In two prior pieces (I, II), I wrote about a case I have been working on where the California State University Employees Union (CSUEU) unilaterally extended the probationary period of one of its staff members. The staff member’s union, the United Auto Workers (UAW) Local 2350, grieved this action and then demanded arbitration over it. CSUEU refused to arbitrate. UAW filed a Section 301 lawsuit against CSUEU to secure an order requiring CSUEU to arbitrate. The parties eventually went to arbitration and the UAW prevailed: the arbitrator determined that CSUEU violated the parties’ collective-bargaining agreement (CBA) by extending the probationary period and ordered CSUEU to recognize the staff member as a permanent employee and provide make-whole monetary relief.
This was a fairly straightforward case. There was never any doubt that CSUEU was violating the CBA. This is what made the case so strange. It was immediately obvious that the UAW would prevail and that CSUEU was wasting a bunch of its members' money on pointless litigation.
The straightforward posture of the case presented another interesting possibility and first for me. By the time the arbitration rolled around, it seemed pretty likely that I would be able to call on AI to draft my brief.
I use AI quite a bit in and around my legal practice. I used it to help build NLRB Research, I sell a suite of AI skills that practitioners can use to quickly research labor law questions, and I use it to produce case summaries for this newsletter. But, for a variety of reasons, I’ve never used it to help generate any sort of significant legal writing such as a brief. I have found it to be useful as an editor for such legal writing. But generating the legal writing in the first instance is a different, more complicated, matter.
This arbitration had some unique features that made me think it was worth a shot:
The case was entirely based on documentary evidence. The CBA is a document. The decision to extend the probationary period was submitted via email. The grievances and arbitration demands were also submitted by email.
The parties had already gone through a federal lawsuit over the question of whether CSUEU was required to arbitrate the grievance. The lawsuit resulted in a docket of complaints, motions, exhibits, and so on that overlapped with the arbitration.
The parties agreed to a stipulated record in the case. This means there was no hearing, no transcript, or anything like that. The record consisted of a short statement of mutually-agreed facts and a PDF containing 18 joint exhibits.
It was a simple contract interpretation case in a labor arbitration. This means it was not really necessary to cite any case law. The arbitrator could just read the CBA and determine what it said about our case.
Given my extensive experience with AI, it was pretty clear to me that, in these circumstances, if you pointed an AI harness at a case file that contained all of these documents, it should be able to draft a decent arbitration brief. So I prompted Claude Code running Fable 5.1 to do so and it did. The resulting brief was not perfect. In general, I’d say it had a tendency of being too exhaustive, meaning that, by default, it would include every conceivable argument. Some lawyers do this, but it is not my style. Of course, it was easy enough to read the brief and prompt the AI harness to cut out stuff I didn’t think was necessary and to make other desired changes.
Below is the brief that I submitted.
One thing I want to highlight in this brief are the citations. There are over 80 citations to the stipulated record and five citations to Elkouri’s How Arbitration Works. These citations were all written by the AI harness. I simply converted the stipulated record and How Arbitration Works into clean markdown files with all of the pagination clearly indicated so that the LLM could bring them into context and cite them accurately. Once the brief was cleaned up, I had the AI harness create a JSON inventory of every single citation in the document and then spawn subagents with clean context windows to double-check every single citation against the source.
As noted already, we won the case. The main reason for this is because we were clearly in the right, not because this arbitration brief is especially brilliant. But the AI brief was clearly adequate for the task. And there are a lot of legal tasks like this where AI writing will be adequate, so long as it is being managed and supervised by someone who has the necessary legal and technical abilities. This is not total automation of legal work. But it is much more efficient legal work that generates the same outcomes in a fraction of the time.