AI POLICY & MARKETS · ENTERPRISE DATA

The Next AI Training Dataset Is a Company's Working Day

The emerging market for enterprise operational data, the economics of human expertise, and the challenge of teaching AI agents how real work gets done.

By Manuel Herranz·

Some time ago, I read that in the future people would be paid for the data they generate simply by living their lives. Every click, every journey, every message. By extension, organizations would monetize not only the products and services they sell, but also the trail of information they leave behind while producing them.

At the time it sounded like science fiction. It belonged to the rain-soaked Los Angeles of Blade Runner, a city of neon and acid rain where corporations manufactured memories and implanted them in others. A world in which everything, even experience itself, had a price.

I thought it was plausible. Every year, a little more so. In our last newsletter we looked at how AI is rewriting the operating system of the translation industry: generating text is becoming cheap, while data, evaluation, adaptation and control are becoming valuable. This week, I realized that the same logic is spreading far beyond language. It is reaching the working day itself.

A question that once seemed hypothetical

Just over an year ago, I had a conversation with a large software company about the kinds of data that might become valuable for future AI systems. At one point, my interlocutor asked me to imagine what it would mean to have access to the corporate correspondence and operational records of a major industrial group (she mentioned a pretty well-known European conglomerate). I explained that we did not represent that company and could not offer access to its records and that we had other types of training data. We both understood it was an example.

But the conversation stayed with me, partly because the example seemed so ambitious. What possible use could justify acquiring years of internal communications, decisions and everyday business activity? I think I am beginning to understand the question rather better now.

The prophecy arrived through a web form

It did not arrive aboard a ship off the shoulder of Orion. It arrived as one more submission on the contact form of our website. A broker in France wrote to tell us that he was lining up French companies willing to license their operational data. Not consumer data. Companies with more than fifty employees, mostly white-collar, that had used the same tools for three years or more. Their emails, documents and threads with clients and partners, mostly in French, with a good share of English. He’s not been the first one.

Before sending any our way, he asked three very merchant-like questions. Do we buy data from French companies? How does the price compare with a US company of the same size? Which functions or sectors are we short of right now? Someone, in other words, is already walking from office to office asking what companies keep in their drawers, so that it can be sold by weight.

Then, almost as if to complete the picture, another unsolicited email arrived. This time, someone was approaching Pangeanic not as a potential buyer of corporate data, but as a potential seller of its own operational history: support tickets, CRM records, internal procedures, project documentation. The offer was simple. Would we like to know what years of our own working records might be worth? Within hours, I had encountered both sides of a market that, not so long ago, barely had a name. A hypothetical question from that earlier conversation was beginning to acquire a recognizable market on both sides.

Behind that message lies a principle that deserves to be stated plainly. Companies may soon earn revenue not only from what they produce, but also from the knowledge that is recorded while they produce it.

A market is taking shape

The broker’s message did not come out of nowhere. Investigative reporting now describes businesses and their advisers exploring the licensing of workplace records: internal correspondence, project histories, operating procedures and the decisions embedded in them. Some potential purchasers want to turn such material into realistic environments where AI agents can learn to perform professional tasks. The market is still young, and announcements of investment or acquisition intentions should not be confused with completed transactions.

One visible example is the proposed sale of internal business data from the bankrupt airline Spirit Airlines. In October, Reuters reported that Google had agreed to pay $10 million for an operational archive containing corporate communications and related records, subject to court approval. A competing offer adding 25% to the price stopped the deal. Passenger booking and loyalty information was to be excluded, while worker representatives and lawmakers raised privacy concerns. As of 13 October 2026, the sale had not received final court approval.[1]

The phenomenon is not confined to large companies. In April, Forbes reported that businesses winding down were licensing years of Slack discussions, Jira tickets and email archives through specialists that help close companies and prepare their digital assets. This is not simply a search for more documents. It is a search for traces of how real organizations function.[2]

AI laboratories and the data companies serving them increasingly want more than additional text. Publicly available writing alone rarely records the full context of a task, the decisions made along the way, the constraints that shaped those decisions and, ideally, whether the work succeeded. That operational context is what makes these archives interesting.[5]

Archives are not experience

There is an important difference between a collection of corporate documents and a record of how work actually gets done. Imagine a French company that has spent five years managing supply contracts. A folder of signed contracts has some value. A dataset that reconstructs the proposed amendments, the reasons behind each change, the internal consultations, the approvals and the final outcome has far more. The first teaches a model what a contract looks like. The second teaches an agent how a contract is negotiated.

The translation industry has walked this road before. For decades we built language technology from bilingual corpora, revised translations and other records of human linguistic work. The same principle is now extending to more complex professional activity. What is being learned is no longer simply a correspondence between languages. It is a sequence of decisions and actions taken under real organisational constraints.

This is where precision becomes very relevant. A growing share of knowledge work consists of observable sequences of decisions and actions that can be partially automated, taught and evaluated. Automating individual tasks, however, is not the same as replacing entire professions, particularly where accountability, expertise and complex judgement remain essential.

Nor is an observed sequence the same thing as a validated training experience. A record shows what an employee did. It does not show what they should have done, and it does not always show what actually happened as a result. Real archives contain shortcuts, mistakes and bad habits alongside good judgement. Turning them into something an agent should learn from requires separating those three layers, which is why expert review and verified outcomes matter as much as volume.

The European question

This is where the film turns darker, and where Europe has something distinctive to say. A company may own its servers, but owning the archive is not the same as being free to license everything inside it. A corporate email thread is full of people who never signed anything: employees, customers, suppliers, the Austrian partner who replied in copy.

Several questions follow, and none of them has a universal answer:

  • Who has the right to sell? Business correspondence contains personal data, trade secrets, third-party information and confidentiality obligations to clients.
  • What about the employees? Their judgement and expertise are precisely what gives the data its value. Depending on how the data is used, and whether it touches on monitoring of staff or their working conditions, employee representatives such as a French works council may need to be involved.
  • Can customer data move to a new controller? Reusing it to train AI is a different purpose from the one for which it was collected, and requires an applicable legal basis and a careful assessment of purpose compatibility and other GDPR requirements. France's data protection authority, the CNIL, has set out how legitimate interest can apply to AI training, and the test is demanding.[3]
  • Is de-identification enough? Not necessarily. De-identification removes direct identifiers; anonymization, in the legal sense, requires that individuals cannot be identified by means reasonably likely to be used. Only truly anonymous data falls outside the GDPR. The European Data Protection Board has made clear that anonymity must be assessed case by case, and the bar is high. Removing names from an email thread does not automatically remove the person (you may be identified as “the CEO’s sister”, “the Founder’s successor” etc.)[4]
  • Who benefits? The company, the broker, the lab, or the people whose experience is being turned into machine capability?

That last question contains a paradox worth stating without ideology. A company may be paid for licensing decades of information generated by its employees, while the models trained on that information may automate part of those same employees' future tasks. There are possible answers: economic participation for workers, profit sharing, contractual rights over how their work is reused, data governance with employee representation, and limits on downstream use. None is settled, and the objections raised in the Spirit Airlines case suggest that these questions are already being contested.[1]

All of this also bears on the broker's question about price. There is no reason for an automatic discount simply because the data is French. Whether genuine French–English working data commands a premium is still a commercial hypothesis, and it will depend on what buyers need. What is certain is that compliance has a cost, and that the final value depends on the use case, the reconstruction of tasks and, critically, on whether the proposed use can be justified and governed lawfully.

From archive to evidence

Raw operational data, taken straight from a mail server, is closer to a liability than an asset. Its value as a trustworthy training or evaluation asset depends on demonstrating at least three things: clear rights of use, verifiable provenance and measurable utility. When a business process crosses languages and jurisdictions, as so many European ones do, there is a fourth dimension of quality: linguistic and cultural fidelity. A negotiation conducted partly in French and partly in English is only useful if both halves, and the shifts between them, are preserved and understood.

That is the work we have been doing at Pangeanic for more than two decades, applied to a new raw material: detecting and removing personal data across languages without destroying meaning, cleaning and aligning bilingual material, annotating and evaluating it, and documenting where every record comes from. There is also a variant that may suit Europe particularly well. A company does not need to sell its data to monetize it. It can use its own records to train or evaluate private agents, without handing that knowledge to anyone else.

There is a further reflection that applies to us. Pangeanic has its own history of operational knowledge: translation memories, human revisions, quality estimations, corrections, linguistic procedures, annotation and evaluation. Much of it contains highly specialized expertise. But it must be separated with care from material that belongs to clients or is subject to confidentiality and restrictive licences. Even if we never sold any of it, the part we can legitimately use is an ideal laboratory for learning how to reconstruct tasks, preserve context and build reliable evaluations. Understanding this market from operational experience, not only from intermediation, is what makes the difference.

Tears in rain

At the end of Blade Runner, the replicant laments that everything he has seen will be lost with him. Every company carries something similar: years of decisions, negotiations and tacit knowledge that nobody ever wrote down in a manual. Much of that knowledge has left with people who retired or changed jobs, or disappeared when a company closed. From now on, it may have a price.

The question is no longer whether companies will license the knowledge recorded in their working day. Some already are. The question is how it will be done without also selling the privacy of employees and customers, and without reducing human expertise to an anonymous raw material.

And there is one question I would leave open, for our readers and for ourselves. If AI models learn from decades of human experience recorded inside companies, who will capture the economic value of that learning?

Sources and further reading

  1. Reuters. Lawmakers raise alarm at Google plan to acquire Spirit Airlines data for AI models. 8 October 2026.
  2. Forbes. AI’s New Training Data: Your Old Work Slacks and Emails. 16 April 2026.
  3. CNIL. Relying on the legal basis of legitimate interests to develop an AI system. 5 January 2026.
  4. European Data Protection Board. Opinion 28/2024 on certain data protection aspects related to the processing of personal data in the context of AI models. 18 December 2024.
  5. Business Insider. AI Labs Are Fueling a Wild Market for Your Emails and Slack DMs. 13 October 2026.