The Work: Who May Hand What to a Machine?

Every handoff already has a rule

Every rule about handing legal work to a machine started as a rule about handing it to a person. Lawyers have long given work to clerks and paralegals, and judges have long taken drafts from law clerks and from the lawyers who appear before them. The law settled who answers for that work well before anyone typed a question into a chatbot, and it settled it differently at the two ends of a case. So the useful question about AI is not what the machine can do. It is where the old line runs at each desk, and why it sits closer to the person at the bench than at the office. The activity puts you on both sides of it: you sort a paralegal’s month of work, then grade an order a chatbot wrote for a judge.

The paralegal’s line: whatever a lawyer will answer for

Start at the office. Rule 4-5.3 of the Rules Regulating The Florida Bar lets a paralegal do the work a lawyer delegates without the lawyer in the room. The cost comes in the same sentence: the lawyer must review that work and is responsible for it. The rule’s comment adds the one limit on what can be handed down. A task that requires the lawyer’s independent judgment and participation cannot be delegated at all. That is the rule on supervision of nonlawyer assistants, and it is drawn around review, not around how hard the task is.

In 2024 The Florida Bar applied the same line to software. Ethics Opinion 24-1 (opens in a new tab) tells a lawyer to review a generative AI’s work the way the lawyer would review a paralegal’s. It also says a lawyer may not delegate to AI anything that could be the practice of law, such as negotiating a claim. A paralegal who uses AI does not step outside the chain of review. The paralegal adds a link to it, and the chain still ends at the lawyer who signs.

Deciding what a machine may touch therefore begins with task decomposition. “Handle the medical records” in an injury case is not one task. Indexing a box of records by date and provider is a job a machine can do, if a person then checks the index against the box. Drafting a first summary of the injuries is a job it can help with, while a person decides what matters. Deciding whether the records support a demand for the policy limits belongs to the lawyer and the client. That is the automate, augment, or human sort, made one task at a time.

The sort reaches the bill too. ABA Formal Opinion 512 gives the plainest example of billing honesty. A lawyer who spends fifteen minutes giving a tool the facts for a pleading may bill those fifteen minutes and the time spent reviewing the draft. The lawyer may not bill the hours the pleading would have taken by hand. A flat fee has to stay reasonable as well, and the opinion warns that charging the same flat fee with the tool as without it may not be. Florida’s opinion forbids inflated time in the same way, and both opinions warn against charging a client while the lawyer learns the tool.

What the numbers say about the paralegal’s side

If machines can take tasks, the next question is whether they take jobs. The evidence so far is early, and it points in more than one direction.

The Bureau of Labor Statistics projects that paralegal jobs will hold flat (opens in a new tab) from 2025 to 2035, moving from 404,900 to 403,800. It names technology, including AI, as the limit on demand, because the tools make paralegals faster at research and at preparing documents. The same page still projects about 40,800 openings a year, most of them left by people who retire or change careers. Florida’s own projections run the other way. FloridaCommerce expects the state’s paralegal jobs to grow about 11 percent from 2026 to 2034, from 38,167 to 42,304. The two forecasts cover different years and rest on different methods. In other words, a forecast of no growth is not a forecast of no jobs, and a national number is not a Florida one.

A team at Stanford looked at payroll records instead of forecasts. In a revision released in August 2026 (opens in a new tab), it found no wave of job losses across the economy. It did find a gap for young workers. Employment of workers aged 22 to 25 in the jobs most exposed to AI, paralegals among them, fell behind. It stood about 19 percent below where it would have been had it kept pace with young workers in less exposed jobs. Experienced workers in the same jobs showed no such gap. The shortfall came mostly from fewer hires rather than more departures, and it was concentrated in jobs where AI substitutes for a worker’s tasks. Where AI complements the worker, employment was flat or rising. The authors call these early signals, not proof that AI caused them.

That is why entry-level work matters more than any one number. The tasks AI tools speed up are the tasks beginners have always learned the job by doing. A firm that stops hiring beginners saves money this year. It may also run short, later, of the experienced people who check the machine’s work.

There is an argument on the other side, and it is an economic one. The “Paralegal Parity” (opens in a new tab) essay weighs a tool against a hire and finds that the usual comparison leaves things out. A paralegal hired to review contracts also takes on the rest of the pile, from confidentiality agreements to supplier paperwork, and the value of that adjacent work comes off the cost of the person. A paralegal also compounds: each week of learning the organization’s people and habits carries into the next task. The essay grants that general-purpose AI has begun to narrow the first gap. It finds no sign yet that a tool compounds.

The judge’s line: help with the writing, never the deciding

Now go to the bench. Judges have always had help with the writing. Law clerks draft opinions, and a Florida judge may ask the lawyers in a case to submit a proposed order. What a judge may not hand over is the decision.

The Florida Supreme Court drew that line in Perlow v. Berg-Perlow (opens in a new tab) in 2004. A trial judge had signed one side’s twenty-five-page proposed judgment without a change, about two hours after receiving it, and without giving the other side a chance to object. The Court sent the case back for a new trial. A party’s proposal can be a starting point, it wrote, but it cannot substitute for the judge’s own independent analysis of the facts, the issues, and the law.

Twenty-two years later a Florida appeals court showed what that analysis looks like when it is missing. In Bell v. Konidaris, decided in September 2026, a trial judge’s order on attorney’s fees tracked the parties’ written closing arguments so closely that it kept both sides’ citation styles. One side had underlined its citations and the other had italicized them, and the order printed both styles side by side in the same paragraphs. The Fourth District reversed. It listed the signs a reviewing court looks for, among them whether the order matches what the judge said in court, how long the judge took, and whether the judge edited the draft at all. An order full of errors and conflicts, it added, looks like the drafter’s work rather than the court’s.

Put a chatbot in the drafter’s chair and every one of those signs still works. So does the worst failure. In Shahid v. Esaam, 918 S.E.2d 198 (Ga. Ct. App. 2025), a Georgia trial court denied a woman’s petition to reopen her divorce in an order that relied on two cases that do not exist. The order itself indicated that the husband’s lawyer had prepared it. The appeals court made no finding about who, or what, put the fake cases there. It vacated the order anyway, because a defect that plain on the order’s face left nothing it could meaningfully review.

The courts’ own AI policies turn these cases into rules. The federal judiciary sent interim guidance to every federal court in July 2025 and has not published it. It cautions against delegating core judicial functions to AI, including deciding cases, and asks every user to verify AI output independently. New York’s interim policy (opens in a new tab) lets judges use approved tools for help, but it says AI must never do the decision-making a judge is ethically bound to do. It also tells judges and staff to assume that anything typed into a public AI tool becomes public. Party names, docket numbers, and dates of birth stay out of any tool the court does not control, and so does any document filed in a case, even a public one. In Florida the rule for judges is still Perlow’s, and it reaches a chatbot’s draft as surely as a lawyer’s.

Why the two lines sit in different places

Set the two lines side by side and the difference is easy to state. At the office, the line runs around review. Almost anything can be handed down, because a lawyer above the work reads it, signs it, and answers for it. At the bench, the line runs around the decision itself. The parties are owed the judge’s own judgment, and no one in chambers stands above the judge to catch a borrowed one. An appeals court can catch it later, as the courts in Bell and Shahid did, but only by sending the case back to be decided again.

That is the work in front of you. The activity starts at the office, with ten tasks from one paralegal’s month on a Volusia County lawsuit, and asks you to put each one on the right side of the paralegal’s line. Then it moves you into chambers on the same case, where a chatbot drafts an order and you decide how much of it a judge could sign. Begin with the three numbers in Part 1, since the note you write at the end has to answer to them.