Building LawManager: putting AI inside legal work, not beside it

Most legal software that added AI added a chat box in the corner. It answers questions about documents you have already read, in a window you have to go to on purpose. Nobody opens it twice.

LawManager is the AI-powered legal practice management platform we built and run. The design decision that mattered was not which model to use. It was refusing to put the AI in its own room.

The problem is volume, not intelligence

A practice does not lose time because lawyers cannot think. It loses time because a matter arrives as forty documents and somebody has to read all of them to find the six things that matter.

That is a volume problem. It is exactly what a model is good at, and exactly the thing a chat box cannot help with, because using it requires already knowing what to ask.

Where the AI actually sits

Inside document review. When documents land on a matter, extraction runs on them. The lawyer opens the file and the key dates, parties and obligations are already surfaced against the source text. No prompt, no separate screen.

Behind a human approval step. Nothing extracted becomes a fact in the system until a person accepts it. In legal work a confident wrong answer is worse than no answer, so the design assumes the model will sometimes be wrong and makes that cheap rather than invisible.

Never as the only record. The extracted value always points back at the page and paragraph it came from. If a lawyer cannot verify it in two seconds, it is not useful — it is a liability.

The rest of the platform, which is most of it

The AI is a feature. The product is a practice management system:

• Cases and matters with the documents, deadlines and correspondence attached to each
• Client communication through a portal, so the thread lives on the matter instead of in someone's inbox
• Billing tied to the work it came from
• Multi-user collaboration with role-based access, because a paralegal, a partner and a client are three very different permission sets
• A complete audit trail, which in this sector is not a feature request, it is the price of entry

What we learned that generalises

Grounding beats model choice. Almost every practical gain came from retrieval over the practice's own documents, not from a bigger model. Answers that cite the source are useful; answers that improvise are worse than nothing.

Design the failure case first. We built what happens when confidence is low before we built the happy path. A low-confidence extraction routes to a human rather than guessing, and the product is trusted because of that, not despite it.

Permissions are the hard part, not the AI. Role-based access across matters, documents and billing touched every screen in the system. The model integration was measured in weeks; the permission model was the thing that had to be right from the first commit.

Where it is

LawManager is live at lawmanager.io and we operate it. It is one of the products we point to when someone asks whether we have actually shipped AI into something people rely on, rather than demoed it.

If you are trying to put a model inside an existing product rather than beside it, that is the work we do in AI integrations.