Cobra Developers
Blog

Legal document automation that lawyers actually use

The first piece of software we ever built generated court filings. An attorney typed a client's name, chose a motion, and got back a formatted document with the right court, judge, case number, and caption already filled in. It saved hours a week, so it grew. Today it is part of a practice management system with 37 document types.

Along the way we learned why most document automation in law firms gets bought, demonstrated, and then quietly abandoned. The reasons are not about the technology.

Why it usually fails

It starts from a blank form. A tool that asks the attorney to type the case number, the court, the judge, and the charges has saved almost nothing. The firm already has that information in its case system. Retyping it is the work the attorney wanted to avoid.

The output needs fixing. If a generated motion has the wrong caption format for that court, or the signature block in the wrong place, the attorney has to open it and repair it. After the third time, they go back to copying last month's motion and editing it by hand, because at least they know where the mistakes are.

It lives somewhere else. Every extra login is a reason not to use a tool. If generating a document means leaving the case file, opening another application, and searching for the client again, it will lose to habit.

What makes it stick

Start from the case, not the document

Our generator opens with a search box. Type a client name, a case number, or a phone number, and it finds the matter. Everything the document needs is pulled from the case record. The attorney's first decision is which document they want, not what the case number is.

It is also available as a tab inside every case, so there is no search at all when you are already looking at the file.

Get the formatting exactly right

Courts have conventions, and attorneys notice immediately when a document breaks them. We ported the firm's existing language word for word, and matched captions, spacing, and signature blocks to what the firm was already filing. Legal text is not something to paraphrase.

The standard we work to is that a generated document can be filed without being opened. Anything short of that is a draft, and attorneys already have drafts.

Use AI narrowly, and say where

Some parts of a filing are fixed: the caption, the standard of review, the certificate of service. Other parts depend on the facts: the middle paragraphs of a motion, the rationale in a request.

We use a language model only for the second kind, and the attorney reviews those paragraphs before anything is filed. The boilerplate never passes through a model, because boilerplate has to be the same every time.

That split is the answer to most questions about AI in legal work. Use it where judgment is being exercised and will be checked by a person. Keep it away from text that must not change.

Put the result where it belongs

A generated document that lands in a downloads folder still has to be renamed, moved, and attached to the case. The generator files it with the matter automatically. Small steps like that are the difference between a tool that saves time and one that moves the work around.

Reducing dependencies is a feature

The original version of our generator relied on five outside services: a hosting platform, a spreadsheet database, the firm's practice management system, a separate database, and cloud file storage. Each had its own login, bill, and way of failing.

When we rebuilt it inside our own practice system, all five went away. Nothing about the documents changed. What changed is that there is one place to look when something breaks, and one bill.

For a small firm, that matters as much as any feature. Software that depends on five vendors is fragile in ways no single vendor will warn you about.

Automation beyond documents

The same thinking applies to the rest of a firm's paperwork. We built a system that reads the firm's incoming email, files attachments to the correct client, routes discovery, and compares any court date mentioned in a message against the calendar. A date that appears in an email but not on the calendar gets flagged.

None of that is glamorous. All of it is work that used to consume a staff member's morning, and every missed calendar entry it catches is a hearing that does not get missed.

How to tell if it is worth building

Ask three questions about any document your firm produces regularly:

  1. Is most of it the same each time? If so, it can be generated.
  2. Does the information that changes already exist in your case system? If so, nobody should be typing it.
  3. Would you file the result without editing it? If the honest answer is no, the template is not finished.

If you answered yes to the first two and want to get to yes on the third, that is the work we do. Tell us which documents eat your week.

guest@cobra: ~