The United States faces an access-to-justice crisis. Legal technology has emerged as a potential solution, with generative AI offering new possibilities for automating access to justice. But questions linger about whether AI can automate legal reasoning and whether such capabilities would expand access to justice or deepen inequities. Automation would seem to require an AI that possesses a “general” legal reasoning skill: the capacity to apply law to facts across different domains. But generative AI presents a conundrum. LLM capabilities form a “jagged technological frontier,” with LLMs excelling at some tasks while failing at others of similar or lesser complexity. LLMs lack general legal reasoning skills, and their outputs are unreliable for tasks that require accuracy and consistency. This might suggest that current AI is incapable of automating legal reasoning work. This Article reframes our conception of legal automation. Because LLM abilities form a jagged frontier, the frontier of potential automation is also jagged. Tasks within the frontier of LLM capabilities are ripe for automation, while tasks at the border or beyond are not. The challenge is identifying which tasks fall where. Because LLMs lack general skills, an LLM’s success with one legal task doesn’t predict its performance on similar tasks—it may succeed in applying the law to one set of facts and inexplicably fail with an almost identical set of facts. Creating reliable legal automation therefore requires extensive, careful testing and redesign. Because these evaluations are specific to individual legal domains and tasks, the jagged frontier of legal automation will become increasingly jagged based upon where investments are made. Investment in open-source research focused on low-income litigants’ legal issues can expand access to justice, while investment in closed, proprietary tools risks exacerbating inequality.