Legal AI in Pakistan: What It Can Do for Advocates, and Where It Stops
Legal AI has arrived in Pakistani practice the way most tools do: quietly, one chamber at a time, before anybody agreed on what it is for. Some advocates use it daily. Some have tried it once, received a confident answer citing a judgment that does not exist, and stopped. Both reactions are reasonable, and neither is the whole picture.
This entry is a plain account of what legal AI can genuinely do for an advocate practising in Pakistan, where it fails, and the habits that separate useful work from risk.
What "legal AI" actually means
Most legal AI today is built on a large language model: software trained on very large amounts of text, which produces fluent writing in response to an instruction. On its own, such a model knows a great deal about language and a limited, unreliable amount about any particular country's law. It has no built-in sense of which statute is in force, which court has jurisdiction, or whether a judgment it recalls was ever reported.
The difference between a general chatbot and a legal AI tool lies in what is built around the model. A tool designed for legal work typically adds three things: a body of real legal material the model is made to work from, such as statutes and judgments; structure, such as templates that fix the parts of a document that must not vary; and checks that test the output before it reaches the advocate. The quality of a legal AI tool is mostly the quality of those three layers, not of the model underneath.
Where it helps
First drafts. A large share of practice is documents with a settled shape: plaints, applications, notices, agreements, petitions. Legal AI can produce a competent first draft of these quickly, leaving the advocate to do what only an advocate can do, which is to judge whether the draft serves the client's case.
Organising material. Clients arrive with a phone full of screenshots, a bundle of photocopies and a voice note. Turning that into a coherent chronology is slow clerical work. AI is well suited to reading mixed material and arranging it into a structured summary for the advocate to check.
Finding the starting point for research. Asked a legal question in plain language, a well built research tool can search statutes and judgments and return the provisions and authorities most likely to matter. The value is not that it answers the question. It is that it shortens the time to the documents that do.
Reviewing a document. A second reading for gaps, omitted reliefs, procedural defects or inconsistent figures is useful on any draft, and more useful on the other side's.
Where it fails
Invented authority. The most serious failure is fluent fabrication: a citation in the correct format to a case that was never decided, or a section number that is nearly right. It reads exactly like a real citation, which is why it survives a quick reading. Any tool that answers from the model's memory rather than from retrieved material carries this risk.
Wrong jurisdiction. Pakistan has federal law, the law of the Islamabad Capital Territory, and the law of four provinces, and the same subject is often governed differently in each. Rent, court fees, stamp duty, forum and procedure all change across the line. General purpose AI tends to blur these, borrowing a provision from one jurisdiction for a matter in another. The document reads well and is filed in the wrong forum.
Stale law. A model trained at one point in time does not know about amendments made after it. Without access to a current text, it will describe the law as it was.
False completeness. Given an incomplete instruction, a model tends to fill the gaps rather than admit them: a figure, a date, an annexure that was never mentioned. The result looks finished, and the gaps are hidden inside it.
Questions to ask of any legal AI tool
- Does it work from real statutes and judgments that you can open and read, or only from what the model remembers?
- Does it know which jurisdiction you are practising in, and does that choice change the court, the forum and the governing law in the output?
- When it cites an authority, can you read the authority itself?
- When it does not know something, does it say so, or does it fill the space?
- What happens to your client's information once it is submitted?
A tool that answers these well is a working aid. A tool that cannot answer them is a writing aid, and should be treated as one.
Habits that keep the work safe
Verify every authority before relying on it. Open the judgment. Read the section. The duty to the court is the advocate's, and it does not transfer to software.
Treat the output as a draft. Read it as you would read a junior's draft: carefully, and with the expectation that something needs to change.
Fix the jurisdiction first. Before anything else, confirm the court, the forum and the governing statute are correct for the matter. Most serious errors begin there.
Protect the client. Do not put a client's confidential material into a tool whose handling of that data you do not understand. Read the privacy terms before the first matter, not after.
Keep your own judgment in charge. AI can arrange, summarise and suggest. Deciding what to argue, what to concede and what to file remains the advocate's work.
The courts' position
Pakistan's higher judiciary has already addressed the use of AI in judicial work, and the standing position is that these tools may support the people who run the justice system but cannot replace the human reasoning the law requires. That position, its constitutional basis and the duties it places on advocates are set out in a separate entry: Proper Use of AI in Pakistani Courts.
Where this leaves the profession
Legal AI in Pakistan is neither a gimmick nor a replacement for the advocate. Used well, it removes a large amount of slow, repetitive work and gives time back for the parts of practice that require judgment. Used carelessly, it produces documents that read perfectly and are wrong. The difference is rarely the model. It is whether the tool was built around real Pakistani law, and whether the advocate using it still checks the work.