AI Training and Data Annotation Recruitment in the US

AI labs pay licensed professionals $75 to $500 an hour to write, review and score model output in their own field. Acceler8 Talent sources and contracts those experts across the US, from engineering and research to clinical and finance. We build the bench, handle the vetting and keep it stocked.

Key takeaways

• US domain experts in AI training average $108 an hour across 95 live specialist listings, rising to $500 for STEM professors and researchers (Handshake AI listing data, October 2026).
• The US data annotation tools market stood at $242.5m in 2023 and is forecast to reach $1,025.2m by 2030, a 22.9% compound annual rate (Grand View Research).
• Mercor reached roughly $2.0bn annualised gross revenue by June 2026, up from about $500m annualised in October 2025 (Sacra; TechCrunch).
• Contract durations across the category run from 3.0 weeks for business projects to 23.5 weeks for language work, so rebenching is continuous rather than occasional (aitrainer.work listings census, September 2026).
• Only 36.0% of Mercor listings disclose weekly hours at all, and 3.3% state a floor of 35 hours or more, which is why buyers struggle to forecast capacity from platform listings alone.

Knowledge workers in AI training produce and grade the expert-grade material a model learns reasoning from. A radiologist writes and reviews diagnostic reasoning. A tax lawyer scores whether a model’s advice would survive a filing. The output is judgement, recorded in a form a training run can use.

That work replaced a different job. Until reasoning models arrived, the bulk of human data work was labelling: drawing boxes, tagging images, transcribing audio. Manual annotation still accounted for 74.4% of the US annotation tools market by revenue in 2023 (Grand View Research), but the money in new engagements has moved up the credential ladder.

What is a knowledge worker in AI training?

A credentialed professional contracted to write, review and score model outputs inside their own licensed or specialist field. Typical profiles are physicians, lawyers, engineers, quantitative analysts and research academics. US rates run from $75 to $500 an hour depending on specialism, and most engagements are part-time and project-based rather than full-time employment.

The term is borrowed from management theory and repurposed. In this market it means something narrow: a person whose professional record is the reason the lab wants them, not a person who can follow an annotation rubric. The distinction decides the rate. Generalist AI trainers average $40.50 an hour against $108 for credentialed specialists, a 2.67 times difference calculated from ZipRecruiter data dated 18 September 2026 and Handshake AI listings from October 2026.

How does expert data differ from data annotation?

Annotation records a fact about data. Expert data records a professional judgement about a problem. An annotator confirms an image contains a tumour; a radiologist explains the differential, ranks the evidence and says what a registrar would get wrong.

The cost structure follows from who can do each job. Annotation can be performed by anyone trained on a rubric, which pushes the rate to the lowest available labour market. Expert work requires a licence, a doctorate or a senior industry record, and that supply constraint is what holds US rates at the levels above. Generalist annotation sits under $25 an hour in most US listings, while DataAnnotation posts $75 to $150 and above for software engineers and data scientists, and $60 to $125 and above for physicians (platform job board, September 2026).

Why are AI labs hiring licensed professionals?

Because the companies holding the data will not sell it. Labs want to automate professional work, and the firms whose value sits in that work have no incentive to hand over the material that would let them.

Mercor chief executive Brendan Foody set out the problem at TechCrunch Disrupt in October 2025, telling TechCrunch that “there is an argument that Goldman Sachs doesn’t love the idea of having models that are able to automate their value chain” and that “their customers don’t want to give them data to automate large portions of their value chains”. Hiring the former senior employee is the route around a data contract that will never be signed. That single commercial fact is what created this labour market, and it is why the demand is for people rather than datasets.

Three vendors account for most of the spend, and their combined revenue grew faster than almost any labour market on record. Mercor went from roughly $500m annualised revenue in October 2025 to about $2.0bn annualised gross by June 2026 (TechCrunch; Sacra). Surge AI reported a $1.4bn run rate in late 2025, and Scale AI recorded $870m in 2024 revenue (Value Add VC).

The tooling market is smaller and better measured. Grand View Research puts the US data annotation tools market at $242.5m in 2023, forecast to $1,025.2m by 2030 at a 22.9% compound annual rate. Tooling is not labour, so that figure describes the platforms rather than the people, and the vendor revenue above is the better proxy for what labs spend on humans.

How big is the US AI training market?

Vendor revenue is the only reliable read, and it reached multiple billions during 2026. Mercor paid out more than $2m a day to contractors in 2026, up from more than $1.5m a day in October 2025 (Sacra; TechCrunch). Annualised, the 2026 figure represents over $730m flowing to individual contractors from one platform.

Scale AI guided to about $1bn for 2026, down from roughly $2bn in 2025 after Meta took a 49% non-voting stake for $14.3bn in June 2025 and several labs reduced their contracts (Value Add VC; CNBC). Revenue did not leave the category when it left Scale, it moved to Mercor and Surge AI. That rotation is useful to a buyer, because it shows the work is contracted project by project rather than locked into a vendor.

Who are the platforms competing for this talent?

Six platforms carried 2,282 active listings in a census of the category taken on 8 September 2026, and one platform held most of them. Mercor accounted for 1,277 listings or 56.0%, Micro1 for 461 or 20.2%, SME Careers for 253 or 11.1%, AfterQuery for 204 or 8.9%, Turing for 82 or 3.6% and Alignerr for 5 or 0.2% (aitrainer.work).

Supply scale varies by an order of magnitude. Mercor reports a pool of over 300,000 professionals with 30,000 or more active weekly, Scale AI reports over 100,000 global contractors, and Surge AI runs roughly 50,000 expert contractors on a staff of around 110 (Sacra; Value Add VC). A platform with 110 employees managing 50,000 contractors is not running a recruitment process, and that ratio explains where the gaps appear.

How much of the market is contract rather than permanent?

Nearly all of it. The permanent alternative pays $47,000 to $87,000 a year with a $63,000 median for an in-house AI trainer (Glassdoor, cited June 2026), against $84,240 annualised for the average contract AI training expert and $224,640 for a specialist at the $108 hourly average.

Almost no listing in the category offers full-time hours. Only 3.3% of the 460 Mercor listings that disclose hours state a floor of 35 hours a week or more, against 24.5% of the 49 disclosing Turing listings (aitrainer.work, September 2026). Expert data work is structured as part-time contract engagement because the experts worth hiring already have jobs, and the same distinction between engineers who build models and engineers who put them into production decides which contractors a lab needs at each stage.

Two bands, and the gap between them is the single most important number in this market. Generalist AI trainers average $40.50 an hour. Credentialed domain experts average $108, with a ceiling of $500.

Everything downstream follows from which band a brief actually needs. A rubric-following rating task priced at expert rates wastes between 60% and 75% of the budget. An expert reasoning task priced at generalist rates produces data a lab will discard.

What do generalist AI trainers earn?

$40.50 an hour on average, $84,240 annualised at 2,080 hours. The median is $35.92, the 25th percentile $28.68 and the 75th percentile $42.66, with the full reported range running from $18.03 to $87.26 (ZipRecruiter, AI training expert, 18 September 2026).

The distribution is tight. 53% of listings fall between $24.32 and $43.20 an hour, and only 1% exceed $80.97. That compression is what makes the generalist tier straightforward to budget and difficult to differentiate on, since paying above $50 buys almost nothing a lab can measure unless the role crosses into a credentialed specialism.

What do domain experts earn by field?

Rates split sharply by field, with healthcare highest on average and research highest at the ceiling. Across 95 live Handshake AI specialist listings in October 2026, healthcare and medicine averaged $180 an hour over nine listings, science and research $131 over fifteen, legal $130 over two, finance $114 over six and engineering $106 over twelve.

Named specialisms set the top of the market: STEM professors and researchers at $500 an hour, radiology, nuclear, ophthalmology and general medical professionals at $300, and physicians at $200. Mercor publishes an average contracted expert rate of $112 an hour as of September 2026, with financial crime and KYC analysts at $75 to $100.

How do US rates compare with offshore annotation?

A US expert costs 2.92 times a South African expert and 81.8 times a Kenyan annotator on the project that produced the original comparison. The arithmetic is $108.00 divided by $37.00, and $108.00 divided by $1.32.

That second figure comes from TIME’s January 2023 investigation into the Sama contract on OpenAI work, where agents took home $1.32 an hour after tax while Sama billed OpenAI $12.50 an hour. Offshore comparators sit at $2.60 an hour for filmed trade experts in India and under $1.00 for Vietnam’s Region I wage floor from 1 January 2026. The spread is wide enough that US expert sourcing and offshore annotation are separate purchases, not two prices for the same thing, and a buyer who treats them as substitutes will get back data a reasoning model cannot use.

Short contracts, undisclosed hours and near-universal screening define how this work is bought. Durations run from 3.0 to 23.5 weeks depending on domain. Hours are disclosed in a minority of listings. Between 97.8% and 100% of listings across the major platforms require an interview (aitrainer.work, September 2026).

Those three facts together describe a market that generates continuous rebenching demand and gives buyers almost no forward visibility. Both are recruitment problems rather than platform problems.

How long do AI training contracts run?

Between 3.0 and 23.5 weeks, with the variation driven by domain rather than platform. Mercor business projects averaged 3.0 weeks across 31 listings, data science 3.6 weeks across 16 and generalist work 6.8 weeks across 10, while language work averaged 23.5 weeks across 64 listings. Turing ran longer on technical work at 10.2 weeks for software engineering across 24 listings and 14.8 weeks for business across 10.

Short cycles create a structural churn problem. A lab running a 3.6 week data science engagement replaces that bench roughly fourteen times a year, and each replacement needs the same credential check as the first. Durations this short make a permanent hire the wrong instrument, and the same arithmetic decides when interim cover returns value faster than a fixed headcount.

How many hours a week do these engagements take?

Most listings do not say, which is the operational problem. Only 36.0% of Mercor listings disclose weekly hours, falling to 4.6% on Micro1, against 59.8% on Turing and 100% on Alignerr’s small listing set (aitrainer.work, September 2026).

Where hours are stated, they are part-time. Just 3.3% of disclosing Mercor listings set a floor at 35 hours a week or more. Capacity planning against that data is guesswork, so a buyer sizing a bench has to work from committed hours agreed per contractor rather than from platform averages, and that commitment is something an agency can contract for and a marketplace generally cannot.

How are AI training contractors classified and paid?

Through contract engagement with the engager carrying the classification risk, and platform fees that often exceed agency markups. Typical US contract staffing markup runs 25% to 75%, with 35% to 50% standard in technology, and the contractor receives 50% to 65% of the bill rate (Kore1, 29 July 2026).

Platform economics are not cheaper. Prolific charges a 42.8% platform fee for commercial customers and 33.3% for academic, and Amazon Mechanical Turk charged 20% rising to 40% on tasks with ten or more assignments before closing on 30 September 2026. Mercor’s recruiting fee of 30% sits below the technology staffing norm, so a buyer choosing an agency is paying for vetting depth, contract handling and bench continuity rather than a lower headline rate.

Worked example at the 95-listing average: a $108.00 hourly pay rate at a 40% markup bills at $151.20, leaving $43.20 an hour of gross margin to cover payroll taxes at 8% to 12% of bill, insurance at 2% to 5%, benefits at 5% to 10%, recruiting and administration at 8% to 15% and net profit at 3% to 8%.

Engineering, research and healthcare carry the volume and the rate. Across the 95 specialist listings analysed in October 2026, science and research held the most postings at fifteen, engineering twelve and healthcare nine, while healthcare led on average rate at $180 an hour.

Finance and legal are real but thin, at six and two listings respectively. Both are named expansion targets for the platforms, which makes them worth building before the volume arrives rather than after.

Which domains pay the most for expert data?

Healthcare and medicine at a $180 hourly average, with a $300 ceiling across radiology, nuclear medicine, ophthalmology and general practice. Science and research follows at $131 average but carries the highest single ceiling in the market at $500 for STEM professors and researchers.

Legal averages $130, finance $114 and engineering $106. The ordering tells a buyer where to expect sourcing friction rather than where to spend, because the highest rates sit in the fields with the hardest licensing requirements and the smallest pool willing to contract alongside clinical practice.

How does engineering expertise transfer into model training?

Directly, because the reasoning a compiler or silicon engineer records is the reasoning labs are trying to teach. A model being trained to optimise kernels needs worked examples of how a human decides between two scheduling strategies, and only someone who has shipped that work can produce them.

Engineering also carries the largest transferable pool, at twelve of the 95 specialist listings and a $106 hourly average rising to $200. The network behind this bench is the one our desk already uses to place contract machine learning engineers across the US, which is why engineering briefs fill faster than clinical ones.

What makes clinical and research experts hard to source?

Licensing, indemnity and the fact that the best candidates are not looking. A board-certified radiologist earning well in clinical practice has no financial reason to answer a platform advert, so supply has to be approached rather than advertised.

Screening adds a second constraint. Between 97.8% and 100% of listings across Mercor, Micro1, SME Careers and Turing require an interview, and verifying a licence, a publication record or a board certification cannot be automated the way a rubric test can. The rates above run higher than the day rates senior ML contractors command on engineering projects, because a licence is scarcer than engineering seniority and takes longer to confirm.

Step 1. Classify the brief into generalist or expert before pricing it. Generalist rating work averages $40.50 an hour and credentialed expert work $108. Pricing the wrong band wastes between 60% and 75% of a budget or produces data the training run discards.

Step 2. Name the credential, not the job title. Specify board certification, publication record or years shipping production systems. Rate bands in this market track the credential, from $106 an hour in engineering to $180 in healthcare and $500 for STEM professors.

Step 3. Commit the weekly hours in the contract. Only 36.0% of platform listings disclose hours at all and 3.3% commit to 35 or more, so agree hours per contractor in writing rather than planning capacity from listing averages.

Step 4. Size the bench against contract duration, not project length. Engagements average 3.0 weeks for business work and 23.5 for language work. Divide the project length by the expected engagement length to get the number of rebench cycles to budget for.

Step 5. Model the bill rate before you approve the pay rate. At a 40% markup a $108 pay rate bills at $151.20, with $43.20 an hour covering payroll taxes, insurance, benefits, administration and margin. Compare that against platform fees of 30% to 42.8% rather than against the pay rate alone.

Step 6. Build the credential check into week one. Verify licences, certifications and publication records before the first task, since between 97.8% and 100% of the category screens by interview and a failed check after delivery means rerunning the work.

Step 7. Contract the replacement path up front. Agree in advance who sources a mid-engagement replacement and how fast, because short contracts and part-time hours make drop-out a scheduling certainty rather than a risk.

What does it cost to hire an AI training expert in the US?

Credentialed domain experts average $108 an hour across 95 live specialist listings as of October 2026, ranging from $22 to $500. Generalist AI trainers average $40.50. At a 40% staffing markup, a $108 pay rate bills at $151.20 an hour.

How is a knowledge worker different from a data annotator?

A data annotator records facts about data and typically earns under $25 an hour with no credential required. A knowledge worker records professional judgement in a licensed or specialist field and earns $75 to $500 an hour. The credential, not the task volume, sets the rate.

How long do AI training contracts usually last?

Between 3.0 and 23.5 weeks depending on domain, measured across 2,282 active listings in September 2026. Business projects averaged 3.0 weeks, data science 3.6, generalist work 6.8 and language work 23.5. Short cycles mean continuous rebenching rather than one-off hiring.

Can you source clinical and legal experts as well as engineers?

Yes. Healthcare and medicine commands the highest average rate in the category at $180 an hour with a $300 ceiling, and legal averages $130. Both require licence verification before first task, which we complete during week one of any engagement.

Is it cheaper to use a platform than a recruitment agency?

Not reliably. Prolific charges commercial customers a 42.8% platform fee and Mercor’s recruiting fee is 30%, against a 35% to 50% technology staffing markup. Platforms price lower on headline fee and higher on unverified supply, drop-out and rebench cost.

Tell us the domain, the credential and the committed hours. We will come back with a shortlist of US experts who hold the licence or record your training run needs, with rate bands, availability and a named replacement path for every contractor on the bench.