On 28 October 2026 the Chief Justice said the judiciary would issue a practice direction on the use of artificial intelligence by lawyers and court officers “within the next few weeks”. Judges already have their own guidelines. His framing was the one every professional body has landed on: technology can speed up the administration of justice, and it must never replace human judgment.
The Malaysian Bar got there earlier. It issued a first advisory on generative AI in 2023 and a substantially expanded circular in July 2025 that separates sensible uses, such as early drafts and brainstorming, from uses to avoid, such as generating legal advice or verifying case law. It updated its guidance again after a High Court rebuke over citations to cases that do not exist.
So the risk everyone is watching is clear: a language model inventing law. The useful question for a firm is where the work is that carries none of that risk.
Two kinds of legal work
Some legal work is judgment. What does this clause mean, will this argument succeed, what should the client do. A model can assist a lawyer here, and the lawyer must check every word.
A lot of legal work is assembly. Take these 200 accounts from the bank’s spreadsheet. For each one, produce the correct notice with the correct name, account number, amount and date. Track which stage each one has reached. Compute the interest that has accrued since judgment. Produce the statement in the bank’s format.
Assembly is not a job for a chatbot. It is a job for deterministic software: the same input always produces the same output, and the output can be proven correct. There is nothing to hallucinate because nothing is being generated from memory.
What we built for a banking litigation practice
Clausework is a platform we built for a Malaysian firm that runs recovery work for a bank. The firm’s previous system was an unstable, locally hosted application, and every letter meant copying fields into a Word template.
Now the bank’s monthly workbook is uploaded and parsed into typed records for each product: mortgage, credit card, hire purchase, personal loan. Notices, summonses, affidavits and bankruptcy documents are rendered from the firm’s own Word templates. If a template refers to a field the record does not have, the document is flagged instead of going out with a blank. For execution proceedings, an interest engine computes post-judgment interest to the cent for conventional and Islamic facilities and exports the statement in the bank’s format.
Filing and checking time fell by 95%. The firm went live six weeks after kickoff.
Where AI does appear, and how it is fenced
There is one AI step in the platform. Statements of account arrive as documents, and a model reads the figures out of them. The extracted numbers go to a person to confirm before they are used in any calculation.
That pattern is the whole discipline:
- AI reads documents into structured fields. It does not write the law.
- Every extracted value is shown to a person next to the source.
- Calculations and document assembly are done by ordinary code that can be tested. The backend has more than 400 automated tests.
- Every generated document records which template, which data and which user produced it.
A practice direction on AI is unlikely to trouble a workflow like this, because no AI output reaches a court without a person and a deterministic system in between.
What to do before the direction lands
- List where AI is already used in the firm. Including the tools individual lawyers use on their own.
- Sort the work into judgment and assembly. Automate assembly with deterministic systems. Keep generative AI as a drafting aid for judgment work, with verification.
- Check where client data goes. A public chatbot is a third party. The amended Personal Data Protection Act now requires breach notification and a data protection officer, and clients will ask.
- Keep a record. Which tool, for what, checked by whom. If a court asks how a document was prepared, the answer should take a minute to find.