AI Document Assistant

This page describes a representative document assistant inside a case system. The client is not named. We do not claim a percentage of time saved. The work was an assistant that drafts a summary of documents already in the matter, for a person on the team to check. It does not give legal advice, and it does not file anything by itself.

The situation

Staff already had the documents. What they lacked was a first reading they could trust enough to correct. A long bundle arrived, and someone had to learn it well enough to tell a colleague what it was. People pasted text into a public chat tool because it was faster than reading. That was the problem we were hired to close, not a habit to automate. Client documents do not belong in a consumer chat box.

The firm wanted "AI" in the same sentence as the case system. A chatbot on the marketing site would not have read the bundle. The assistant had to sit where the documents already lived, under the same login, and it had to refuse questions about matters the user cannot open.

What we took on

The build is described on our AI integration page, and for a law-shaped workflow also on legal AI. We connected the model to the document store the firm already used. The assistant receives the text of files in that matter, not a scrape of the open internet, and not a shared corpus of every client's papers. A question is answered from those files, with the file name it used. If the files do not contain the answer, the draft says so.

A person remains in the loop. The summary is a draft in the matter. It is not an email to the client. It is not a note on the court file. Someone who knows the matter reads it, corrects it, and only then decides whether it is fit to use. We wrote that rule into the screen, not into a policy PDF nobody opens.

We logged what was asked and which documents were sent to the model, because a firm has to be able to say, later, what left the building. Retention of that log was the client's decision. We did not turn the log off to make the demo feel lighter. The first users were two people who already read bundles, not the whole firm on day one. A wider rollout waited until those two had rejected a bad draft and we had seen why.

What shipped

From a matter, a staff member can ask for a short summary, a list of dates mentioned, or a question about a document in that matter. The answer shows its sources. The draft can be edited and saved beside the file. The model is not offered a button marked "send to client." That button is how a confident wrong sentence becomes the firm's letter.

What we refused

We did not train a model on the firm's archive. We did not let the assistant browse other matters. We did not describe the output as advice. A summary can omit the sentence that matters, and a date extraction can misread a scanned page. The check is the product. We also did not quote a time saving. Anyone who wants a number needs a before-and-after study on their own files. We did not have one, and we will not invent one.

What a similar project needs from you

A document store with permissions that already mean something. A written rule about what may be sent to a model. One partner or manager who will read ten drafts and say which errors are unacceptable. If those three are missing, the project is not ready, however urgent the demo feels.