Local AI for law firms: working through a case file

A chronology you can check against the file

A useful application lets the person reviewing a timeline open the source page from an entry. They still check the facts, but they have less searching to do. A fluent summary offers limited help if tracing a date means reading through the entire PDF again.

Keep the first trial focused on documents from one matter that the firm has approved for this use. The task is to collect dates, explain what each one refers to, and preserve links to the sources. Searching across the firm's matters is a separate project, with decisions about who can access each file and when information from one matter can be used in another.

What three different dates actually tell you

Suppose a letter is dated May 3 on page 1. Page 2 contains a receipt dated May 5 and a handwritten “Date mailed” field that is illegible. The PDF's metadata, the information stored about the file, gives its creation date as May 6.

The application should keep those details separate:

The file creation date tells you what was recorded about the PDF. It cannot supply the missing mailing date or replace the date on the letter. The application is recording what the materials show. Determining the legal significance of those dates or calculating deadlines is outside this task.

Source references make the result easier to review. Someone can open page 2 and see why the mailing date remains unknown. Those references need checking too. If a citation leads to a page that does not support the entry, the reviewer is left searching for the evidence again.

When the scan is hard to read

A scan is an image of a page. Optical character recognition, or OCR, turns that image into text that software can search or pass to a model. An error can happen before the model starts organizing the chronology. If OCR reads a 3 as an 8, the model may receive the wrong date.

During a trial, the reviewer needs to see both the extracted text and the original page image. Include difficult scans and check whether the application identifies unreadable passages and makes them easy to revisit. If the source itself is illegible, someone needs a clearer copy or access to the original. A larger model cannot supply information that is missing from the material.

When something goes wrong, save the incorrect entry, the source the application cited, and the correct information. That gives the developer a concrete way to investigate whether the problem came from reading the scan, confusing the type of date, or linking to the wrong page. It also shows the firm how much correction and rereading the draft requires.

Will one person use it, or will the team share it?

For a person working through a file on their own, an application running on their computer may be a suitable starting point. Test the beginning of a normal work session, including launching the application and waiting for the model to load. Starting the trial with the model already running leaves that part of the experience unanswered.

A shared application has to handle several people asking it to work at once. Test that situation on the hardware intended for the firm, and record the equipment and number of users. Otherwise, it is difficult to tell whether the trial reflects how the team will use it. Our comparison of a workstation and a shared server for local AI covers that choice in more detail.

Foundry Local is one example of software intended to run a model on a user's device. Microsoft describes it as designed for one person accessing the model at a time. It is not designed to serve a whole firm simultaneously; Microsoft directs that use to dedicated server software (Microsoft Foundry Local documentation).

Running the model locally does not, by itself, establish the confidentiality of the whole process. The firm also needs to know where OCR reads documents, where originals and temporary files are stored, what appears in logs, and what connected applications receive. Model and component downloads need to be accounted for as well. For Foundry Local, Microsoft says that prompts and responses are processed on the device, while also describing network downloads and optional sharing of diagnostic logs (Microsoft documentation).

Deciding whether to commission an application

For a short, legible document, reading it manually or using an approved search tool may be enough. A local application is worth comparing on a larger file that requires repeated searching and organizing. If the firm permits a hosted tool, include that option using the same approved material.

Compare the complete job, including review. A draft produced quickly may offer little benefit if someone then has to find the sources and correct many of its dates. By the end of a trial, you should be able to see which documents the application handles well, where it needs help, and whether checking its work is easier than your current method.

A project with Syntalith could start with a local AI application that prepares a chronology for one matter. We would check with your team whether a lawyer can open the right page from an entry and trace any errors back to their cause. This would let you assess the tool on your own hardware. Agree on a scope that includes a working example and a record of where the application processes documents, the conditions it was tested under, and the problems found.

For an initial conversation, describe the part of reading a case file that takes the most time and whether one person or a team would use the tool. Selecting documents and agreeing on how they can be shared for a trial can come later. Service information is also available on Syntalith's pricing page.