Somewhere between the bar election that requires lawyers to drive 1,000 km to vote and the contract review eating another weekend, the legal AI industry decided the answer was one more English-first assistant. We took a close look at HAQQ, the Beirut-built platform behind the Justinian engine that treats Arabic as a drafting language and civil-law systems as the baseline rather than the localization step. Inside: what 15,000 firms across 80+ countries actually bought, why its published benchmarks and dated per-country corpus are the hardest thing to copy in legal AI, and the one number Harvey, Legora, and Clio all keep getting wrong.
Start with the evening, not the demo. A managing partner in Beirut or Riyadh has a commercial agreement to review by morning, and the firm's matters, research, time tracking, and billing sit in spreadsheets, plugins, and disconnected tools where nothing connects and deadlines slip. Generic AI makes it worse in a very specific way. In customers' words, ChatGPT "sounds plausible but misses the risk my client will actually face," because it doesn't know the jurisdiction, the precedent, or the firm's voice. HAQQ's own retrieval testing found that Arabic queries surfaced 9x more English primary law than Arabic law, alongside silent wrong-country errors: the model answers confidently about the wrong legal system, and nobody notices until the client does. So HAQQ flipped the baseline. Most legal AI learns US and UK law first, then gets localized. HAQQ takes civil-law and Sharia-influenced systems as the starting point and treats Arabic as a drafting language rather than a translation layer.
The wedge: jurisdiction depth you can check, plus the whole stack
Two moves, and they reinforce each other. The Justinian engine scored 49 out of 50, the highest of 19 models tested on HAQQ's published benchmark of 250 complex legal tasks, and 1,372 long-horizon tasks have been run end-to-end and graded all-pass. More interesting than the score is the disclosure around it. Thirty-three countries have a dedicated, dated legal corpus, last reviewed 2026-08-26, and the coverage page states depth per tier, so a buyer in Doha knows before purchasing whether Qatar is handled deeply or generally. Rivals publish either nothing about geography or a list of countries they refuse to serve. HAQQ publishes the opposite. Then the bundle: Legal AI Chat for drafting and research, eFirm for matters, billing, and time tracking, eBar for bar associations, and a Legal OS deployment for enterprises. The wedge is hard to copy for unglamorous reasons. Arabic-first retrieval is corpus work, not prompt work, and it takes years. The bar-association channel is institutional trust, not ad spend. Even the ROI calculator admits its defaults are assumptions, which is a trust position a bigger competitor would find awkward to match overnight.
The ICP they actually win: MENA firms from solo to mid-size, and the associations that govern them
The headline number is 15,000+ firms and in-house teams across 80+ countries. The traction is concentrated in Lebanon and the GCC, and the buyer splits cleanly. Managing partners buy eFirm, which starts at $60 per user per month with a two-seat minimum. Individual lawyers buy the Chat tier, Starter at $30 a month for 100 credits, with top-up packs from $1 to $100 and a best rate around $0.227 per credit. Students get 60 credits for $15 with a university email, which is how the next generation of the ICP gets seeded early. The named logos are honest about the segment: Keserwani & Associates ("more organized and efficient"), Kawtharani Law, Yakan Law Firm, and BOXXER, whose CEO Ben Shalom credits it with handling commercial agreements faster. The institutional layer is the interesting one. The President of the Beirut Bar Association has endorsed HAQQ by name as "the one company that truly understands what AI is, how legal work should flow," and the Tripoli Bar Association runs on eBar because lawyers were traveling 1,000+ km to vote or collect a certificate. From there, enterprise deals reach banking, telecom, and government buyers who require the SOC 2, ISO 27001, ISO 42001, and GDPR posture, zero training on customer data, and the 99.9% uptime commitment. A 5.0 rating across 77 Google reviews is the kind of proof anyone can verify in ten seconds.
What the category still gets wrong: it counts countries, not depth
HAQQ names its competition directly: Harvey, Legora, Spellbook, LexisNexis, Clio, and generic ChatGPT or Claude. The contrarian read is that nearly all of them optimise for the same wrong number. The AI tools benchmark in English, then market the score as global coverage, when an English benchmark tells an Arabic-speaking practitioner almost nothing about whether the model retrieves Qatari commercial law or Texas case law. The practice-management incumbents count seats and billable-hour capture, when the leak a ten-lawyer firm actually feels is the disconnected stack where a deadline slips between the research tool and the spreadsheet. Country counts are the vanity metric. Corpus depth per jurisdiction, in the language the practitioner drafts in, with a review date attached, is the number that predicts whether the output is usable. "Built for MENA jurisdictions, not translated into them" is a claim about retrieval infrastructure, and the published inclusion list with per-tier depth is the receipts. The $25K-per-lawyer-per-year savings claim is aggressive. At least the calculator shows its assumptions.
The takeaway
Defensibility in vertical AI is moving from model access, which everyone now has, toward corpus depth, workflow lock, and institutional channel, which almost nobody has. HAQQ holds all three, backed by a $3M round covered by Forbes Middle East and Wamda, in a region the big legal-AI players treat as an afterthought. If you sell into a non-English jurisdiction, ask any vendor for their corpus review date. Then watch the silence.