The American legal profession has always run on a simple bargain: young lawyers suffer through years of document review, and in exchange, they learn the craft. Sifting through thousands of pages of contracts, emails, and memoranda in the discovery phase of litigation was tedious, yes, but it was also how associates learned to spot the smoking gun, to understand how corporate malfeasance actually looks in the wild. That bargain is now breaking down.
Artificial intelligence has been quietly transforming e-discovery for over a decade, but the latest generation of large language models has accelerated the shift from augmentation to replacement. What once required teams of associates billing hundreds of hours can now be accomplished by a handful of senior lawyers supervising machine-generated document classifications. The economics are irresistible: a task that might have cost a client two million dollars in associate time can be completed for a fraction of that sum.
The pyramid problem
Big law operates on a leverage model. Partners generate business; associates do the work; the ratio between them determines profitability. Discovery was the foundation of this pyramid—reliable, voluminous, and difficult to offshore. When predictive coding and technology-assisted review first emerged, firms marketed them as cost savings for clients while quietly maintaining staffing levels. The new AI tools are harder to disguise. They don't just find relevant documents faster; they can summarize depositions, draft privilege logs, and identify patterns across millions of records that human reviewers would miss.
The uncomfortable question is what happens to the associates who no longer spend their first years in document dungeons. Some firms argue that freeing junior lawyers from drudgery allows them to do more substantive work earlier. Critics counter that there is no substitute for the pattern recognition that comes from reading thousands of bad documents before you encounter a good one. The legal profession has always been an apprenticeship trade, and apprenticeships require menial tasks.
What AI still cannot do
For all its capabilities, AI remains poorly suited to the aspects of litigation that actually win cases. It cannot read a room during a deposition. It cannot sense when a witness is lying in a way that doesn't show up in the transcript. It cannot make the judgment call about whether to pursue a line of questioning that might alienate a jury. The strategic and interpersonal dimensions of trial work remain stubbornly human.
The technology also struggles with novelty. AI systems trained on existing case law can identify precedents with remarkable speed, but they are less adept at constructing genuinely creative legal arguments—the kind that change how courts interpret statutes. The most consequential litigation often turns on persuading a judge to see a familiar problem in an unfamiliar way. That remains the domain of human ingenuity, at least for now.
Our take
The legal profession's anxiety about AI is really anxiety about its own business model. For decades, law firms have charged for inputs—hours worked—rather than outputs—problems solved. AI exposes the inefficiency that clients have long suspected but couldn't prove. The firms that thrive will be those that figure out how to train lawyers without discovery drudgery and how to charge for judgment rather than time. The technology isn't eliminating lawyers; it's forcing them to articulate what, exactly, makes their judgment worth paying for. That conversation is overdue.
