Why AI labs pay for operational records when the public web is exhausted
Models have read the public internet. What they have not seen is how a real company runs. That gap is why buyers now license operational records, and why yours may be worth more than you think.
Every frontier model has read the public internet. The articles, the forums, the documentation, the open-source code: that material has been collected, cleaned, and trained on many times over. The research group Epoch AI has published a widely reported projection that the stock of public, human-written text could be fully used by training runs between 2026 and 2032. The exact year is debatable. The direction is not.
What no crawler has ever seen is how a real company runs. Not what it says on its website, but what happens when a request comes in on a Tuesday afternoon, who decides, what the exception was, and how it turned out. That knowledge sits in support desks, inboxes, chat channels, and CRMs inside established businesses. It is the one large body of human work that is both unread and still being produced. That is why buyers now pay to license it.
The public web shows what companies say
A marketing page is the clean, finished version of a company. A press release is the version the company wants remembered. Even a detailed knowledge base describes the policy, not the practice. All of it is useful to a model, and all of it has already been learned.
What that material lacks is the middle of the work. A refund policy says refunds are issued within ten days. It does not show the handful of cases where the policy was overridden, the reason a manager approved the override, or the message the agent wrote afterwards. A model that has read every policy on the internet still does not know how a policy is applied by people under time pressure.
What a model learns from a resolved ticket
Take one support ticket from a company that has run the same desk for six years. The customer describes a problem in their own words. An agent asks two clarifying questions. The issue is tagged, routed, escalated once, and closed with a resolution note. Later, a satisfaction score arrives.
That single record contains a problem statement, a diagnostic conversation, a decision, and an outcome. Multiply it by tens of thousands of tickets and you have something the public web cannot provide: a dense, labeled record of how problems are actually solved in one domain, with the failures included. Buyers describe this as workflow data, decision data, or outcome-labeled data. The plain version is that it shows work being done rather than described.
The same is true of a sales pipeline with closed-won and closed-lost outcomes, an engineering tracker with the review discussion attached, or a billing system that shows which disputes were upheld. The common thread is that a human made a judgement, and the record shows what happened next.
Why the buyers changed
Three things happened at roughly the same time.
First, the free supply thinned. Labs have been open about the fact that high-quality public text is a finite resource, and more of what remains online is itself model-generated, which is of limited use for training the next model.
Second, licensing became normal. Reuters reported in 2024 that Reddit had agreed to license its content to Google for roughly $60 million a year. News publishers signed reported multi-year agreements with model builders. Paying for a source stopped being unusual.
Third, the market reached operational data. In 2026 the business press reported a bankruptcy auction for the data assets of the former Spirit Airlines, described in reports as roughly 100 million emails, 500 million chat messages, and 30 million lines of code. The opening bid was reported at $5 million; competing bids from Google and AI data companies were reported well above that, subject to court approval. The point is not the airline. The point is that a company's ordinary working records were treated as a distinct asset with a market price, and more than one buyer showed up.
What operational means in practice
The records buyers ask about are not exotic. Email and calendar. Documents, proposals, and standard operating procedures. Team chat. Recorded calls, where consent exists. CRM activity and deal notes. Support and service tickets. Billing, disputes, and collections. Engineering workflows, including the discussion around code review.
You do not need to be a technology company. A regional accounting firm, a logistics operator, a property manager, or a specialty manufacturer has all of this. What matters is that the same workflows have run in the same systems for a few years, so the records are consistent and deep enough to learn from. Buyers are grouped by what they build rather than by name: frontier labs, specialist data labs, model builders, and enterprise AI teams. Each wants a slightly different slice, which is why the same records can draw different numbers from different buyers.
What this means for an owner
A few practical consequences follow.
Your records are an asset you already own, and no one else can produce them. A competitor cannot generate your six years of tickets, and a lab cannot crawl them.
The value depends on depth and continuity more than on industry. Our qualification is 20 or more desk staff, three or more years of records, and data in three or more systems. Companies below that line are not turned away; they are told what to start documenting so the records are worth something later.
You do not have to send anyone data to find out what it is worth. Buyers price from a one-page manifest, an inventory of what exists and where, and later from a small de-identified sample you approve. The dataset itself stays in your systems until a contract is signed, and then moves directly to the buyer, never through an intermediary.
And the price is not fixed. The same manifest can draw very different numbers from different buyers, depending on what each one is building this quarter. That is the case for pricing the records before they go to market, and for running a competitive process rather than accepting the first inbound number.
The public web was the first supply. Operational records are the second, and they belong to the companies that produced them. If you want to know what yours are worth, start with the manifest. A first offer follows within 24 to 48 hours of a complete one. It is indicative and buyer-benchmarked, and binding terms follow sample review.
Questions worth asking
Why would an AI lab want records from a company that has nothing to do with AI?
Because the records show work being done, not described. A model trained on public text knows what a refund policy says. A model trained on five years of support tickets knows how refunds are actually decided, escalated, and closed, including the exceptions. Buyers pay for that second kind of knowledge because it is the part the public web cannot teach, and it exists only inside companies that have run the same workflows for years.
Is the public web really running out?
The widely reported projection from the research group Epoch AI is that the stock of public, human-written text could be fully used by frontier training runs somewhere between 2026 and 2032. Whether the exact year is right matters less than the direction. The easy, free supply is finite, more of what remains online is itself generated by models, and labs have responded by paying for licensed sources. Reported deals with publishers and forums are the public evidence of that shift.
Does this mean every company's data is valuable?
No. Buyers want depth and continuity: a few years of history, enough volume to learn from, and records that show outcomes rather than just activity. The qualification we use is 20 or more desk staff, three or more years of records, and data in three or more systems. Below that line, the right move is usually to start documenting consistently now so the records are worth something in a couple of years.
Measured from a complete data manifest. A first offer is a process target, not a guarantee; buyers decide eligibility, price, and terms. Indicative value ranges mentioned on this site are illustrative, not quotes. Nothing here is legal advice; have your own counsel review any agreement before data moves.