Blog · Trends, tools and stories

Why AI Companies Now Pay Domain Experts to Teach Their Models

Public deals from 2025 show a shift: general data labeling is giving way to specialists who know a field well enough to judge the answers.

By Dijae Dadula · November 18, 2026 · 6 min read

Dark matte grey illustration of a lone green-accented expert figure beside a stack of document cards and a checkmark panel, no text

For a few years, the common picture of AI training was a large crowd of general workers labeling images and grading short replies. That picture is changing. In 2025, several public deals showed AI companies paying for specialists: people who know finance, medicine, law or software well enough to tell whether a model's answer is actually right.

This guide explains what the public record shows, why the shift is happening, and what a business owner can learn from it. It uses only reported facts with dates and sources. Where the reporting is thin, it says so.

What you will get
  • Four public 2025 data points that show the shift to domain experts
  • Why general labeling stops being enough once models answer hard questions
  • A simple test to check whether a vendor's “expert” claim holds up

1. What changed in 2025

Four public reports show the pattern. Each one is a fact reported by a named outlet, with the date given.

  • June 2025: Meta buys into Scale AI. Meta announced a $14.3 billion investment for a 49% stake in Scale AI. Scale supplies human workers who label and evaluate training data for AI developers. The deal placed a large data supplier inside one of the biggest AI programs in the industry. Fortune, June 13, 2025.
  • September 2025: xAI reshapes its training team. Engadget, citing Business Insider, reported that at least 500 workers on xAI's data-labeling team were cut on September 12, 2025. The company described a “strategic pivot”: it said most generalist tutor roles were no longer needed and that it would surge its specialist AI tutor team by 10 times. Engadget, September 2025.
  • October 2025: OpenAI hires former bankers. Bloomberg reported that more than 100 former investment bankers were recruited at about $150 per hour to write prompts and build financial models for training. Fortune, October 21, 2025.
  • October 2025: Mercor raises $350 million. The AI hiring and expert marketplace Mercor announced a $350 million Series C at a $10 billion valuation, led by Felicis Ventures. TechCrunch, October 27, 2025.

Read together, the four reports tell one story. The work is moving from broad, short tasks toward longer tasks that need a real domain background.

Timeline of four 2025 public reports on expert data workA horizontal timeline from June to October 2025 with four dated markers: Meta and Scale AI, xAI specialist tutor pivot, OpenAI banker project, and Mercor Series C.TIMELINE · 2025JunMeta invests inScale AISepxAI shifts tospecialist tutorsOctBanker projectat OpenAI reportedOctMercor raises$350M at $10BDates are the publication dates of the cited reports.
Figure 1. Four public reports from mid to late 2025. Each marker is one report, placed by its publication date.

2. Why general labeling is no longer enough

Labeling a photo as a cat is a task most people can do in seconds. Judging whether a tax answer follows the current rules, or whether a clinical summary misses a contraindication, is different. The reviewer must know what a correct answer looks like before they can mark a wrong one.

That is why the public deals focus on specialists. The model is no longer only learning to be fluent. It is learning to be right in a narrow field, and the people who can check that work are rare and costly to find.

Dijae's rule Before you assume an AI product is accurate in your field, ask who checked its answers, what they knew, and how the checks were recorded. If nobody can say, treat the output as a draft.

3. How the work flows from company to expert

Public reporting describes a chain with three parts: an AI developer that needs specific knowledge, a supplier that recruits and manages experts, and the experts who review model output against clear instructions. The supplier usually handles recruiting, scheduling and quality control. The developer sets the task and receives the results.

Flow from AI developer to supplier to domain expertsThree boxes connected by arrows: the AI developer sets the task, the supplier recruits and manages experts, and experts return reviewed output to the developer.FLOW · HOW EXPERT WORK REACHES A MODELAI developersets the taskSupplierrecruits and checksDomain expertsreview outputReviewed results return along the same path to the developer.
Figure 2. A simplified flow based on public reporting. Real arrangements vary by company and by agreement.

Each step in that chain adds a question a business owner should ask. The developer has to define the task clearly enough that an expert can judge it. The supplier has to keep the reviewers qualified and consistent, which is hard when reviewers work in many countries and time zones. The reviewers have to follow written criteria, because two experts who judge the same answer differently produce data that teaches the model nothing useful. The public reports say little about how these steps are run, so the chain above should be read as a general model, not a description of any one company.

4. How to test an expert claim

Vendors often say their model is “expert-trained”. Treat that as a claim to test, not a guarantee. Use this check on any AI product you are considering for your business.

QuestionWhat a good answer looks like
Who reviewed the output?Named role and field, such as a licensed accountant or a practicing clinician
What did they check?A written list of accuracy criteria, not a general “quality” label
How are errors tracked?A log of failures that is reviewed and fed back into the product
What can it not do?Stated limits and cases that go to a person
Success metric Pick three real questions from your business. Score each answer as correct, partly correct or wrong. Accept a tool only if it scores at least 90% correct on your own questions, with every wrong answer caught before a customer sees it.

There is also a quality question on the other side of the table. A model trained on expert judgment still needs an owner who decides when its answer is good enough to show a customer. That owner is usually the business, not the vendor, so the review step belongs in your own process as well as the vendor's.

5. What the public record does not tell us

An honest limitation: the figures above come from press reports, not from audited company filings. The xAI figure is an “at least” count reported through Business Insider, and the Mercor numbers are as reported at the time. The exact scope of each program is not public. Do not treat these reports as a market size or a forecast.

Frequently asked questions

Are AI companies still using general data workers?

The reporting shows both. The xAI case is about moving away from generalist tutor roles toward specialists, not a complete end to general work.

Is this relevant to a small business?

Yes, as a buyer. The same question applies to any AI tool: who checked it, and against what standard. Ask that before you rely on it for customer answers.

Does this mean experts will replace AI?

No. The public record describes experts improving models, and the models still need people to set tasks, review results and make final decisions.

Build an AI workflow with checks built in. If you want help designing review steps for your own AI assistant, book a free AI Automation Blueprint call: www.dijaedadula.com/book. Or join the free weekly newsletter for one practical idea each week: Subscribe.

References

  1. Fortune, “Meta's Scale AI investment and the superintelligence team,” June 13, 2025. fortune.com
  2. Engadget, “xAI reportedly laid off at least 500 AI tutors working on Grok,” September 2025 (citing Business Insider). engadget.com
  3. Fortune (citing Bloomberg), “OpenAI recruits former bankers for finance work,” October 21, 2025. fortune.com
  4. TechCrunch, “Mercor quintuples valuation to $10B with $350M Series C,” October 27, 2025. techcrunch.com

Keep reading

Get one practical idea like this every week

Join the free newsletter on AI agents, automation and turning your skills into income. Prefer video? Watch the walkthroughs on YouTube.

Subscribe free
Book a free call

Want a system like this built for your business? Book a free AI Automation Blueprint call, or see the done-for-you services at my consultancy, IRD Digital Marketing Services.

DIJAE DADULA

Dijae Dadula · AI automation and systems specialist · Cavite, Philippines · Founder of IRD Digital Marketing Services

Some links on this site are affiliate links. If you buy through them, I may earn a commission at no extra cost to you.

© 2026 Dijae Dadula. All rights reserved.