Paralegal — tasks, one by one

First-pass document review

Relevance classification over large document sets is exactly the shape of problem language models handle well, and the legal industry had already adopted predictive coding for e-discovery before the current generation of models. The remaining human step is verifying the model's misses on a sample, not reading everything.

Privilege calls and anything that will be sworn to still route through a human, because the professional liability does not transfer to a vendor.

Drafting from precedent

Generating a competent first draft from a precedent bank and a fact pattern is now fast, but the value was never in the typing. It sits in knowing which precedent, which clauses the other side will fight, and what this client has agreed before — context the tool only has if someone maintained it.

The context that makes the tool useful — which precedent, what this client agreed before — is firm knowledge that someone has to keep maintaining. Firms that automate the drafting without funding that maintenance get worse drafts faster.

Running the matter

Practice-management software tracks the calendar, but the calendar is not the job. Knowing that a particular registry rejects filings for a formatting quirk, or that opposing counsel always asks for an extension on Fridays, is the operational knowledge that keeps matters moving, and it lives in people.

Operational knowledge lives in people until someone writes it down — and the pressure to write it down is exactly what automation creates. This task is protected by an undocumented state that will not stay undocumented.

Client and witness contact

The barrier is trust and confidentiality rather than language. People disclose awkward facts to a person they have met, and a firm cannot delegate the duty of confidentiality to an interface it does not control.

Confidentiality duties constrain which interface may be used, not whether a person is needed. Firms are already procuring tools that satisfy those duties, and when they do, the argument moves.

Checking what the tools produced

Fabricated citations in filings have already produced sanctions in several jurisdictions, and firms responded by making a human sign off on machine-assisted work. That sign-off is a new task, and it lands on the person who already checks everything else.

Whether this task is paid as skilled work or absorbed as unpaid diligence depends on the firm, and that decides whether it is an opportunity or just more load.