AI legal research in Kenya: what works, and what to never trust

Reading long documents is where AI earns its place. A 60-page judgment, a 30-page lease, a bundle of payslips and warning letters — it can pull out the holding, the obligations and the dates in seconds, and point you to the paragraph each one came from.

It is also good at translation of register: turning statutory language into a sentence a client understands, and turning a client's story into the legal questions worth researching.

Most general chatbots were trained on far more American and British material than Kenyan material. Ask about notice periods or succession and they will often answer confidently with another country's rule, or cite a section number that does not exist.

They also invent case names. A citation that looks perfectly formatted is not evidence that the case is real — only opening it is.

The fix is not a bigger model, it is grounding: the tool should search actual Kenyan legislation, judgments and your own uploads, then answer only from what it found. That is how CHAT CLYDE SLP is built — the answer is assembled from retrieved sources, and those sources are shown to you.

A grounded answer from a modest model beats an ungrounded answer from the biggest model available, every time, for legal work.

Open the cited source. Search for a distinctive phrase from the answer. Read the sentence around it. If it matches, you can rely on it; if the source says something else, you have caught the error before it reached a client.

Do this every time and AI becomes a research assistant rather than a risk.

Judgement, strategy and responsibility. AI can tell you what the Employment Act says about a hearing; only you can decide whether to file, negotiate or advise your client to wait.