Where AI Actually Excels
The tasks where legal AI has achieved genuine excellence are well-documented: first-draft generation for standard agreements, legal research on well-established legal questions, contract clause comparison, discovery document review, and translation of legal documents across languages.
What these tasks have in common: they are primarily lookup and pattern-matching operations where correctness is measurable and training data is abundant. The patterns are learnable. The performance is excellent.
Where AI Still Fails
The work lawyers find hardest — and charge the most for — is not lookup and pattern-matching. It is judgment under uncertainty with high stakes: advising a board on whether to litigate or settle a bet-the-company case; structuring a cross-border transaction; counseling a client in a regulatory investigation where the law is unsettled.
AI is poor at these tasks for reasons that are not easily fixed. The training data does not reflect the full texture of professional judgment — it reflects the written outputs of professional judgment, which is different. A merger agreement reflects a deal lawyer's judgment, but not the forty strategic conversations that shaped it.
The Narrow Path to Specialization
The legal AI companies that will close the specialization gap are the ones investing in three things: practitioner feedback loops that capture tacit professional knowledge; matter outcome data that lets the AI learn what advice correlated with good outcomes; and extended context architectures that can hold the full complexity of a long matter.
These are solvable problems, but they require access to data that is either privileged, proprietary, or not yet digitized. The companies that figure out how to get that data will have an insurmountable advantage.
A Realistic Timeline
Based on current trajectories, our view: legal AI will achieve genuine competence on straightforward advisory tasks within 18-24 months. On genuinely complex advisory work — the tasks that currently command the highest rates — the timeline is five to seven years, if the data access and feedback loop problems are solved. If they are not solved, it may be a decade.