breakdown.
A Mexican labour-law ontology replaced 300 MCP tools with 7 and cut cost per message from $0.60 to $0.02. The method, the Stanford hallucination data and a build plan.
We ran 3 parallel simulations with 72 AI agents and 1,543 interactions to score legal AI's future: Harvey IPO odds, AI malpractice settlements, BigLaw cuts.
Claude's legal plugin didn't replace the legal stack, it exposed the weak layer. What it actually replaces, what it can't touch, and where it fits.
A federal judge ruled 31 AI-generated documents are not privileged. Why AI chats are discoverable, and what lawyers must change in engagement letters now.