Why settlements still run through a person
Arbitration and mediation run on consent and authority. Two sides pick a neutral, agree on what that neutral may decide, and then have to live with the result. That structure sets the limit on what software can carry on its own.
Look at the actual day. A mediator sits with parties who do not trust each other, holds separate private caucuses, tests what each side will really accept, and walks them toward a deal neither one loves. An arbitrator runs the hearing, rules on objections, weighs sworn testimony against the documents, then writes findings that have to hold up if someone challenges them. The surrounding work is far more routine: notices, scheduling, exhibit indexes, pulling up the governing rules and prior decisions.
The question of AI replacing arbitrators also meets the enforcement problem. An award only matters if it sticks. When it is challenged, someone has to own the reasoning and answer for it. Parties also choose a specific neutral for a reason, and that choice is part of what makes the outcome acceptable.
This is a small field with steady demand. The Bureau of Labor Statistics counted about 9,210 arbitrators, mediators and conciliators in the United States, with median pay of $75,530 and projected growth of 4.7% from 2025 to 2035 (BLS, 2025). Small fields can still see hiring slow when the easy parts of the work shrink, which is the pressure worth watching here.
What AI does, what it helps with, what stays with people
The clearly automatable end of this job is paperwork and preparation. Setting up hearings and settlement conferences, logging filings and tracking deadlines are tasks current tools handle end to end. In our task split, the share of task time in that group is 6%.
The larger slice is assisted work. Researching the applicable statutes, rules and prior awards, and producing a first draft of a written opinion or a settlement memorandum, both move faster with a model in the loop while the neutral checks every citation and every finding. The share of task time in that assisted group is 50%. You can read how we split task time on the coverage method page.
Then there is the work that does not move. Conducting the hearing, interviewing claimants and their representatives, and getting disputing parties to agree to terms are tasks that depend on reading a room, holding confidences and being trusted with a decision. The share of task time our scoring leaves with a person is 44%. Our overall coverage figure for this job, meaning how much task time AI could handle today, is 31 out of 100.
How strong is the evidence
Thin, and we say so plainly. Our quality-parity grade here is D, which means no one has run a clean head-to-head test of an AI system against qualified arbitrators or mediators on real disputes. Because of that, we publish no parity number for this job. A grade is not a guess dressed up as data; it tells you how much to trust the picture.
What would move that grade: a blind study where experienced neutrals and an AI system handle the same set of closed cases, scored by panels on reasoning quality and on whether the award survives challenge; or published outcome data from an institution that runs AI-assisted awards alongside conventional ones. Settlement rates, time to resolution and vacatur rates would all be measurable. Until something like that exists, claims about machine neutrals being as good as people are arguments, not findings. Our scoring method explains how evidence grades feed the headline score.
The score itself is 70 out of 100 (higher is safer), and it reflects both the task mix above and how little has been tested.
When the picture could change
Most likely between 2036 and 2049 (8 in 10 of our scenarios). The replacement-year method sets out how that window is built and what it does and does not claim.
Two things could pull it earlier. First, cost: running a model over filings costs a fraction of paid neutral time, which is a strong pull in small-value, high-volume disputes such as consumer, insurance and online marketplace claims. Second, no hardware is needed. This is desk and hearing-room work, so there is no robot to build, test or certify, and software adoption can move as fast as institutions allow.
Two things hold it back. Enforceability comes first: an award has to withstand challenge, and parties who lose look hard for grounds. Accountability comes second. Confidential caucuses, candid admissions and the judgment to hold one side back from a bad deal all rest on trusting the person in the chair, not the tool behind them. You can line this job up against a neighboring one on our job comparison page.
How to stay needed in dispute resolution
Lean into the tasks that are not drifting. Run the hearing well, including control of objections and a clear record. Do the caucus work: interview each party and their representative so you know the real interests behind the stated positions. And own the closing move, getting disputing parties to agree to terms they will actually honor.
Two skills pay off. One is drafting that survives review, where you use a model for the first pass and verify every authority yourself. The other is subject depth in a specific field such as construction, labor, employment or commercial contracts, because parties pick neutrals who already understand the dispute.
What to do: Build case management and research habits around the assisted tasks, so your billable hours shift toward hearings, caucuses and decisions rather than scheduling and first drafts.
If you want the wider view, the lawyers, judges and related workers family and the law firms sector page show how neighboring roles score. Closest by work are administrative law judges and hearing officers, judges and magistrates and lawyers. For context on which kinds of work hold up best, see the list of jobs that mostly need a person, or search every occupation in the full rankings.