Will AI Replace Paralegals?

AI will not replace paralegals, and what changes instead is the task mix inside the role. Paralegal and legal assistant employment is projected to show little or no change from 2025 to 2035 according to the Bureau of Labor Statistics, which projects about 40,800 openings each year over that decade. What AI replaces is a set of tasks, including first-pass document review, contract redlining, file indexing and routine research. The tasks that remain carry judgment, and the lawyer who supervises the output remains responsible for it under ABA Model Rule 5.3.

AI can do first-pass document review, redlining against a playbook and routine research today, and paralegal employment is projected flat rather than falling. Accuracy holds only where the tool is grounded in the firm’s own playbooks and precedents, which is the constraint law firms staffing this work run into first. Supervision and billing stay with the supervising lawyer under the ABA model rules, and the AI skills that keep a paralegal valuable are the reviewable ones.

[cta-1]

Key takeaways

1. Paralegals will not be replaced by AI, and paralegal employment is projected flat rather than falling. Median pay was $62,890 per year in May 2025.
2. AI takes the first pass on repeatable work: document review, redlining against a recorded position, indexing and routine research. Deciding what each flag means stays with a person.
3. ABA Model Rule 5.3 keeps the supervising lawyer responsible for work produced under the lawyer’s authority, and ABA Formal Opinion 512 applies that framework to generative AI.

What paralegal tasks can AI do today, and which ones it cannot

Where the line actually sits

The line is not between simple work and complex work. It sits between work whose output can be checked against a source and work that requires a decision. Everything in the middle column produces something a reviewer can hold against the document in front of them, which is why first-pass document review moved first and intake moved last.

What the third column has in common

Each item in the third column is a decision with professional responsibility attached:

1. Which flag matters for this counterparty and this deal.
2. Which precedent fits the terms actually agreed.
3. What the summary left out that the matter needs.
4. What the client did not say, or did not know to say.

Nothing in the middle column produces those decisions; a tool that did would be making the call rather than preparing it.

What the employment data says about AI and paralegal jobs

AI will not take over paralegal jobs on the current projections: paralegal and legal assistant employment is projected to show little or no change from 2025 to 2035, according to the Bureau of Labor Statistics Occupational Outlook Handbook, which projects about 40,800 openings for paralegals and legal assistants each year over that decade. The same handbook reports median pay of $62,890 per year for the occupation in May 2025.

How to read a flat projection

A flat projection is not a verdict on AI, and it matches the wider shift to AI in the legal sector: the number of roles holds while the tasks inside them change. The projection does not say why the number holds, which tasks sit inside the role at the end of the decade, or how the work is distributed:

1. Why the number holds: a projection estimates headcount; it does not explain it.
2. Which tasks sit inside the role at the end of the decade: an occupation can be the same size and a different job.
3. How the work is distributed: a flat national figure is consistent with growth in one practice area and contraction in another.

Read against the task table above, the projection is better treated as a statement about headcount than as evidence either way about automation.

How accurate is AI on paralegal work?

AI paralegal accuracy depends on what the tool is grounded in. A general-purpose chatbot asked to find authority can fabricate citations, which is why the court in Mata v. Avianca sanctioned counsel over a filing containing invented cases. A tool grounded in the firm’s own playbooks and precedents is answering a narrower question, and its output is checkable against the source document in front of the reviewer.

What verification looks like

That difference sets the verification step. Output drawn from a defined source set can be checked line by line against that set, while output from a general-purpose chatbot has to be verified against the authority itself, which is the accuracy limitation behind the question of whether ChatGPT can replace paralegal work. The check is the one a supervising lawyer already applies to a junior’s draft:

1. Open the source rather than the summary of it.
2. Confirm the authority exists and says what the draft claims.
3. Check that a quoted passage is quoted from the version in force, not a superseded one.
4. Record that the check was done, and by whom.

Existence and accuracy are separate checks, and the second is the easier one to skip. A citation that resolves to a real case still has to say what the draft says it says, and a tool that is confident about a holding it has summarized loosely will pass an existence check and fail a reading. Where a paralegal is working to a filing deadline, the practical control is to verify before the draft circulates internally, not after a partner has read it.

[cta-2]

How lawyers supervise AI work a paralegal used to do

Supervising AI paralegal work falls under the same ethics rules that govern supervising a person. The duty comes from ABA Model Rule 5.3, which requires a lawyer with direct supervisory authority over a nonlawyer to make reasonable efforts to ensure that the person’s conduct is compatible with the professional obligations of the lawyer. The rule also makes the lawyer responsible if the lawyer orders the conduct or, with knowledge of the specific conduct, ratifies it. That framework was applied to generative AI in ABA Formal Opinion 512, issued July 29, 2024, which names Model Rules 1.1, 1.6, 1.4 and 1.5 as the rules that principally govern its use.

What the supervision record should carry

Reasonable efforts is a standard rather than a checklist, so what satisfies it depends on the matter and the tool. A workable record of the supervision carries:

1. What the tool was asked, kept with the matter rather than in a chat history.
2. Which lawyer reviewed the output before it left the firm.
3. When in the workflow the review happened, so that the review sits before the work is filed, sent or relied on.

Keeping that record is what makes the supervision reviewable afterwards, whether the question comes from the client, a court or the firm’s own risk committee. Model Rule 5.3 does not specify any of those steps, and reading it as a fixed procedure would overstate it.

Can a firm bill for AI-assisted paralegal work?

A firm can bill for AI-assisted paralegal work, subject to the reasonableness test in ABA Model Rule 1.5. ABA Formal Opinion 512 gives the worked example: a lawyer who spends 15 minutes entering information into a generative AI tool to draft a pleading may bill that time, along with the time spent reviewing the draft for accuracy and completeness, but in most circumstances may not bill the client for learning how to use the tool.

What the billing change does to leverage

The consequence for leverage follows from the arithmetic rather than from the opinion. A task that took a paralegal a morning and now takes under an hour is billed at the time it takes, so the revenue attached to it is likely to fall even though the work still gets done. The recovered hours go somewhere:

1. Into the work that still needs a person, which is the third column of the task table.
2. Into more matters through the same headcount, so that volume replaces the hours.

Opinion 512 addresses neither choice. It is scoped to the fee question in front of it, so nothing beyond its own example carries over.

Why paralegal AI rollouts get dropped after the first month

Spellbook’s own observation from customer rollouts is that the AI legal assistant platforms paralegals adopt rather than drop share one trait: they sit inside the document the paralegal already has open. A tool that requires exporting a contract, pasting it into a browser, and copying the answer back adds steps to a workflow that was already measured in minutes, and it is abandoned quietly rather than canceled.

Signals that a rollout is failing

The first month is where that shows. A rollout that starts on one repeatable use case, usually first-pass review on a single contract type or the kind of routine document automation a matter already requires, produces a before-and-after a supervisor can see. A rollout that starts with open-ended access produces anecdotes instead.

A supervisor can watch for:

1. Document volume through the tool in the first weeks, which falls before anyone complains.
2. Whether the output is edited or deleted, because a redline that gets deleted rather than amended was not grounded in the firm’s positions.
3. Who asks for access next, which is the adoption signal that is not self-reported.

What AI skills for paralegals are worth building

The AI skills for paralegals worth building are the specific, reviewable ones: writing a prompt that carries the firm’s playbook position, checking an AI-proposed redline against the governing clause, verifying every citation the tool produces, and recording what the tool was asked so the supervising lawyer can follow the chain.

The four skills, and the check on each

Each of those produces something a supervisor can check:

1. Prompting from the playbook: the prompt names the firm’s recorded position on the clause, not just the clause.
2. Checking a proposed redline: the revision is read against the governing clause before it reaches a partner.
3. Verifying citations: every authority the tool returns is opened and confirmed to say what the draft claims.
4. Keeping the record: what was asked and what came back is saved with the matter.

A paralegal who can produce the prompt behind a redline, and the clause the redline was checked against, has produced reviewable work. The skill underneath all four is knowing what the firm’s position on a clause actually is, which is what separates a generic request from the kind of AI prompt a lawyer would write.

What AI cannot replace in paralegal work

What AI cannot replace in paralegal work is the part that depends on reading people and context. A paralegal interviews the client who is withholding the fact that matters, decides what a counterparty’s silence on a clause signals, or knows which partner wants an issue raised in a call rather than a comment bubble.

What each of those judgments costs

Each of the three is a decision that can be made wrongly:

1. The fact the client did not volunteer changes the advice that follows from it.
2. Silence on an indemnity is either an oversight or a position, and treating it as the wrong one costs a round of negotiation.
3. Raising an issue in the wrong channel can cost a week.

Two long-lasting skills sit behind all three: reading a document closely enough to notice what is missing, and writing a sentence a partner can act on without a follow-up call. A tool can surface the clause and the omission. Deciding what either one means in this matter, for this client, is the work that stays.

How Spellbook fits the paralegal work AI can take on

Spellbook is an AI platform for commercial legal work, and the paralegal contract work it covers is the review, redlining and drafting rows of the table above. It works in Word and Google Docs, where that work already happens.

What each feature covers

It is a contract specialist rather than a research or litigation platform, so the coverage tracks the contract rows of the table and stops there:

1. the Review feature reads a third-party contract against the firm’s positions and returns risks and proposed revisions as native Word tracked changes, which is the first-pass review row.
2. the Draft feature produces a first version from a precedent and the deal terms supplied, which is the drafting row.
3. the Compare to Market feature shows how a term compares against 20M+ real contracts, which is what turns “off-market” from an impression into something a paralegal can put in a memo.
4. the Playbooks feature records the firm’s positions, so that the review and the redlines come back in the firm’s own terms rather than generic ones.

A commercial lease example

A lease renewal shows the difference. A paralegal reviewing it has to know whether the rent escalation is standard for that market and asset class. Knowing where the number sits takes a comparison set, and the Compare to Market feature supplies it while the paralegal decides whether to flag the term. The supervising lawyer still signs off on what goes back to the counterparty.

Pricing is custom per seat by team size, with a free seven-day trial, so a firm can test the first-pass review step on its own paper before committing to seats.

Try Spellbook free to see how first-pass review and redlining run on your own contracts in Word.

[cta-3]

Frequently asked questions

Will AI replace legal assistants?

The Bureau of Labor Statistics counts paralegals and legal assistants as a single occupation, so the employment projection above covers both titles. The distinction that matters in practice is not the title, it is the task mix. Where a role is mostly substantive legal work performed under a lawyer’s supervision, the exposure is to the first-pass tasks in the table above. Where it is mostly document production, scheduling and file handling, more of the day is made up of the repeatable work AI does first. Exposure follows the tasks, not the job title.

Will AI replace paralegals in personal injury firms?

No, though the proportion of the day that moves is larger than in a transactional practice. Personal injury work is heavy on intake, medical record review and demand packages, which is high-volume repeatable document handling, so more of the first pass can be automated. What does not move is the part that decides: which records support the claim, what an adjuster’s silence on a figure signals, and how the demand is framed. The task boundary is the one in the table above; the share of the day on each side of it differs by practice.

Does a paralegal need to disclose AI use to a client?

There is no standalone disclosure rule. ABA Formal Opinion 512 names Model Rule 1.4, the communications rule, among the rules that principally govern generative AI use, which makes this a communications question rather than a separate disclosure duty. Model Rule 1.4 asks a lawyer to explain a matter to the extent reasonably necessary for the client to make informed decisions, and that is a reasonableness standard applied matter by matter. The duty also sits with the lawyer rather than the paralegal. Nothing in Opinion 512 turns it into a blanket disclosure rule.