Language, Memory, and Jurisdiction: Where Legal AI Breaks Down in a South Florida Practice

A Tuesday morning in a mid-sized Coral Gables firm. The intake coordinator is on the phone with a new client in Spanish, taking notes in English. A paralegal is triaging a document production that arrived overnight from opposing counsel in Buenos Aires, mostly in Spanish with some English correspondence threaded through. An associate is drafting a response on a hurricane insurance dispute involving three separately-held properties, two of them owned by a Haitian American family through a limited partnership. The managing partner is reviewing a retention letter for a French Canadian snowbird with assets in Quebec, Florida, and Saint-Martin.

This is a morning. Not a special matter. Just a morning.

Now picture the pitch deck from any major legal AI vendor. The demo firm is in Chicago or Dallas or London. The matter is a domestic contract dispute. The documents are in English. The client is a corporation with a single general counsel. The associates all use the same DMS. Every exhibit has been OCR'd to 99% fidelity.

Those two worlds are not the same world. And the gap between them is where most legal AI implementations quietly fail when they land in South Florida.

Why this matters now

The number of legal AI products pitched to small and mid-size firms has roughly tripled in the last eighteen months. Harvey raised at valuations nobody would have believed in 2023. Spellbook, Legora, Paxton, Evisort, and a dozen others compete for the same shelf space. Thomson Reuters and LexisNexis have rebuilt their flagships around generative AI. Clio now ships an assistant embedded in the practice management layer.

The pressure to pick something is real. Managing partners who ignored this space in 2024 are being asked by junior associates, by clients, and by their own billing committees why the firm is still paying for human-hour document review in a multiform world. Pilots are getting funded. Trials are running.

A predictable pattern is showing up in tri-county firms. The trial starts well. Documents in English get summarized fluently. First-draft memos read respectable. Someone in the firm gets excited, schedules a demo for the partners, and the vendor's sales engineer flies down from New York to close the deal.

Then the tool meets the actual work.

The language asymmetry

The major legal AI platforms are trained and optimized on English-language corpora. The model can handle Spanish. It can handle French. Its handling of Portuguese is uneven. Its handling of Haitian Creole is poor, and in many products Creole is not officially supported at all.

This matters in South Florida in a way that it does not matter in Chicago.

More than two thirds of Miami-Dade residents speak a language other than English at home. Broward County has one of the largest Haitian populations in the United States. A personal injury firm on Pines Boulevard takes new client calls in Spanish, English, and Creole every day. A commercial litigation practice in Brickell reads correspondence from Bogotá, São Paulo, and Buenos Aires. An immigration firm in Coral Gables works with source documents in whatever language the issuing country used.

What the vendors do not tell you, and what shows up in the trial, is that model performance degrades sharply once you move off English. Summarization quality drops. Citation accuracy drops. Tone mismatches multiply. A summary that reads well in English may flatten legal nuance in Spanish and introduce outright hallucinations in Creole. On a translated document production, the compound error rate across a single matter can run high enough that a careful paralegal spends more time verifying AI output than she would have spent reviewing the originals.

You do not see this in the demo. The demo uses clean English documents. The failure mode only appears when your intake coordinator hands the system a phone call transcript that switched between Spanish and English seven times, and the system confidently produces a narrative that is plausible but wrong in both languages.

The memory problem

Legal work is compounding. A single matter can run eighteen months and generate forty thousand pages of discovery. A single client relationship can span a decade and three unrelated litigations. A single firm tracks hundreds of open matters, each with its own fact pattern, its own opposing counsel, its own judge, its own procedural history.

Most general-purpose AI tools reset at the end of a conversation. Some legal-specific tools offer document-level storage. A smaller number offer workspace or matter-level persistence. Very few offer real semantic memory across a matter's lifespan, across an associate's handoff to another associate, across the firm's institutional knowledge about how a particular opposing counsel negotiates or how a particular judge handles motions in limine.

The result is that firms adopting AI at scale end up in one of two failure modes.

In the first mode, the associate treats the AI as a fresh-start drafting tool. Every session begins from zero. The associate re-uploads documents, re-explains the matter, re-establishes context. The time savings are real but modest, maybe twenty percent on a given task, and the context-building tax eats most of that back.

In the second mode, the associate tries to carry context forward and runs into the boundaries of whatever persistence layer the tool offers. The tool forgets that the client has a prior related matter. It forgets that a particular statute was already analyzed in a different filing. It forgets the firm's preferred citation format. It produces content that is internally consistent but inconsistent with work the firm produced six months earlier on the same issue.

Neither mode gets the firm to the productivity step that actually justifies the investment. The real lift from AI in a legal practice comes from compounding context: a system that knows, cumulatively, what this firm has already decided, written, and argued. Without that, legal AI is a faster typewriter.

The jurisdictional confusion

South Florida law offices routinely handle matters where one party, one asset, or one cause of action sits outside US jurisdiction. International estate planning. Cross-border commercial disputes. Immigration work, which by definition involves foreign facts. Maritime and admiralty matters tied to the ports. Hurricane insurance litigation involving non-US insurers on properties held through Caribbean holding companies.

National legal AI products are, with few exceptions, trained and benchmarked on United States law. Some have limited UK or Canadian coverage. Very few handle civil law jurisdictions fluently, and a practice that regularly works with Argentine corporate codes, Venezuelan property law, or Brazilian family law will hit the walls of these tools quickly.

The failure mode here is the most dangerous of the three. In the language case, the tool produces visibly bad output. In the memory case, the tool produces consistent but shallow output. In the jurisdictional case, the tool produces confident, fluent, well-formatted output that is substantively wrong. A partner reviewing a memo on a Panamanian corporate structure may not have the background to catch a US-law-flavored mischaracterization of how bearer shares actually work under current Panamanian law. The AI does not flag its own blind spot.

This is the case that most firms underestimate on the way in and overestimate on the way out, after the third or fourth incident.

What actually works

The firms getting real productivity from AI are not the firms that bought a single product. They are the firms that assembled an architecture.

An architecture has three layers. The first is a domain-specific engine that handles English-language US law work at a high standard: document review, cite-checking, memo drafting. That layer is commoditizing fast, and it is where most of the noise in the legal AI market lives.

The second layer is a multilingual handling capability built around the language pairs the practice actually uses. For a firm that works with Haitian Creole, this is not a feature of the main platform. It is a separate model, or a separate orchestration step, designed around the languages the firm actually touches. The quality of this layer determines whether AI is a net positive in intake, client communications, and document production.

The third layer is persistent memory. Semantic storage of case facts, client context, firm preferences, and institutional knowledge that does not reset between sessions and does not rely on the associate to re-establish context every morning. This is the layer that turns AI from a drafting assistant into an operational system.

None of the major legal AI vendors ship all three layers in a form that works for a twenty-five-attorney practice in Fort Lauderdale. They will claim to. The trial will reveal the gaps.

Assembling the architecture is consulting work. It requires someone who has seen how these tools behave under the actual load of a South Florida practice, not the sanitized load of a vendor demo. It requires deciding which tool runs which layer, how they hand off between layers, and how the firm's own data stays inside the firm rather than leaking into a shared training corpus.

Questions any legal AI vendor should have to answer

A pilot is cheap. The decision to make one of these tools the firm's central research or drafting platform is not. Five questions to ask on the second call with any vendor, before the pilot and not after.

First, on language. What is the documented degradation in accuracy between English and Spanish on your citation-heavy tasks? If the answer is a range or a shrug, the vendor has not measured it. Move on.

Second, on Creole. How does your system handle Haitian Creole source documents? If the answer involves translation as a preprocessing step, ask what the translation quality is on legal Creole specifically and who measured it.

Third, on memory. How does the system treat a matter that spans eighteen months and four associate handoffs? What persists? What resets? Where does the firm's accumulated context actually live on your infrastructure?

Fourth, on jurisdiction. Produce a sample output on a problem involving Argentine commercial law or Panamanian corporate structure. Do not let the vendor pick the example. Pick one from the firm's actual open matter list.

Fifth, on data. Where do our documents go when we upload them? Does your model train on firm data, either during this engagement or in aggregate? What does the contract say, in specific terms, about retention, deletion, and jurisdictional data sovereignty?

The vendor that answers all five cleanly is rare. The vendor that deflects on any of them should be running a deeper pilot, not a larger rollout.

The local angle

South Florida legal practices are not a smaller version of New York legal practices. They are a different shape. The linguistic profile is different. The client base is different. The jurisdictional complexity is different. A national legal AI product can be part of the solution, but it cannot be the whole solution, and pretending it is is how firms end up with a year-long contract they do not use.

The firms that get this right will not be the ones that wait for the perfect all-in-one platform, and they will not be the ones that buy the first thing that impressed the managing partner's golf partner. They will be the ones that treat AI adoption as an architecture decision, tested against the actual shape of their work, with help from advisors who understand both the technology and the local market.

The rest of the industry will catch up. For firms that move before that happens, there is a window.