Will AI Replace Paralegals?

What the employment figures say

The US Bureau of Labor Statistics counts 376,200 people in the category it calls Paralegals and Legal Assistants (SOC 23-2011) in 2024, and projects 0% employment change for 2024-2034 alongside roughly 39,300 openings a year, mostly from workers transferring out or leaving the labour force ^{(U.S. Bureau of Labor Statistics 2025)}. Flat headcount sits oddly beside a Critically Exposed task score, and the gap is real rather than an error: BLS states that paralegals and legal assistants are expected to see the strongest employment impact among legal occupations from generative-AI productivity gains, while also stating that lawyers must still review the outputs ^{(U.S. Bureau of Labor Statistics 2025)}. The category is also narrower than the exposure driver — legal secretaries, the role named in the fastest-declining evidence, are classified separately, so this series is not a proxy for legal support staff as a whole ^{(U.S. Bureau of Labor Statistics 2025)}. BLS does not attribute the whole difference between its earlier +1.2% outlook for 2023-2033 and the 0% figure for 2024-2034 to AI, because the periods and the products differ ^{(U.S. Bureau of Labor Statistics 2025)}.

What the research says about paralegals

Now

The exposure sits in the middle of the job rather than at its edges. Employers name legal secretaries among the fastest-declining roles, and document review and legal research are highly automatable ^{(World Economic Forum 2025)}. Everlaw, which sells the review software and chose the framing of its own case study, reports that a three-attorney team at an unnamed Am Law 100 firm processed more than 126,000 documents in under 24 hours after validation, with 50-67% less review time and one-quarter of the personnel estimated for conventional managed review, at reported accuracy above 90% ^{(Everlaw 2025)}. Allen & Overy's launch announcement with Harvey — a law firm and a software vendor with a shared interest in the story — described more than 3,500 lawyers across 43 offices asking about 40,000 questions covering contract analysis, due diligence, litigation and regulatory compliance, with outputs requiring careful lawyer review ^{(Allen & Overy 2023)}.

What none of those deployments disclose is a paralegal headcount effect. Financial Times reporting on in-house teams, where corporate interviewees have an interest in showing adoption working, describes Unilever legal delivery centres in Barcelona, Mexico City and Bengaluru with 85 staff working on contracts and tools saving lawyers an average of 30 minutes per day, with planned role cuts in a European office expected to have minimal impact on the legal team ^{(Financial Times 2024)}. ASML's ContractMatrix rollout in early 2024 and Repsol's Lexia system, used alongside Harvey for more than 12,000 prompts in under six months across a team of 160 lawyers, were reported without any disclosed productivity gain or headcount reduction ^{(Financial Times 2024)}. Pay has not moved either: the median annual wage was $60,970 in May 2023 against $61,010 in May 2024, a nominal difference of $40 ^{(Allen & Overy 2023)}.

The next few years

The relevant question is shifting. Earlier exposure indices asked whether a model could speed up a task with a person prompting and checking each step; agentic systems now plan, use tools, spawn subagents and run multi-step work unattended on a schedule, which turns the question into whether the work can be completed without you ^{(Nous Research 2026) (Anthropic 2026)}. Routine reporting and data enrichment are already delegating into API workflows, while interpretation, framing and ownership of the decision lag behind ^{(Anthropic Economic Index 2026) (Microsoft Research 2025)}. Taken together, the occupation-specific record supports high task-level exposure across review, coding, retrieval, summarization, standard drafting and matter administration, alongside flat official employment, substantial replacement demand, growing complex responsibilities and no disclosed paralegal layoffs in any cited deployment ^{(U.S. Bureau of Labor Statistics 2025)}.

Longer term

Over the ten-year view, the official series expects the number of paralegal and legal-assistant jobs to hold rather than fall, at 0% change for 2024-2034 with about 39,300 openings a year arising mostly from turnover ^{(U.S. Bureau of Labor Statistics 2025)}. That is a downgrade from the previous projection round, which had employment rising from 366.2 thousand to 370.5 thousand, a gain of 4.3 thousand jobs or +1.2% for 2023-2033 ^{(U.S. Bureau of Labor Statistics 2025)}. The two figures come from different periods and different products, so the shift should not be read as a measured AI-caused decline — but it is the direction of travel the statistical authority now expects ^{(U.S. Bureau of Labor Statistics 2025)}.

Because the confidence here is high and the driver is task content rather than a single forecast, the practical reading is that the work changes shape faster than the headcount does.

‎

What is exposed, and what protects it

● Where the exposure is

The exposed core is the document layer: first-pass review, discovery coding, retrieval, summarization and standard drafting, which is exactly what legal AI is built to do ^{(World Economic Forum 2025)}. A vendor case study puts the scale of that compression at more than 126,000 documents in under 24 hours with 50-67% less review time and one-quarter of the estimated personnel, though the vendor sells the software and the firm is unnamed ^{(Everlaw 2025)}. Large-firm adoption is broad rather than experimental, spanning contract analysis, due diligence, litigation and regulatory compliance ^{(Allen & Overy 2023)}. In-house legal teams are running the same playbook on contract drafting and review, with the reported time savings measured on lawyers rather than on support staff ^{(Financial Times 2024)}.

○ What holds

The protection here is regulatory and accountability-shaped rather than physical: a licensed lawyer still has to review and stand behind the output, which keeps a human in the loop by design ^{(U.S. Bureau of Labor Statistics 2025)}. Verification is not a formality — Stanford researchers benchmarking Lexis+ AI, Westlaw AI-Assisted Research, Ask Practical Law AI and GPT-4 on more than 200 open-ended legal queries reported incorrect answers above 17% for two tools and above 34% for another, concluding that every proposition and citation must be checked ^{(Stanford Institute for Human-Centered Artificial Intelligence 2024)}. Practitioners quoted by Bloomberg Law, a publisher with its own commercial stake in legal technology, describe AI as useful for first drafts and memo or contract analysis while client conversations, discovery walkthroughs, meetings and people skills stay with them ^{(Bloomberg Law 2026)}. The trade body NALA, which sells certification and promotes the report as a member benefit, states that 20% more paralegals reported handling complex responsibilities than in 2020, with certification, education and experience associated with higher compensation ^{(National Association of Legal Assistants 2024)}.

Which tasks go first

What AI already covers is the volume end: first-pass document review and coding, retrieval, summarization, template-based drafting and routine matter administration ^{(U.S. Bureau of Labor Statistics 2025)}. In a paralegal class reported by Bloomberg Law, students using ChatGPT finished a drafting exercise about 15 minutes faster than students working from a traditional template, though repeated prompting was needed for completeness and accuracy — a classroom exercise, not a labour-market study ^{(Bloomberg Law 2026)}. The agentic shift moves the next layer: routine reporting and data enrichment run unattended inside API workflows, while interpretation and decision ownership stay behind ^{(Anthropic Economic Index 2026) (Microsoft Research 2025)}. What stays human is the part with a signature and a person attached — client conversations, discovery walkthroughs, meetings and the judgement calls a supervising lawyer must own ^{(Bloomberg Law 2026)}. Checking the machine is itself becoming a task, given reported incorrect-answer rates above 17% on benchmarked legal research tools ^{(Stanford Institute for Human-Centered Artificial Intelligence 2024)}.

What to do about it

Move toward the work that carries accountability

Client contact, discovery walkthroughs, witness and meeting preparation, and the coordination a supervising lawyer signs off on are the parts practitioners describe as still theirs ^{(Bloomberg Law 2026)}. Paralegals reporting complex responsibilities have grown as a share of the profession, which suggests the ladder upward exists inside the role rather than only outside it ^{(National Association of Legal Assistants 2024)}. Documenting the matters where your judgement, not your throughput, changed the outcome is how that shift becomes visible at review time.

Become the verification layer, deliberately

Benchmarked legal research tools returned incorrect answers above 17% of the time on more than 200 queries, and above 34% for one of them, with the conclusion that every proposition and citation needs checking ^{(Stanford Institute for Human-Centered Artificial Intelligence 2024)}. Vendor deployments themselves state that outputs require careful lawyer review ^{(Allen & Overy 2023)}. A paralegal who can design and run that check — sampling, citation validation, escalation rules — is holding the part of the workflow the tools cannot close.

Bank credentials and transferable technical skill

NALA, which has a commercial interest in saying so, reports that certification, education and experience are associated with higher compensation ^{(National Association of Legal Assistants 2024)}. Because the official projection is 0% change with about 39,300 openings a year driven by turnover, mobility within and out of the occupation is normal rather than a crisis signal ^{(U.S. Bureau of Labor Statistics 2025)}. Skills in structured data work and AI tooling travel well into compliance, contract operations and litigation support if you decide to move.

Tools and courses that fit this situation

Below are paid tools and courses that fit the situation described here: verification-heavy research work, document workflows, and the option of moving sideways into more technical legal support. None of them removes the exposure; they address specific parts of it.

Some links above are affiliate links. If you buy through them, Jobs After AI may earn a commission at no extra cost to you. It does not change which tools appear here.

Common questions

Is this about paralegals losing jobs, or about the tasks changing?

It is about tasks, first and foremost. The score measures how much of the occupation's task content current AI can do, and for paralegals that content — review, retrieval, summarization, standard drafting — is heavily exposed ^{(World Economic Forum 2025)}. Official employment is projected at 0% change for 2024-2034, and no cited deployment discloses a paralegal layoff ^{(U.S. Bureau of Labor Statistics 2025)}.

What is AI still bad at in this job?

Legal accuracy without supervision. Benchmarking of Lexis+ AI, Westlaw AI-Assisted Research, Ask Practical Law AI and GPT-4 across more than 200 open-ended queries found incorrect answers above 17% for two tools and above 34% for one, with the finding that every proposition and citation must be verified ^{(Stanford Institute for Human-Centered Artificial Intelligence 2024)}. Client conversations, discovery walkthroughs and meetings also remain with people ^{(Bloomberg Law 2026)}.

Are junior paralegals affected first?

The entry-level work is where the compression is reported. A vendor case study describes conventional first-level review being done with one-quarter of the personnel estimated for the job, at accuracy the vendor reports as at or above first-level attorney benchmarks — the vendor sells the software and the firm is unnamed ^{(Everlaw 2025)}. Meanwhile the share of paralegals reporting complex responsibilities has risen relative to 2020, which is the other half of the same story ^{(National Association of Legal Assistants 2024)}.

How fast is this actually happening?

Adoption is already broad in large firms and in-house teams, covering contract analysis, due diligence, litigation and compliance ^{(Allen & Overy 2023)}. What has not moved is pay: the median annual wage was $60,970 in May 2023 against $61,010 in May 2024 ^{(Allen & Overy 2023)}. Institutional pace, not capability, is what sets the speed here.

How is the 30/100 score calculated?

It combines the employer-reported decline of legal secretary roles with the automatability of document review, discovery, drafting and research ^{(World Economic Forum 2025)}. It is a task-exposure measure over roughly a three-to-five-year view, not a forecast of headcount, which is why it sits beside a 0% official employment projection without contradicting it ^{(U.S. Bureau of Labor Statistics 2025)}.

Related occupations

The evidence behind this page

How this score was built

The AI-Proof Score is an estimate of how exposed an occupation's core tasks are to AI over roughly the next three to five years, on a 0–100 scale where higher is safer. It measures the exposure of tasks, which sits upstream of, and is softer than, job loss — it is not a probability that this job disappears, and not a claim about any individual doing it. The full derivation, the four research families behind it and its limitations are set out in the methodology, and every occupation is ranked in the AI Job Impact Report 2026.