Short answer
Legal AI is worth having when it retrieves the actual text before it answers, shows you the source of every statement, and refuses when it has no basis. A general assistant does none of those three, which is why it will produce an article of the Code civil that reads perfectly and does not exist. The engineering that fixes this is not exotic, but it has to be built in from the start.
Most articles about AI for lawyers are written by people who have not shipped one. We built Kateb, an AI legal assistant for Algerian notaires. It is live, it is used by practising professionals, and you can go and try it before reading another word of this.
So this piece is not a forecast. It is what we found out.
The failure that matters
Ask a general assistant for the article of the Code civil that governs a specific situation. It will give you an article number, a confident paraphrase, and a tone of complete certainty. Some of the time the article does not exist, or exists and says something else.
In most fields that is an annoyance. In legal work it is a professional liability event, because the citation is the thing that carries the weight. A wrong citation delivered confidently is worse than no answer at all, because no answer prompts you to go and look.
The useful question about a legal AI tool is not what it knows. It is what it does when it does not know.
What actually fixes it
- Retrieve before answering. The system searches an indexed corpus of the actual texts and pulls the relevant passages. It does not answer from memory.
- Answer only from what was retrieved. If the passages do not contain the answer, the correct output is that they do not, not a plausible reconstruction.
- Verify every reference before it is displayed. Each cited article is checked against the corpus. A reference that does not resolve never reaches your screen.
- Show the source next to the claim. Not a bibliography at the bottom. The specific passage, next to the specific sentence it supports, so checking takes seconds.
- Make refusal a designed output. "I have no basis for this" is a feature that has to be built, tested and measured like any other.
What this looks like in a practice
Works well
- Research across texts, with sources shown
- First drafts of routine acts and letters
- Extracting parties, dates and obligations from a contract
- Checking a dossier against a required-documents list
- Finding the precedent in your own archive
- Summarising a long document for review
Do not
- Advise a client without a lawyer reading it
- File anything unreviewed
- Rely on a citation you have not opened
- Feed privileged material into a consumer chatbot
- Assume foreign legal training applies here
That last one is the point most vendors skip. A tool trained mostly on French or American material is not neutral about Algerian law. It will reach for the wrong instrument confidently. The corpus is the product, which is the argument in "AI for Africa" is 54 different products.
If you want to see the standard we hold ourselves to, Kateb is live and free while in beta. Try it, then tell us what your practice needs.
Tell us what you are trying to fixHow to test a legal AI tool in ten minutes
Whoever you buy from, including us, run these before you commit.
- Ask it something that has no answer. Invent a plausible article number and ask what it says. It should decline. If it explains the invented article, stop the demonstration.
- Click a citation. You should reach the actual text, not a summary of it. If there is nothing to click, the citation is decorative.
- Ask the same question twice. Materially different answers to the same question mean it is generating rather than retrieving.
- Ask about a recent change. It should tell you the date its corpus was last updated. Any legal tool that cannot state this is not a legal tool.
- Ask where your documents go. Privileged material has to be treated as privileged, contractually and not just verbally.
In-house legal teams
For an in-house team the shape is different. The volume is in contract review, obligation tracking and answering the same internal questions repeatedly. The highest-value build is usually an assistant over your own contract archive that answers "what did we agree with this supplier about X" with the clause attached, rather than anything touching external law.
That is a narrower system, it is cheaper, and it is far easier to prove. We would generally build that before anything more ambitious.
Why we are a reasonable firm to ask
Not because of a claim about expertise. Because there is a product you can open right now that was built to this standard, in this jurisdiction, in Arabic and French, and it either behaves the way this article describes or it does not. That is a harder thing to fake than a case study.
Frequently asked questions
Can AI be trusted for legal research in Algeria?
Only if it retrieves the actual text before answering, verifies every citation against the corpus before displaying it, and refuses when it has no basis. A general assistant does none of these and will produce article numbers that do not exist. Ask any vendor to demonstrate a refusal, not just a good answer.
How do I test a legal AI tool before buying it?
Ask it about an invented article number and check that it declines. Click a citation and confirm you reach the real text. Ask the same question twice and compare. Ask when its corpus was last updated. Ask, in writing, where your documents are processed and what is retained.
Is it safe to put privileged client material into an AI tool?
Not into a consumer chatbot. In a properly arranged system you can establish where processing happens, which provider is involved, and what retention applies, and have that written into the contract. If a vendor will not put it in writing, treat that as the answer.
Will a tool built for French or American law work for Algerian law?
No. Algeria is a civil law jurisdiction with its own codes and procedure, and a system trained mostly on foreign material will apply the wrong instrument confidently. The corpus of local texts is most of the product, so it has to be built rather than translated.
What is the best first project for an in-house legal team?
An assistant over your own contract archive that answers questions about what you agreed and attaches the clause. It is narrower than legal research, cheaper to build, and much easier to prove before you extend it.
