The record
Two records were examined, along with a wider question.
The first is the working file of an Indian accountancy firm: the client's ledger balances, the items the auditors tested, the queries they put to the client, review notes by the partner, the audit opinion as signed, the tax return and any later ruling by the tax office.
The second is what an outsourcing firm in India or the Philippines holds for clients abroad: processed invoices, reconciliations, medical-billing claims and payroll runs, with the client's acceptance or rejection of each.
The wider question is whether labs that are not at the frontier, meaning American labs a step behind, Chinese labs, and model-builders backed by a region or a state, buy local professional work at all.
An illustration, not a real file: a firm in Bengaluru audits a small manufacturer. A junior notices that year-end stock is 20% higher than the year before. The manager asks for the count sheets. The partner requires slow-moving items to be written down. The client objects, and the opinion goes out with a paragraph drawing attention to the point. A year later the tax officer's ruling accepts the valuation. From first decision to outside result takes more than a year.
Why it might matter for training AI
Accounting work with a later outside result is the kind of record labs are said to want. The hope was that labs behind the frontier would buy it more cheaply from countries where a great deal of such work is done.
Nothing found supports that. What the evidence shows instead:
- Chinese labs are reported to buy tasks written by experts, and packaged datasets already sold to American labs, from the very American vendors that supply the frontier labs 1. The reporting describes tasks and packaged sets, not company archives.
- India's model-builders work from customers' own data and state programmes. Sarvam sells tuning on a customer's own data and says that data is not pooled or used to train Sarvam's models 2. BharatGen, which is state-funded, is building a store of Indian-language data and cites more than 15,000 hours of annotated speech across 22 languages 3.
- In the Gulf, Mistral and Humain announced a collaboration which they put at several hundred million euros and which covers Arabic models among other things 4. No data purchase was disclosed.
- For Japan, Korea, Southeast Asia and Europe, one search found nothing on professional-work data. That is weak evidence of absence.
Who holds it and who may share it
Indian accountancy firms. The firm keeps its audit file. An earlier screen for this project recorded that the Indian accountants' institute bars members from disclosing client information without specific authority. The institute's code was not re-opened for this dossier, so its exact wording is not established here.
Outsourcing firms. The client abroad owns the records. The outsourcer is a processor, meaning it handles them only on the client's instructions.
India's data-protection law. The rules under the Digital Personal Data Protection Act were notified in November 2025, and its central duties, covering consent, notice and purpose, begin in May 2027 5 6. A law firm's summary says that an Indian firm processing foreigners' personal data under a contract with a party outside India is exempt from most of those duties 6. The Act itself was not read. If that summary is right, an outsourcer need not meet Indian requirements for its foreign clients' data. The exemption does not hand it any right to that data.
The Philippines. Its privacy law was not opened.
Australia. For Australian content, a single buyer has named a budget. Sovereign Australia AI says it has earmarked $10 million to acquire data, paying news services and buying books and music 7. The Copyright Agency reported in October 2025 that Australian AI start-ups, Maincode among them, had pledged to pay for the Australian content they use 8. For neither company was a signed deal or a sum actually paid found. No buyer of Australian professional-work data was found.
Who is already there
American task vendors sell into China 1. Vendors with staff in India sell expert labour and tasks to American labs; LH2 AI Labs, for one, says it supplies frontier labs and enterprises with rights-cleared datasets drawn from under-represented populations 9. It publishes no prices, customers or list of datasets. It sells to frontier labs from India, and that tells us nothing about follower labs buying Indian files.
Chinese domestic suppliers are emerging. One of them, Benyuan Zhishu, organises professionals in mathematics, coding, law and medicine 10.
Nobody was found selling professional files from India or the Philippines. The simple reason is that the holders are not free to sell them.
What would have to be stripped
Indian tax and audit files carry national identity and tax numbers, bank details and salaries. Once those go, little is left. Outsourcers' archives carry other countries' personal data, including American patients' records, under contracts whose breach would end the outsourcer's business.
What stands in the way
- The holder has no right to sell: the profession's code and client contracts stand in the way.
- No buyer has been shown for foreign professional files. Follower labs want what frontier labs have already tested.
- Brokers of company records that state a preference ask for English-language, American companies first.
- Take the identifiers out of an Indian tax file and little of substance remains.
- Selling toward China carries a cost in standing with American buyers.