AfterQuery Hit 3.2 Billion Dollars Paying Experts to Write AI Training Data

AI training-data startup AfterQuery has reportedly raised a round valuing it at $3.2 billion, five months after announcing a $30 million Series A at a $300 million valuation in April, TechCrunch reported, citing Forbes as first to break the round. At Metaintro, the valuation is the least interesting part. What matters for careers is how the company makes money, which is by paying working professionals like doctors and lawyers to encode how they actually do their jobs, turning years of accumulated judgement into a directly sellable asset for the first time.

What does AfterQuery actually do?

It trains models and agents on how professionals work rather than only on whether answers are factually correct. The company describes this as encoding the patterns, decisions and reasoning of the world's best practitioners, which is a meaningfully different product from labelling images or checking outputs. The distinction is between teaching a model what a correct answer looks like and teaching it the route a competent professional takes to reach one.

To do that it employs knowledge professionals including doctors, lawyers and other specialists to do model training work. That places it in the same category as Scale and a handful of other firms building expert data operations, but pointed at process rather than accuracy. The company has named Nvidia, Legora and the Korean AI lab Motif Technologies among its customers, and said in April that it had reached an annualised revenue run rate of $100 million.

Why did it grow this quickly?

Because the constraint in frontier AI development has moved. Compute and general web text are widely available, while structured records of how expert work is actually performed are not, and they cannot be scraped because they mostly exist in people's heads. A company that can reliably produce that material is selling into extremely concentrated demand from a small number of very well-funded buyers.

That concentration cuts both ways for anyone considering the work. A handful of buyers with enormous budgets keeps rates high while they are competing, and leaves contributors exposed if one of them changes strategy or decides it has enough data in a given field. Treat it as a strong supplementary income with an uncertain duration rather than as something to build a plan around, and take the work while the rates reflect scarcity rather than assuming they will hold.

The pace is genuinely unusual. More than a tenfold valuation increase in under half a year, and according to a Y Combinator partner the fastest any startup has gone from launch to unicorn status in the accelerator's history. The founders are 22 and 23 and went through the Winter 2025 cohort just 18 months ago. That combination should prompt some caution as well as interest, since valuations set in a scramble for scarce supply are the ones most exposed if that supply becomes less scarce.

What is the actual job on offer here?

Structured expertise capture, done part-time and paid by the hour or by the task. In practice it means working through cases in your own field and documenting not just the conclusion but the reasoning, the alternatives you rejected, what additional information you would have wanted, and where a less experienced practitioner would go wrong. It is closer to writing teaching material than to data entry, which is why it requires practitioners rather than annotators.

That difference explains the economics. Annotation can be distributed to a very large pool of people with short training, which drives rates toward the floor. Expert reasoning capture cannot, because the supply is limited to people who already hold the qualification and the experience, and no amount of demand creates more of them quickly. Scarcity of supply rather than generosity of buyers is what holds the rates up, and it is also why this market may narrow as the most codifiable specialities get covered.

That is a different proposition from the annotation work that has characterised this industry so far, which has typically been low-paid and precarious, and which carries its own risks as 720 Meta annotators facing job cuts showed. Expert data work sits at the opposite end of that market on rates, because the pool of people who can supply it is small. The same dynamic in a different sector is visible in robotics firms paying Indian factory workers to teach machines their own tasks.

Should professionals actually take this work?

For most, it is worth evaluating seriously as supplementary income rather than dismissing. The rates for genuinely scarce expertise are considerably above general freelance work, the hours are usually flexible, and unlike most side income it uses the qualification you already hold rather than requiring you to build something new. For a doctor or lawyer looking at side income options, it is one of very few that pays at anything near professional rates. Most supplementary income available to high earners either pays far below their hourly rate or demands the evenings and weekends they do not have, which is why so little of it gets taken up. Work that pays close to professional rates and can be done in ninety-minute blocks is a genuinely different proposition.

There are real considerations before signing anything. Check whether your employment contract or partnership agreement restricts outside work in your professional capacity, since many do. Check what you are permitted to use, because no real case material, client information or patient data should ever enter this work, and the obligation to prevent that sits with you rather than the platform. Check who owns what you produce and whether you are granting a perpetual licence to your own methodology. And be clear-eyed about what you are selling, which is the transferable part of your professional judgement.

Professional bodies are beginning to take positions on this. The American Medical Association and the American Bar Association both publish guidance on AI in professional practice, and while neither currently prohibits this kind of work, the ethical framing around confidentiality and competence applies to it directly. Checking your own regulator's current position takes a few minutes and is considerably cheaper than discovering the answer afterwards.

Are you training your own replacement?

That is the honest question and it deserves a straight answer rather than reassurance. You are contributing to systems designed to perform parts of your work, and pretending otherwise would be dishonest. The counterargument is that this happens whether or not you personally participate, and declining only removes your influence over how well it is done and forfeits the payment.

The more useful frame is which parts of a profession are actually being encoded. The codifiable, repeatable, procedural portion is what transfers, while the parts requiring physical examination, live negotiation, accountability and judgement under uncertainty do not, and those are also where most of the value in senior professional work sits. This tension is already visible inside workplaces, where a third of workers are hiding what they know from the AI they were told to train, and it is the same calculation arriving as a paid opportunity rather than a company mandate. It also produces the paradox that companies are replacing the experts AI still needs to learn from.

There is a version of this decision that is straightforwardly reasonable. If the encoded expertise makes a competent practitioner faster rather than replacing the practitioner, contributing to it is closer to writing a textbook than to digging your own grave, and professions have always transmitted knowledge that way. The uncomfortable difference is that a textbook needs a reader who has trained for years, and a model does not, which compresses the time between your contribution and its consequences.

What does this mean for your career?

The immediate takeaway is that deep domain expertise now has a second market. Historically your professional knowledge was only sellable through practising, employment or consulting, all of which convert your time into money at a fixed ceiling. Expert data work is a third route, and while it is unlikely to replace a professional income it can supplement one materially and it values seniority directly, since the whole point is that you know things a junior does not.

The longer-term implication is more important than the side income. "In combination with layoffs over the recent years plus the rise of required AI skills, experience is no longer enough," Lacey Kaelani, chief executive of Metaintro, told Fortune. What this market suggests is a refinement of that. Experience that can be written down is being bought and encoded. Experience that cannot, because it lives in physical skill, relationships, accountability or judgement in unclear situations, is becoming comparatively more valuable. Knowing which kind yours mostly is, is a better guide to your next decade than any general prediction about AI.

There is a timing argument too. Being early to this market is worth more than being thorough about it, because the first contributors in a speciality set the templates and are paid to define the standard, while later contributors are paid to fill it in. If your field is one of the obvious targets, the window where the work is well paid and intellectually interesting is now rather than in three years, and it closes as the speciality gets covered.

Which professions are most affected?

The ones with high-value, documentable reasoning and clear right answers, which is why medicine and law appear first. The Bureau of Labor Statistics reports 2025 median pay of $151,160 for lawyers, while physicians and surgeons sit among the highest-paid occupations it tracks, and expensive professional hours are exactly what a buyer wants to automate.

That said, the effect is uneven inside each profession. Legal research, document review and first-draft work are more exposed than advocacy and client relationships. Diagnostic reasoning from records is more exposed than examination, procedures and the conversation with a frightened patient. Finance, engineering, accounting and architecture all contain the same split. Frameworks published by the National Institute of Standards and Technology increasingly shape how these systems must be documented and evaluated in regulated settings, which is itself creating professional work for people who understand both the domain and the standards. Occupation-level pay and outlook data across all of these fields sits in the Occupational Outlook Handbook if you want to compare where your own speciality stands. The practical exercise for anyone in these fields is to divide your own week into work that could be captured in a written procedure and work that could not, then deliberately grow the second half, which is the same conclusion reached in the AI paradox facing companies replacing their own experts.

One caution about that exercise is worth stating. People consistently overestimate how much of their work is irreducibly human, because the parts that feel effortful to them are not the same as the parts that are hard to specify. A task that takes concentration is not therefore uncodifiable, and a task that feels routine may depend on context you have stopped noticing you apply. Doing the audit honestly usually means asking a colleague to describe what you do, since the things you cannot see about your own work are exactly the things a written procedure would omit.

Related Articles

• Summer Associate Cuts Hit Law Schools, What JD Grads Should Do
• Tesla Offers Competitive Pay for Data Collection Operators
• Physician Board Certification and What Each Specialty Requires
• The Skills and Know-How AI Still Cannot Replace
• What Primary Care Doctors Earn and Why the Number Still Falls Short
• What Law School Grads Need to Know as AI Reshapes Legal Careers
• An AI Law Firm Just Won Its First Court Case
• Want to Be a Leader, Technical Expertise Is No Longer Enough
• The Career Advantage AI Still Cannot Copy
• How to Show AI Skills on Your Resume and in Interviews

People Also Asked

Q: What is expert AI training data work?

A: It is paid work where qualified professionals document how they reach decisions in their own field, so that models and agents can be trained on the reasoning rather than only on correct answers. AfterQuery describes this as encoding the patterns, decisions and reasoning of the world's best practitioners, and it employs doctors, lawyers and other specialists to produce it.

Q: How much does this kind of work pay?

A: Rates are not published and vary by field and platform, but expert data work sits well above general annotation work because the supply of qualified contributors is small. It is normally structured as flexible, task-based or hourly contract work, which makes it viable as supplementary income alongside a professional role rather than as a replacement for one.

Q: Does doing this work threaten your own profession?

A: Partly and honestly, yes, since you are helping systems perform parts of what you do. What transfers is the codifiable, procedural portion of professional work. Physical examination, advocacy, negotiation, accountability and judgement under genuine uncertainty do not transfer well, and those are where most senior professional value sits.

Expertise that can be written down is now something you can sell twice, and expertise that cannot is worth more than ever. Metaintro indexes live roles across professional fields so you can see how the market is currently pricing what you know. Create your free profile and benchmark your experience against what employers are paying for now.