
After reading a comparison of 8 legal AI tools, I went back and tweaked WorkBuddy's batch contract review config again
A few days ago I came across that 2026 legal AI tool comparison, 8 mainstream products, covering everything from retrieval to document review, case analysis, and drafting. I read the opening line twice: domestic legal AI tool users have passed 5 million, with 72% penetration in law firms. The specs look good too — AlphaGPT has 190 million cases and 5.8 million-plus statutes connected, and it's completed filing.
After reading it I didn't plan to switch tools. Not out of laziness — it's that I care more about something else: once you have the retrieval results in hand, that organizing step is the biggest time sink. This kind of comparison answers who's smarter, but it doesn't answer what the output of that step should look like.
I've been using WorkBuddy for over two months, and the pitfalls I've hit are basically all on this. Today I'll write out my batch contract review setup to save some trouble for people just getting started.
How I got it wrong at first. I had a labor arbitration case with a dozen-plus related contracts to go through. The first time, to save effort, I dumped the whole requirement in one sentence: extract risk clauses, give revision suggestions, classify by level, export a summary table. What came out was a document that was neither fish nor fowl, with a different format every time, and risk levels sometimes written as high, sometimes as major.
Later I figured it out — the problem wasn't the model. If you don't lock down the output format, you might as well not have written anything. It doesn't know whether you want a table or a paragraph, so it improvises.
Here's how my setup goes. The environment is the Windows desktop version of WorkBuddy; interfaces may differ slightly by version, so it may not apply to everyone.
Open WorkBuddy, in the left Projects column click New Project, name it "Contract Review - Labor Arbitration". Inside there's a Materials area — drag the PDFs to be reviewed straight in, batch drag is supported. Don't skimp on this step; get the files organized first and you'll save a lot of trouble later.
Then the key part: click Instruction Templates in the top right → New Template. In the template I fix three things. The output format is a table with a fixed five-column header: clause location, original text excerpt, risk level (only fill in High/Medium/Low), revision suggestion, basis. File naming follows {contract number}_review_{date}.docx, with the number taken from the original filename. Execution granularity is one thing at a time — first produce the risk list, and after I confirm, revise item by item per the list.
Once the template is saved, in Tasks select this template, check the entire materials folder under Execution Scope, and click Start. The expected result is one table per PDF, with field order completely consistent. The first run produced 13, formats aligned, but it still missed two spots: the dispute resolution clause and the notice service clause. This is exactly the same problem I ran into when organizing invoices before — edge information like the remarks column is the easiest to skip. So I later added a line at the end of the template: these two types of clauses must be listed whether or not there's risk; if none, write "none".
In project settings there's Collaboration. I locked template editing permission to the admin, i.e., myself. If anyone can edit the template, it's the same as no one editing it. Assistants and interns only get Can view and Can comment — they can see results and raise opinions, but can't touch the fields. This logic is the same as how I split duties for storing tokens in 1Password: narrow the permissions, and responsibility becomes clear.
In daily use, every Monday I spend ten minutes on two things: go through last week's output to check for missing fields; and log a version number each time the template changes, noting what changed. Last month I got to the third version, and looking back at the first version's notes, I finally understood why I configured it that way back then.
If I were to give advice: tonight, just create a project, throw three similar contracts in, and first let it do only one thing — extract risk clauses — and see whether the output format is right. If it doesn't match, change the template, don't rush to change the prompt. Once the template is right, only then does batch efficiency become a real conversation.
Physix Frontier