There is a temptation, in any discussion of legal automation, to treat the arrival of large language models as having settled the question. It has not. It has made the question sharper, because the two available models fail in opposite directions.
Three waves, not one
Document automation in legal practice has moved through three reasonably distinct phases. The first was the static template: a document with blanks, filled by hand, offering consistency but no logic. The second was rule-based boolean logic: conditional clauses, decision trees, validation rules – a system that produces exactly what its author specified and nothing else. The third is generative: a model that produces fluent legal prose on request, without having been told the rules.
Each wave did not replace the last. All three are in production use today, frequently in the same firm.
What each model guarantees
A deterministic system offers something valuable and unusual: if it is correctly specified, its output is predictable. The same inputs produce the same document, every time, and the document contains only clauses a lawyer approved. What it cannot do is handle a situation its author did not anticipate. Confronted with novelty, it produces nothing useful.
A generative system inverts this. It handles novelty gracefully and drafts around gaps in its instructions. But it offers no guarantee that the clause it produced is the clause you would have approved, and it can be confidently wrong in ways that read perfectly. In a context where a single mis-stated obligation is a liability, fluency is not the property you most need.
The hybrid case
The practical conclusion is that the optimal industrial arrangement is a hybrid: deterministic logic for the parts of a document where certainty is the requirement – parties, obligations, amounts, governing law, anything with legal consequence – and generative assistance for the parts where flexibility genuinely helps: summarising, adapting register, drafting the covering explanation, handling the unanticipated variant.
The design question is therefore not “which model” but “where is the boundary”. Drawing that boundary is a legal judgment, not a technical one, which is precisely why it cannot be delegated to engineers alone – and precisely why we teach it as an interdisciplinary exercise.
A note on incentives
It is worth naming the barriers honestly. Automation collides with the billable hour; training on client material collides with privilege; and institutional inertia is a rational response to professional liability. These are not technical problems and no model release will resolve them. They are the reason adoption lags capability by years, and they belong in the analysis from the beginning.