Why the drafting moves faster than the job
Paralegal work sits close to text. Researching statutes, regulations and case law is reading and summarizing. Preparing pleadings, motions and discovery responses is assembling language from precedent and a file. Software is good at both, which is why the question of whether AI will replace paralegals comes up in almost every firm that has tried a legal research tool.
The rest of the job is less portable. A paralegal interviews clients and witnesses, and the facts change mid-case. Deadlines and filing rules differ by court, and a missed one has consequences a tool cannot absorb. Someone has to sign, and someone has to answer to the supervising attorney when a citation turns out not to exist.
There is also almost no physical component here. Our robotics view puts this job in the tier that needs no hardware at all, so the pace is set by software and firm policy, not by machines. That cuts both ways: adoption can arrive quickly, but it arrives as a change in how hours are spent rather than as an empty desk. The sharper pressure is on the entry rung, where billable research and document review used to be the training ground. What AI is doing to entry-level work is the part worth watching.
Which tasks AI handles, which it assists, and which stay with you
Start with the share of task time our scoring puts in the automation column: 10%. That is the first-pass stuff. Keyword and citation searching across case law databases. Sorting and indexing large document sets so the relevant material surfaces first. Both are jobs where speed and volume matter more than judgment, and both used to eat whole afternoons.
A larger slice sits in the assist column: 38%. Drafting a motion or a discovery response is the clearest example — the model produces a serviceable skeleton, the paralegal checks every cite, fixes the facts and shapes it for the court it is going to. Summarizing depositions and medical records works the same way: a draft summary saves time, and a human reads the transcript to make sure nothing load-bearing was dropped.
Then the share that still needs a person: 52%. Interviewing a client who is distressed or unsure. Gathering and arranging evidence in a form that will survive challenge. Calling a clerk’s office about a local rule that is not written down anywhere useful. If you want the definition of the task-time figure, the coverage method page explains how it is built.
Good to know: the cost panel above compares tool spend against staffing spend for the same slice of work, and the gap is the main reason firms pilot these tools even when quality is unproven.
What has actually been tested
Not much, in this job specifically. Our evidence grade is D, which is our marker for work where no study has tested AI output against the output of qualified paralegals on real caseloads. So we publish no parity number for this occupation, and you should treat any figure claiming one with care. The quality-parity method page sets out what each grade means.
What would settle it is specific and testable: a blind comparison on real discovery sets, with supervising attorneys scoring draft pleadings and deposition summaries they do not know the source of, plus rework and error rates tracked over months rather than a single sitting. Citation accuracy would need to be measured, not assumed. Until something like that is published, the honest answer is that the tools demonstrably help and have not been shown to match a trained paralegal end to end.
The labor market numbers are firmer. The US had about 392,880 paralegals and legal assistants, with median pay of $62,890 (BLS, 2025), and the federal projection for 2025 to 2035 is roughly flat at -0.3%. That is a profession holding its size while its daily work changes, not one in retreat.
When the balance could shift
Most likely between 2036 and 2050 (8 in 10 of our scenarios). The replacement-year method page explains what that window is measuring and how the spread is produced.
Two things could pull it earlier. Legal research platforms are embedding drafting features directly into the systems firms already pay for, which removes the procurement hurdle. And billing pressure from clients who refuse to pay for first-pass review pushes firms to automate exactly the tasks in the first column above.
Two things hold it back. Accuracy liability is unresolved: fabricated citations have already drawn sanctions, and a supervising attorney’s signature means a human has to verify everything. And the work is procedural as much as textual — court-specific filing rules, service requirements and deadline tracking vary enough that a general tool gets them wrong in ways that cost a case. Firm culture and malpractice insurers move slowly for good reason.
How to stay needed as a paralegal
Lean into the tasks in the human column. Client and witness interviewing, where you get facts nobody has written down yet. Evidence handling and trial preparation, where organization under pressure is the whole skill. Court and clerk relationships, including the procedural knowledge that keeps filings from bouncing.
Two skills are worth real effort. First, verification: knowing how to check a cite, a date and a quote fast enough that a draft from a tool becomes safe to send. Second, case management judgment — owning a matter’s deadlines, documents and priorities well enough that attorneys route decisions through you.
Nearby work worth comparing: title examiners, abstractors and searchers, judicial law clerks and lawyers. The wider legal support workers family shows how the scores differ across the support roles, and the law firms sector page covers the setting most of these jobs sit in. You can also put this job beside any other on the compare tool, check where it falls on our list of jobs most at risk, or read how all three scores are built in the methodology.