In this release, Paralegals and Legal Assistants have an exposed task share of 33.6%. Lawyers have 22.0%. The paralegal role is more exposed than the role that supervises it, by a margin of eleven and a half points.
The obvious explanation is seniority: junior work goes first. The obvious explanation is wrong, or at least incomplete, and the task-level data shows why.
Legal work splits along a line that is not the org chart
Sort every legal task in the index by exposed share and a pattern appears immediately. The exposed end is research, summarisation, document review, drafting from precedent, and records management. The unexposed end is appearing, advising a client in the room, negotiating, taking instruction, and testifying.
Representing a client in legal proceedings scores 1 out of 4 on capability and carries near maximum presence and accountability. Testifying scores 0. These are not difficult tasks in an intellectual sense. They are tasks where a human being has to be the one doing it for the act to count at all.
Paralegal work is concentrated at the first end. Lawyer work is spread across both, and the presence-bound portion pulls the weighted average down.
Which means the split runs through the lawyer's own day
The more useful way to read this is not as a comparison between two job titles. It is a statement about which hours in a legal career are in play.
A lawyer whose practice is predominantly research and drafting is, in the terms of this index, doing paralegal-shaped work regardless of what the business card says. A lawyer whose practice is predominantly advocacy and client counsel is in a different position entirely. The index scores occupations because that is what the underlying taxonomy provides, but every occupation page publishes the task-level detail precisely so the occupation number can be taken apart.
What this does not settle
The index measures what current systems can produce and what stands in the way. It does not measure how many paralegals a firm will employ, which depends on whether firms pass efficiency gains through as reduced headcount or as more work done, and on the billing model, and on professional rules about supervision.
There is also a well-documented failure mode specific to this field. Systems that produce fluent legal text with fabricated citations have already caused sanctions in real proceedings. That is captured in our verification cost dimension, which rates legal research tasks near the top of the scale: the output is cheap to produce, expensive to check, and damaging when the check is skipped. High capability and high verification cost together do not mean the work is safely automated. They mean the work is cheaply produced and expensively wrong.
Figures from release v2026.Q3, built on O*NET 31.0. Task-level scores for both occupations are published in full on their pages, and every rating decomposes into six components described on the methodology page.