"Legal AI" has become one phrase for three quite different things, and most of the confusion about it comes from treating them as one. A managing partner asking whether the firm should use AI is usually asking about all three at once, and they deserve different answers.
This guide separates them, explains what each is genuinely good at today, where each goes wrong, and what to ask before you let any of them near a client file. It is written for firms in the UAE and Saudi Arabia, where the paperwork arrives in two languages and the consequences of a wrong date are measured in appeal windows.
The three kinds of legal AI
1. Research tools
These answer legal questions by drawing on statutes, regulations and case law: what the limitation period is, how a court has read a clause, which authorities support a position. Their value depends almost entirely on the body of law behind them and on whether that body covers your jurisdiction. A tool trained mostly on common-law judgments has little to say about a Dubai Court of Cassation ruling or a Saudi circular.
2. Drafting and review tools
These work on documents you are producing: first drafts of contracts and letters, clause comparison, redlines against a standard position. They save the most time on high-volume, well-structured work and the least on anything genuinely bespoke.
3. Case-file tools
These work on the documents you already hold for a matter: the case bundle, the correspondence, the filings and the judgments. They do not tell you what the law is. They tell you what your file says: what happened and when, who is involved, and where two documents disagree. For a litigation practice this is often where the hours actually go, and it is the kind least discussed.
Where legal AI genuinely helps today
- Reading volume. A new matter that arrives as four hundred pages can be indexed and made searchable in minutes rather than read cover to cover before anyone can answer a simple question.
- Chronologies. Pulling every date out of a bundle and putting the events in order is exactly the patient, mechanical work machines do well, and exactly the work a junior lawyer spends a week on.
- Finding contradictions. Two documents giving different dates for the same event is usually the most important thing in a file. A careful tool surfaces it; a careless summary hides it.
- Mapping the parties. The same company written five different ways across a bundle is one party, and a tool can say so.
- Working across Arabic and English. In the Gulf a single matter routinely mixes both. Being able to ask in one language and search documents written in the other is a practical advantage, not a novelty.
Where it goes wrong
It invents things. The best-known failure is also the most instructive. In 2023 a federal court in New York sanctioned two lawyers who had filed a brief citing cases that did not exist; a chatbot had produced them, complete with plausible names and quotations. The tool did not malfunction. It did what such tools do when nothing constrains them: it produced fluent text that looked like an answer.
The defence against this is not a better prompt. It is a tool that will only say what it can point to, and that says "the documents do not answer this" when that is the truth.
It flattens disagreement. A summary is, by design, one account. If the file contains two accounts, a summary picks one, and you will not know which it dropped.
It forgets your corrections. If a lawyer fixes a misread date and the next run over new documents quietly puts the wrong one back, the tool is creating work rather than saving it.
It is wrong about jurisdiction. Research answers are only as good as the law behind them. Ask where a tool's legal content comes from and whether it covers the courts you actually appear before.
Courts in several jurisdictions have begun issuing guidance on the use of generative AI in proceedings, and more will follow. Check the current position for your own forum before relying on AI output in anything you file.
Eight questions to ask before you trust any legal AI tool
- Is every statement cited to a source I can open? Not "based on your documents", but a file and a page.
- What happens when the answer is not in the material? The right answer is an honest blank, not a plausible guess.
- Does it show me where sources disagree? Or does it resolve the disagreement for me without saying so?
- If I correct it, does the correction stick? Including after the analysis is re-run over new documents.
- Does it work in Arabic as well as English? In both directions, and on scanned documents, not only clean text.
- Where does my client's data go? Which country it is processed and stored in, and who can see it.
- Is my data used to train the vendor's models? Get the answer in writing, and check whether it covers the vendor's own AI providers.
- What is it actually for? Research, drafting or your case files. A tool that claims to be all three should be tested on all three.
Ask these of every vendor, us included, and ask for the answers in writing.
How CasEngine approaches it
CasEngine's AI is the third kind: it works on a case's own documents inside the matter, rather than answering questions about the law. It is not a legal research tool, and we would rather say so than let you find out.
In the AI case workspace, the documents for a case are indexed once and read three ways:
- A chronology. Every date in the bundle becomes an event carrying its file and page number. Where the same fact appears in several documents it is one event listing all of them.
- A map of the parties. Companies, individuals and authorities mentioned across the bundle, with the same company written different ways merged into one.
- Answers to questions. Drawn only from the case's documents, with every sentence cited. A sentence that cannot be cited is removed before you see it, and where the documents do not answer the question, it says so.
Where two documents disagree about the date of the same event, the timeline flags it rather than choosing one. When a lawyer corrects an event, the correction keeps its citation, records the original value, and is not overwritten when the analysis is re-run. A question can be asked in Arabic and answered from English documents.
That is a narrower promise than "legal AI" usually implies. It is also one a lawyer can check, page by page, which is the only kind worth making about a client's file.
If you would like to see it on a real bundle, talk to us.