Skip to content
man bot
Compare

A closer look

Paralegal

AI can do the first pass on ten thousand documents. Someone still has to spot the one that doesn't fit.

Drafting legal documents and preparing property and share transfers are where AI could take a fair chunk. Keeping order in court and investigating theft it barely touches, which pulls the average back up.

61 / 100

Mostly holds up against AI

AI can take a few chores off your plate, but most of the day is still yours.

Not the odds you keep your job. It's how much of the everyday work still needs a person, out of 100.

← Robot's turf Human's turf →

The ILO's official verdict: Exposed: Gradient 1

12 of the 20 jobs up close

Why this score?

The ILO rated each task in this job for how much of it AI could do. They average 39.1 out of 100, so the job holds up at 100 − 39.1 = 60.9.

ILO group “Exposed: Gradient 1”: The mildest of the ILO’s four “exposed” groups.

AI helps JSA 2025
59
Usually good news: more of the work gets quicker without the job going away.
AI takes over JSA 2025
37
This is what most people mean by “exposed”. We keep it apart from helping on purpose.
Hands-on O*NET 31.0
14.7
Software can't do hands-on work, but robots might one day. How far they've got isn't on this site, because nobody measures it well enough yet.

from the source Working Paper 140, Generative AI and Jobs 2025 index O*NET 31.0 Our Gen AI Transition (Gen AI Capacity Study) September 2025 detailed data release data 0.8-observed-use

Man v Bot · compare

Who holds up better?

Paralegal 61

v

Go on, try your partner's job. Or your sibling's. We won't tell.

Man v Bot

Where the bot helps, and where the human wins

Bot

What AI takes

The routine stuff. You may not miss it.

  • Document review and first-pass discovery.
  • Drafting standard correspondence and bundles.
  • Citation checking and precedent searching.

Man

What people still do better

Judgement, trust and skilled hands. Hard to download.

  • Noticing the document that does not fit the pattern.
  • Managing a client who is frightened and under-informed.
  • Carrying responsibility for an error a court will see.

Both

What's arriving

New work that only exists because the AI does.

  • Verifying machine-drafted work to an evidential standard.
  • Running matters at volumes a firm could not previously staff.
  • Explaining to a client what was automated and what was not.

Written by a human, from the task data. Yes, we see the irony. Read it as one person's take.

Meet your two rivals

They're not the same machine. AI software can only reach the parts of a job done on a screen or in words. A chatbot can't climb a ladder. A robot might, one day, if it doesn't fall off.

The AI

Lives in a screen. Never had to fix a boiler.

Threat level: Medium

37 could do without you

It could take over some of the work, and help with more. It could help with 59 of 100.

The robot

The wildcard. Has hands, still working out stairs.

How much it matters here: Lower

15 how hands-on

Not much of this job needs hands, so the robot can sit this one out.

So could a robot do the physical part?

Less than half of this job is hands-on. Robots still matter, but for a smaller part of the work, so we don't sort it any further.

Robots hate fiddly work in awkward spaces. On that measure, this job scores 12 out of 100. The higher it is, the longer the robot waits.

12 How fiddly
and cramped

Kind of hands-on work Less than half hands-on

The small print

It isn't a robot score, and nothing on this page is. Nobody measures how far robots have got well enough to put a number on it. This is the closest thing we have: how much fiddly handwork and cramped, awkward work the job takes, which is what Frey and Osborne found hardest for machines.

The numbers come straight from the US job database's entry for Paralegals and Legal Assistants. Its wider group, Legal and related associate professionals, lands in the same place.

finger skill 19.7 · hand skill 14.3 · cramped spaces 0.8 · hands-on 14.7 · caring 2.04 of 7 · rule: hands-on under 50

from the source O*NET 31.0 data 0.8-observed-use measures chosen by Frey & Osborne (2017) · the grouping is a sorting, not a measurement

Meanwhile, in real life…

Everything else here is about what AI could do. This is what people actually did. When they asked Claude for help with this kind of work, did they hand the whole task over, or roll their sleeves up and work through it together?

Out of every 100 conversations

63
37

worked through it together handed it over

The small print

Anthropic publishes this about its own product, and it doesn't feed into any score here. These are people who use Claude, not workers in general, so little use here doesn't mean AI can't reach the job. This job makes up 0.2% of the conversations.

These two add up to 100. The bars elsewhere on the page don't, because each is scored on its own, so we don't put them side by side.

observed use Anthropic Economic Index release 2026-06-26, period beginning 2026-05-01 data 0.8-observed-use 2026-05-01 · Anthropic, about its own product

So, should paralegals be worried?

Short answer: not much. Some of the chores will change, and good riddance.

The work is shifting. Drafting and document review rate Medium, and they're a big part of the day, while keeping order in court and investigating rate Very Low. Expect more checking of machine-drafted work, bigger caseloads, and new routes into legal engineering and AI governance.

Worth remembering: we score tasks, not jobs. When AI takes on a task, the job usually reshuffles around it rather than vanishing. Bank tellers survived the cash machine. The job just got more chatty.

Where this job could lead

Fancy a rematch?

Put paralegals up against another job, then send the result to someone who needs a little humbling.

Where this page's numbers come from