Law, Medicine & Finance

If you are a licensed lawyer, a practicing physician, a CPA, a CFA charterholder, or any credentialed professional in law, medicine, or finance, you can currently earn between $60 and $250 per hour working part time from your laptop, teaching frontier AI models how your field actually reasons. This is not data entry and it is not survey filler. AI labs like OpenAI and Anthropic have discovered that the last mile of model quality (getting a legal memo, a differential diagnosis, or a discounted cash flow model correct) requires feedback from people who have actually done that work, and they are paying real professional rates to get it. The company Mercor alone has assembled a network of more than 30,000 doctors, lawyers, bankers, and other white-collar professionals to do exactly this, per TIME and Metaintro.

This guide is written for the person actually deciding whether to do this. It covers what the work is day to day, the specific platforms hiring right now and what each really pays (broken out by platform, seniority, and country), the credentials you need, how to pass the assessments, realistic monthly earnings, the tax and payment logistics, the scams to avoid, and an honest look at the downsides (including the November 2025 controversy where Mercor cut some contractors' pay by roughly a quarter). Where a number appears, it is sourced. Where the picture is messy, this guide says so.

What "domain expert" AI training actually is

The plainest way to describe the job: an AI lab is building a model that gives legal, medical, or financial answers, and it needs a real expert to judge whether those answers are correct, safe, and reasoned the way a professional would reason. You are the human standard the model is measured and trained against.

In practice the work falls into a handful of recurring task types, which line up with how reinforcement learning from human feedback (RLHF) pipelines work:

• Output evaluation and ranking. The model produces two or more answers to a realistic prompt (say, "draft a demand letter for breach of a commercial lease" or "interpret these labs for a 58-year-old with new-onset dyspnea"). You rate each one, or rank them against each other, on dimensions like correctness, completeness, safety, and reasoning quality. This is the bread-and-butter A/B comparison that RLHF is built on.
• Rubric and gold-standard creation. You write the "right answer" (the gold-standard response a domain expert would give) so the model has something authoritative to learn from and be graded against. This is the highest-value, highest-paid work because it captures how a professional actually thinks, not just a thumbs-up or thumbs-down.
• Prompt and scenario writing. You invent hard, realistic problems the model should be able to handle, including deliberately ambiguous edge cases (a fact pattern with a hidden statute-of-limitations issue, a financial model with a subtle circular reference) that separate a competent model from a superficial one.
• Error spotting and critique. You read a model's long-form output and flag exactly where it went wrong: an invented case citation, a misread of GAAP revenue recognition, a contraindicated drug interaction, an unsupported legal assumption. Then you explain why, in the language a supervising professional would use.
• Benchmark task authoring. The most prestigious version of this: building the evaluation set that decides whether a frontier model is "good enough" at your profession. Mercor reportedly spent more than $500,000 building a benchmark called APEX with 200 tasks that test whether AI "can perform knowledge work with high economic value" across law, medicine, finance, and consulting, per TIME's reporting.

The common thread is judgment. A generalist can label a stop sign in a photo. Only a securities lawyer can reliably tell whether a model's answer about a Regulation D exemption is subtly, dangerously wrong. That gap is precisely what you are paid for, and it is why domain expert rates sit an order of magnitude above generalist annotation. As one platform framing puts it, the gap between a $15/hr generalist and a $200+/hr specialist "is entirely your existing knowledge," per RemoWork.

Why the labs pay professional rates for it

Two forces set the price. First, the supply of people who can do this is small: you need a real bar admission, medical license, or finance credential plus years of practice, and most such people already have well-paying day jobs. Second, the demand is enormous and growing. The AI data labeling market is roughly $2.32 billion in 2026 and is forecast to reach $6.53 billion by 2031, a 22.95% CAGR, per Mordor Intelligence. The broader human-in-the-loop AI market grows from $5.4 billion in 2025 to $6.73 billion in 2026 at a 24.7% CAGR, per Research and Markets. The money behind the buyers is staggering: Scale AI was valued above $29 billion after Meta's 2025 investment, Surge AI reportedly cleared $1 billion in revenue in 2024 while bootstrapped, and Mercor raised $350 million at a $10 billion valuation in 2025, per that same market roundup. When those companies compete for a scarce pool of credentialed experts, the price of an hour of your judgment goes up.

How much you can really make: pay by platform

Pay in this field is genuinely wide, and averages are misleading because they blend $12/hr generalist tasks with $200/hr specialist tasks. What matters for a credentialed professional is the specialist band on each platform. Here is what the major platforms hiring for law, medicine, and finance actually pay in 2026.

A few honest caveats about this table. The high numbers (Mercor's $200+, Braintrust's $300, AfterQuery's $250) are real advertised rates but they are the top of the band, reserved for the most senior specialists on the most demanding projects. The realistic middle for a solidly credentialed professional (a licensed attorney with a few years of practice, a board-certified physician, a CFA) is closer to $80-$150/hr on the premium platforms. And "advertised rate" is not the same as "effective rate": on task-based platforms like Outlier, unpaid time spent reviewing instructions, waiting for work, and redoing rejected tasks drags the real hourly figure down. A 2025 AlgorithmWatch study cited by reviewers found roughly a third of AI gig workers' time goes unpaid.

Pay by seniority

Within any single platform, your rate is a function of credentials and experience. Mercor's published tiering is the clearest public example and generalizes well across the industry, per Mercor's own salary breakdown:

The jump from mid to expert is the single most important line for a professional reader: it is where your license or advanced degree roughly triples your rate. And within the expert tier, sub-specialty matters enormously. Mercor has posted UK labor and employment lawyer roles at $225-$250/hr, transactional/corporate legal experts at $100-$150/hr, and IP/patent experts at $100-$150/hr, per its lawyers page. The rarer and more commercially valuable your niche, the higher the number.

Pay by geography

This work is remote and global, but pay is emphatically not flat across countries. US-based experts command the top of every band. The same task can pay a US attorney several times what it pays a contributor in a lower-cost market, and some entry projects in the Philippines have been advertised at $2.00-$2.50/hr for L1 work, while specialist AI training roles there range $15-$75/hr, per RemoWork.

Two practical points. First, several of the best-paying entry platforms are geo-restricted: DataAnnotation accepts workers only from the US, Canada, UK, Ireland, Australia, and New Zealand, per multiple reviews. Second, jurisdiction-specific expertise pays a premium because it is scarce in that jurisdiction: a US-barred securities attorney or a UK-qualified employment solicitor is exactly what a lab building a jurisdiction-aware legal model needs, and they will pay for the passport-plus-credential combination.

The credentials and skills you actually need

For the premium law, medicine, and finance work, the gate is a real, verifiable professional credential. Platforms verify these before assigning specialist projects. What qualifies:

• Law: an active bar admission and/or a JD, with professional experience on complex matters. Litigation, transactional, corporate, compliance, regulatory, IP, real estate, tax, and bankruptcy backgrounds are all in demand, and judicial clerkship experience is valued, per Mercor's lawyers page. An LL.M. helps for the highest tiers on some platforms like Braintrust.
• Medicine: a medical license (MD/DO), plus board certification and clinical experience. Nurses (RN, NP), pharmacists (PharmD), and other licensed clinicians also qualify for many medical-reasoning projects.
• Finance: a CPA, CFA, MBA in finance, or substantial professional experience in investment banking, equity research, accounting, financial modeling, or corporate finance. The work leans on your ability to reason through valuations, statements, and regulatory frameworks the way a practitioner would.
• Adjacent credentials: a PhD in a relevant field, or 5-10+ years of documented senior experience, can substitute for a formal license on many platforms per Mercor's tiering.

Beyond the credential, the platforms are screening for a specific skill that many excellent professionals lack: the ability to articulate your reasoning in clear, structured writing. Being right is not enough; you have to explain why a given answer is right or wrong in a way a model (and the lab's researchers) can learn from. Strong written English, precision, and comfort with detailed rubrics and instructions are effectively mandatory. If you can write a clean, well-reasoned memo or a tight clinical assessment, you already have the core skill.

Tools you will actually use

The toolset is deliberately lightweight. Most work happens in a browser-based platform interface where prompts, model responses, and rating rubrics are presented side by side. You will encounter comparison and ranking UIs (rate A vs B), long-form text editors for writing gold-standard answers, and structured forms for tagging errors. Some coding-adjacent projects use notebooks or IDEs, but pure law/medicine/finance evaluation is mostly reading and writing. Time-tracking software (Hubstaff, for instance, on Invisible) appears on some hourly contracts. You do not need to know how to code or how machine learning works; you need to know your profession and be able to type.

How to get started and pass the assessments

The path is consistent across platforms, and understanding it removes most of the friction.

1. Pick the right platforms for your credential and country. For a US/UK/CA/AU/NZ/IE professional, apply to Mercor, AfterQuery, Braintrust, and DataAnnotation as your premium tier, with Outlier and Surge as volume backups. Apply to several: work availability is lumpy, and having multiple approved platforms smooths out the dry spells.
2. Build a credential-forward profile. These platforms hire on verifiable expertise. Lead with your license, your firm or institution, your specialty, and concrete matters or cases you have worked. Mercor uses an AI-driven application and interview process; micro1 uses an AI interviewer called "Zara"; the profile and interview are where you prove your tier, per micro1.
3. Complete the qualification/assessment task. Almost every platform gates paid work behind an unpaid or nominally paid sample task in your domain. For a lawyer this might be evaluating two AI-drafted contract clauses and writing which is better and why. This is the single most important step: it sets your rate and your access to projects. Treat it like a bar exam answer. Follow the rubric to the letter, show your reasoning explicitly, and do not rush.
4. Calibrate to the rubric, not to your instinct. The most common reason strong professionals get rejected or downgraded is that they answer the way they would in practice instead of the way the rubric asks. Read the instructions twice. If the rubric says to penalize unsupported claims, penalize every one, even ones a real client would forgive.
5. Start small, build a quality record, then negotiate. Your first projects establish a reviewer-quality score. High agreement with gold standards and clean writing move you into higher-paying projects and unlock rate increases. On some platforms you can request a higher rate once you have a track record.

A practical note on the assessment: expect it to take one to three hours and to feel harder than the work itself. That is by design. The platforms would rather over-screen than ship low-quality expert data to a lab paying millions for it.

Realistic earnings: what a month actually looks like

The honest answer is that this is excellent part-time income and an unreliable full-time income. Work is project-based and comes in waves. A credentialed professional working 10-15 focused hours a week on premium platforms at, say, $100/hr is looking at roughly $4,000-$6,500 a month when projects are flowing. Push to 25-30 hours a week during a busy project and the top end climbs well past that. Mercor advertises senior domain expert earnings equivalent to about $400,000 a year at full-time volume, per Forbes, though sustaining full-time hours at that rate depends entirely on continuous project availability, which no platform guarantees.

Set expectations accordingly. Most people in this work treat it as a high-value side income stacked on top of a primary job or practice: a physician doing 8 hours on weekends, an attorney billing evaluation work between matters. The people who try to make it a sole income source run into the structural problem of the field: you are a contractor on someone else's project pipeline, and pipelines dry up. Diversify across platforms, and never assume this month's project volume repeats next month.

Payment and tax logistics

You will almost always be classified as an independent contractor (1099), not an employee. That has concrete consequences you should plan for from day one.

• How you get paid. Most platforms pay weekly or per-project via PayPal, Stripe, or direct deposit, after your work passes review. Payment timing is generally reliable on the established platforms, though task-based platforms only pay for approved work, so rejected tasks are unpaid.
• You owe self-employment tax. As a 1099 contractor you pay both halves of Social Security and Medicare: 15.3% total (12.4% Social Security + 2.9% Medicare) on your net earnings, on top of ordinary income tax, per the IRS. No employer withholds anything for you.
• Quarterly estimated taxes. Because nothing is withheld, you are expected to pay estimated taxes quarterly using Form 1040-ES to avoid an underpayment penalty, per the IRS gig-work guidance. Set aside roughly 25-35% of what you earn.
• The forms you will receive. A platform that pays you directly issues a 1099-NEC; for 2026 that reporting threshold rose to $2,000. If you are paid through a third-party processor like PayPal, the payment is reported on 1099-K, and after the July 2025 "One Big Beautiful Bill" the 1099-K threshold reverted to more than $20,000 and 200 transactions, per tax reporting guidance. Critically: you must report all income even if you never receive a form. No 1099 does not mean no tax.
• Deductions. You report income on Schedule C and can deduct genuine business expenses (a portion of internet, a home office, hardware). Keep records.

If you hold a professional license, one more consideration: check your employer's moonlighting policy and your professional conduct rules. Evaluating anonymized AI outputs is generally fine, but a physician or attorney should never input real patient or client confidential information into these platforms, and some employment contracts restrict outside work.

Red flags and scams to avoid

The legitimate version of this work is real and well-paid, which is exactly why scammers imitate it. The single clearest rule, per multiple 2026 fraud analyses: money only ever flows from the company to you. A real AI training platform never charges you.

Treat any of the following as a hard stop:

• Any upfront fee. Registration fee, training fee, background-check fee, "starter kit," equipment deposit: all scams. Legitimate platforms (Mercor, DataAnnotation, Braintrust, etc.) are free to join.
• A check you must deposit and partially forward. The classic overpayment scam. The check bounces and you owe the bank.
• Requests for banking passwords, or a photo of your ID before you have done any work or spoken to a real person.
• Off-platform recruiting to WhatsApp or Telegram, especially with pressure to "act today."
• Mismatched domains. A recruiter emailing from @google-careers.com or @microsoft-hr.net is not from Google or Microsoft.
• Deepfake or glitchy "video interviews" and pressure tactics ("the role closes at 5 p.m.").

The safe pattern looks like this: you go directly to a known platform's website, apply, complete a qualification task, and start receiving paid tasks after review, with no fee at any point. Stick to the established names and you avoid essentially all of this.

The honest downsides

This is a genuinely good opportunity, but it is not frictionless, and a serious reader deserves the full picture.

Pay can be cut with little notice. In November 2025, Forbes reported that Mercor cut some contractors' pay from $21/hr to $16/hr, a roughly 24% reduction, after abruptly cancelling a months-long project ("Project Musen") that had affected over 5,000 contractors, locking them out of Slack and offering the lower rate hours after termination, per Forbes. One worker called it "a slap in the face." $16/hr is below minimum wage in states like California and Washington. That episode is a useful reminder: you are a contractor, the labs and platforms control the pipeline, and rates and projects can change unilaterally. (Note this affected lower-tier annotation work, not the top specialist bands, but it shows the structural risk.)

Effective pay is lower than advertised on task platforms. As noted, unpaid instruction-reading, waiting, and rework can consume roughly a third of your time on platforms like Outlier, dragging a $40/hr advertised rate toward the high $20s in reality, per Breaking Even and the AlgorithmWatch findings cited there.

Work is inconsistent. Projects start and stop. A great month can be followed by a dead one. This is why diversifying across platforms is not optional.

The work can be repetitive and mentally taxing. Reading and grading dozens of near-identical model outputs against a strict rubric is more grinding than practicing your profession, and satisfaction data for some platforms has declined as rates and access tightened, per Outlier reviews.

There is a quiet irony. You are, in part, training the models that will automate portions of the very work you do. The current window exists because AI is not yet good enough at law, medicine, and finance to do without expert oversight. That is precisely why the pay is high now.

The 2026 outlook

The near-term trajectory strongly favors people with real professional credentials. The whole industry is shifting from cheap, high-volume generalist labeling toward scarce, high-value expert feedback, because the frontier labs have largely saturated the easy data and now compete on quality in exactly the domains that require licensed judgment. The market numbers back this up: the AI data labeling market grows at nearly 23% a year through 2031 per Mordor Intelligence, and the human-in-the-loop segment at 24.7% per Research and Markets, with billions in fresh capital flowing to the platforms that broker this work.

For a credentialed law, medicine, or finance professional in 2026, the realistic read is: this is one of the best-paying flexible remote opportunities available to you right now, it is legitimate at the established platforms, and the demand curve is pointing up. The rate you can command is directly proportional to how scarce and commercially valuable your specific expertise is. The smart move is to get onto two or three premium platforms, ace the qualification tasks, build a quality record, and treat the income as high-value but variable. Do that, keep your confidential professional data out of it, set aside money for self-employment tax, and you have turned your license into a $100+/hr side income for as long as the models still need you to teach them.

Quick-start checklist

• Confirm you hold a verifiable credential (bar admission, medical license, CPA/CFA, or PhD/senior experience).
• Apply to a premium platform matched to your country: Mercor, AfterQuery, and Braintrust for most; add DataAnnotation if you are in the US/CA/UK/IE/AU/NZ.
• Add Outlier and Surge as volume backups.
• Treat the qualification task like an exam: follow the rubric exactly, show your reasoning, do not rush.
• Never pay a fee, never move to Telegram/WhatsApp, never share banking passwords.
• Set aside 25-35% for taxes; plan on quarterly estimated payments.
• Diversify across platforms and expect project volume to fluctuate.