8 min read
By
Anand Sivadasan
Founder, Das Nexus
The Operating Record Is Now An Asset
For two decades, I negotiated partnerships at Goldman Sachs, PayPal, Citi and Capital One. Protecting data rights was often one of the toughest parts of getting a deal done.
Now I’m having conversations I would barely have imagined a year ago: what would it take to license the internal records of a business to an AI developer? I hear from a new company about this almost every day.
The amounts are getting attention. In August, Google submitted the winning $10 million bid for Spirit Airlines’ de-identified business records, with Mercor named the alternate bidder at $7.5 million. The proposed package included email, collaboration records and data from core business systems. As of October 15, court approval was still pending, with a hearing scheduled for November 6. ¹ ²
Earlier this year, cielo24’s CEO told Forbes that its 13-year history of Slack messages, Jira tickets, email and shared drives brought in hundreds of thousands of dollars as the company wound down. On October 9, micro1 announced a commitment to spend $1 billion over the following year acquiring and licensing enterprise operational data. ³ ⁴
These records capture things that rarely appear in a manual: why a quote was changed, how a disputed invoice was resolved, when a customer deserved an exception. AI developers want examples of those decisions to help their systems do the work.
For business leaders, that raises questions beyond price. What could your records teach? What have buyers already learned elsewhere? And would licensing them surrender an advantage you would rather keep?
This article introduces the Value Trail framework, a way to work backward from the economic value of an AI capability to what a company’s records contribute and what it might receive in return.
The devil’s dilemma
Sell your operating history, and you may help train systems that compete with your business. Refuse, and a competitor may supply similar records, collect the payment, and leave you facing much the same technological change.
That is the devil’s dilemma. Neither choice guarantees protection.
It is sharpest where substitutes exist. If competitors can provide much the same knowledge, withholding your archive may do little to slow the technology. If your records contain expertise that others cannot readily supply, refusing a license may preserve real leverage. Some records should never be offered, regardless of price.
Why the records matter
Explaining a freight brokerage is different from running one. To quote a load, an AI agent may need to read an email, check a transport system, compare carriers, recognize a missing constraint and negotiate an alternative. A clean document rarely captures that whole sequence. A connected operating history sometimes does.
Developers use these records as training examples, as material for practice environments in which agents attempt tasks and receive feedback, and as private evaluations of whether a system can do the job. Think of a flight simulator built from the decisions, tools and exceptions of an actual business.
An intermediary layer has formed to acquire and prepare the material. Mercor and micro1 run company-data programs alongside their expert-data work. Pewter and Polyshares arrange or enter operating-record licenses. Ooak Data turns enterprise records into practice environments. SimpleClosure and Sunset help monetize records from businesses winding down. These are overlapping businesses with different products and rights. ³ ⁸ ⁹
The broader suppliers are already substantial. Mercor reportedly exceeded $2 billion in gross annualized revenue in June; micro1’s CEO reported an annual run rate above $1 billion in October; Proximal reported $200 million in annualized revenue in September. Those figures cover more than historical archives, and run rates are not contracted recurring revenue. They show the scale of the surrounding industry, not the size of the market available to an individual seller. ⁵ ⁶ ⁷
Follow the value backward
Most conversations begin with the archive: how many years, how many files, which systems. I would begin with the work the buyer wants AI to perform.
The Value Trail framework works backward from the economic value an AI capability could create, to what a company’s records contribute, to what the owner can capture.
1. Capability value: What useful work could AI perform?
Start with a specific capability: quoting freight, coordinating maintenance, answering leasing inquiries or resolving a claim. Estimate the work involved, its cost, the share AI could handle, adoption and the oversight that would remain.
2. Your contribution: What do your records add for this buyer?
Years of linked decisions and outcomes may be valuable. So may a narrow set of difficult exceptions. A large archive with repeated examples of the same routine process may contribute less than its size suggests.
3. Your capture: What could you earn, and on what terms?
The price depends on the buyer’s alternatives, competing offers, preparation costs and the rights granted. A one-time license for a defined use is a different product from a perpetual grant with broad resale rights.
Keep two permissions separate: whether the owner can license the same records again, and whether this buyer can sublicense them to several labs. Either can change the economics. So can a requirement to supply new records, build tasks or help evaluate results.
What the numbers can tell us
Published examples provide reference points, but few disclose enough detail to establish a price list.
Operating record | Published amount | Basis |
|---|---|---|
Spirit Airlines | $10M winning bid | Proposed Google transaction; approval pending as of October 15 ¹ ² |
cielo24 | Hundreds of thousands of dollars | Completed sale reported by its CEO to Forbes ³ |
Industrial distributor | $1.2M | Pewter-published example ⁸ |
Freight brokerage, about 100 people | $650K | Pewter-published example; license scope not disclosed ⁸ |
Engineering firm, about 70 people | $450K | Pewter-published example; license scope not disclosed ⁸ |
Closed startups through SimpleClosure | Typically $10K–$100K per company | Range reported by its CEO to Forbes ³ |
These are selected examples, not a representative sample of transactions.
We built a capability-by-capability illustration anchored to the $650,000 freight example, then applied the same approach to a hypothetical apartment operator at the scale of roughly 100,000 operating units. The apartment case is scaled to the public disclosures of a large apartment owner-operator, with estimated workflows; it does not value any company’s actual data. ¹⁵
Illustration | Base gross license payment | After substantial comparable acquisition |
|---|---|---|
Freight brokerage | $650K, fixed by calibration | About $90K |
Large apartment operator | About $5.6M | About $0.6M |
The freight payment is the anchor, not independent validation. The apartment figure is a cross-sector scenario that depends heavily on assumed buyer coverage, the contribution assigned to real operating records and the seller’s share. The same capture rate is used in both sectors as a simplifying assumption. Cash proceeds after costs and risk-adjusted economic value are separate outputs in the working model.
Your edge is what buyers haven’t learned yet
The first useful examples of an unfamiliar workflow may close an important capability gap. Further examples of a well-covered process may add less. That gives owners a reason to understand buyer demand before close substitutes become widely available.
Language-model research supports diminishing returns from additional and increasingly repetitive data under the conditions studied. Limited repetition can still be useful, and larger training efforts can need more data. These findings inform the logic; they do not establish archive prices or the rate at which those prices fall. ¹¹ ¹² ¹³ ¹⁴
The advantage is about capability coverage, not the calendar. A buyer’s purchase doesn’t prove it has learned a workflow, a competitor’s rare failures may stay valuable after its routine records don’t, and another lab may have different gaps.
In our illustration, later comparable acquisitions sharply reduce value, but how sharply is an assumption, not a measured relationship. If additional records keep adding value at a similar rate, timing matters far less. Rising adoption, new buyers and harder tasks could also offset the decline.
For large operators, more volume eventually stops adding much; breadth of workflows and exceptions matters more than size. Fresh records and ongoing evaluation work may support a relationship beyond a one-time sale.
The agreement matters as much as the archive
Spirit illustrates how scope and safeguards can shape a transaction. Passenger transaction systems, refunds, inflight purchases and WiFi sales were excluded from the proposed sale. An independent agent would de-identify the remaining unstructured data and certify the procedures before transfers. Employee-data objections remained unresolved. These are features of a proposed transaction, not evidence that processing or delivery has occurred. ²
De-identification does not settle ownership or contractual permission. Nor does a buyer’s scale eliminate security risk: Mercor confirmed that it was affected by a LiteLLM supply-chain attack in March 2026. ¹⁰
Before sharing records, establish which material can be licensed, what must be excluded, who can access raw data and where it will be processed. Negotiate permitted uses, sublicensing, retention, deletion, acceptance and payment. Understand how a license could affect a future sale of the business.
Begin with an inventory of systems, dates and record types, rather than a bulk export. Compare buyers against the same defined scope. A first offer tells you one buyer is interested; it does not establish market value.
A decision worth making deliberately
I expect buyers to become more selective as they acquire comparable records. Owners who understand their workflows and rights will be better placed to decide what to offer, what to retain and when to say no.
The devil’s dilemma will not disappear. Value Trail makes the economics more explicit; it cannot decide how much competitive advantage an owner should surrender.
I have spent much of my career protecting data in commercial partnerships, and I am still working out what the right answer is here. If you have received an offer or are weighing one, I’m glad to compare notes and walk through the working model behind these figures.
Let’s compare notes
I have spent much of my career protecting data in commercial partnerships, and I am still working out what the right answer is here. If you have received an offer or are weighing one, I’m glad to compare notes and walk through the working model behind these figures. GET IN TOUCH → to walk through it.
This article is for information only and is not legal, tax or investment advice. Das Nexus may earn referral fees from some companies mentioned here, and discloses any such arrangement to clients.
About the model
The figures are illustrative scenarios, not market valuations. We anchor the model to one published freight transaction and apply it to a hypothetical apartment operator using public company and labor data. Workflow volumes, buyer coverage and seller economics are assumptions. The freight result matches the transaction by construction; the apartment result is an extrapolation. Figures shown are gross license payments before seller costs. We’re happy to walk through the model with anyone interested, hear where the assumptions could be improved, and help develop it for your company’s operating record.
Sources
3. Anna Tong, AI’s New Training Data: Your Old Work Slacks And Emails, Forbes, April 16, 2026.
5. micro1, end-of-Q3 update, October 12, 2026. Company-reported annual run rate.
6. Dealroom, Mercor doubles to $2B gross revenue run rate as AI labs buy expert data, June 2026.
7. Proximal, New Frontiers, September 29, 2026. Company-reported annualized revenue.
8. Pewter, published deal examples.
9. Program descriptions: Mercor, micro1, Polyshares, Ooak Data, SimpleClosure.
10. Mercor, Update on Mercor security incident, June 25, 2026.
11. Kaplan et al., Scaling Laws for Neural Language Models, 2020.
13. Lee et al., Deduplicating Training Data Makes Language Models Better, ACL, 2022.
14. Google DeepMind, An empirical analysis of compute-optimal large language model training, 2022.