Stop asking AI for cases. Start asking it to do the grunt work.
Ask a lawyer what AI is for and most will say "research". It is the one task the technology is worst at, and the reason many firms have tried it, come away disappointed, and quietly stopped.
Research is the worst thing you can ask it to do
Researchers at Stanford tested the leading models on specific, checkable questions about real cases. The models invented answers between 58 and 88 per cent of the time, and they did so with no sign of doubt.
The same team then tested the purpose-built legal research products that vendors sell to firms, the ones grounded in real case databases and marketed as the safe option. Lexis+ AI produced false or misleading answers more than 17 per cent of the time.
We all know why a fabricated citation is dangerous.
Look at what the tools actually do well
Anthropic, the company behind the Claude models, publishes data on how people actually use them. The single most common task across every conversation is writing and fixing code.
Measured across millions of real uses, automating work is what the model does best.
That matters for lawyers because of a quiet change most of the profession has missed. Until recently, writing a script to rename 5,000 files, pull the dates out of 200 contracts, or reconcile a discovery list meant finding someone who could code. Scripting sat behind a programming background, and most lawyers did not have one. That gate has gone. You describe the job in plain English, the model writes the script, and you run it. A controlled study of one such tool, GitHub Copilot, found developers finished a set task about 56 per cent faster with it than without. The point for lawyers is broader than coding speed.
The skill that used to require a programmer is open to anyone who can explain what they need in a clear sentence.
Why scripting is safe and research is not
The reason this work carries so little risk comes down to one difference between code and prose. Code can be checked. A script either runs and produces the right output or it fails in front of you. You can open the result, count the rows, and read the files it renamed. There is no confident lie buried in a column of dates. When the model writes a pattern to find every ABN across a folder of PDFs, you test it on ten documents and you know within a minute whether it works.
Put your AI use on the verifiable side and the worst case is a script that does not run, which costs you ten minutes. Put it on the trust side and the worst case is a fabricated authority in a document with your name on it, which costs you your standing.
What this looks like in practice
Take a job that lands on litigation desks all the time. You have 300 bank statements saved as PDFs, and you need a spreadsheet that lists, for each one, account number and date. By hand that is a day of dull work and a guaranteed crop of transcription errors. With AI it is a short conversation. You describe the fields you want and the format you want them in, the model writes a script that reads each PDF and writes a row, and you run it over the folder. Then you check it the way you would check any junior's work. Open ten of the PDFs, compare them against the spreadsheet, and confirm the script caught what it should. If a field is wrong you can see it, and you tell the model to fix the rule. The output is testable at every step, and nothing depends on trusting the machine.
When you run a script like this, AI actually has no involvement (except for writing the script). Turns out, the scripts and tools for work like this has existed for decades. Programmers have vast databases of scripts for automating aspects of their work without AI. Now lawyers can create these as well.
Where to start
The practical move is to reverse the usual order. Most firms began with research and drafting, the high-risk end, and either got burned or grew nervous and stopped. Start instead with one document-heavy job that has been waiting for someone to find the time. A document set that needs sorting. A back catalogue of precedents that needs tagging. A monthly report someone rebuilds by hand every cycle. Describe the task, let the ChatGPT or Claude write the script, test it on a small sample, and check the result yourself.
And before you trust any output, ask one question. Can I check this deterministically? If the answer is yes, you are on the safe side of the line. If the answer is no, treat it as research and verify it with your own eyes.
The mistake sits in the aim, not the tool. Point a model at the one task it does worst, trust it without looking, and it will fail you. Point it at the grunt work that fills your week, where every result can be checked, and you take most of the value with almost none of the risk. The lawyers who work this out first will not be the ones writing clever prompts for legal arguments. They will be the ones quietly automating the jobs everyone else still does by hand.