Which Enterprise Data Types AI Developers Value, and Why

1. Introduction#

Language models are built from data in stages, and each stage uses data with different properties. For several years the main input was text collected from the public web. That input is now the subject of two concerns. One is quantity, since forecasts suggest the supply of high-quality public text could be exhausted within this decade.¹ The other is access, since a growing share of web sources restrict automated collection.² Against this background, several AI data companies and model developers have begun to publish pages inviting companies to license their internal data.³⁴⁵⁶⁷

This review asks three questions. First, how is training data used across the stages of model development, so that the demand for specific data types can be placed in context? Second, which enterprise data types do AI developers and their suppliers say they want, and what reasons do they give? Third, what attributes of a dataset are associated with higher value, according to the research literature and the public record of transactions?

The scope is deliberately narrow. We do not estimate what any dataset is worth, we do not assess legal rights to sell data, and we do not evaluate de-identification methods. Those topics are covered in other papers in this series. We also do not claim that any particular company buys any particular category of data unless a cited source says so.

2. Methods and sources#

We conducted a structured reading of four groups of sources between September and October 2026.

1. Research literature on how training data is used and how its properties affect results. We included foundational papers on pretraining and scaling, instruction tuning and learning from human feedback, data filtering and de-duplication, synthetic data and model collapse, software-engineering benchmarks, and agent benchmarks. Most are preprints on arXiv, which we label as preprints. Findings in preprints may differ from versions later accepted by journals or conferences.
2. Forecasts of data availability, principally from Epoch AI, and a large audit of web restrictions by the Data Provenance Initiative.
3. The published data-acquisition pages of five AI data companies: Mercor, micro1, Scale AI, Handshake AI and Turing (Project Lazarus). We read these pages as statements of what each company wants, not as evidence of what it has purchased.
4. News and trade reports of completed or attempted purchases of company data, including the sale of failed startups' archives and the bankruptcy auction of an airline's corporate data.

Inclusion required that we opened and read the source in full or, for long documents, the abstract and the sections relevant to the claim. Where sources disagree we say so. Company pages change frequently, so all statements about them are as of October 2026. Terms such as "wants" and "values" are used strictly for what a source states. A page that lists a data type is evidence of stated interest. A press report of a sale is evidence that a transaction was reported, usually from anonymous sources, and does not establish its terms.

3. Findings#

3.1 How training data is used#

Pretraining is the stage in which a model learns general language and world knowledge from very large corpora. The compute-optimal scaling analysis known as Chinchilla found that, for a fixed compute budget, model size and the number of training tokens should grow together, so that doubling model size calls for doubling the training data.⁸ Practice has since moved beyond that rule. Epoch AI notes that models are often trained on more tokens than compute-optimal because the resulting smaller model is cheaper to run, and it describes Llama 3-70B as trained about ten times beyond the compute-optimal point.¹ Pretraining is therefore the stage with the largest appetite for sheer volume.

Post-training adapts a pretrained model to follow instructions and behave helpfully. It has several components, and the literature treats them separately.

• Supervised fine-tuning (SFT) trains the model on demonstrations of desired responses. FLAN, an early instruction-tuning study, tuned a 137-billion-parameter model on more than 60 tasks phrased as natural-language instructions and reported that it outperformed zero-shot GPT-3 on 20 of 25 evaluated tasks.⁹
• Reinforcement learning from human feedback (RLHF) uses human rankings of model outputs to refine a model after SFT. In the InstructGPT study, human evaluators preferred outputs from a 1.3-billion-parameter model trained this way to outputs from the 175-billion-parameter GPT-3.¹⁰
• Reinforcement learning against a verifiable task, increasingly run inside simulated software environments, trains a model by rewarding it when a task is actually completed. Trade press describes these "RL environments" as training grounds that simulate what an agent would do in a real software application.¹¹

Evaluation is a distinct use of data. Benchmarks need realistic tasks with checkable outcomes. SWE-bench, for example, built 2,294 software-engineering tasks from real GitHub issues and their corresponding pull requests in 12 Python repositories, and reported that the best model it tested, Claude 2, resolved 1.96 percent of them.¹² OpenAI's GDPval built tasks from work products created by professionals averaging 14 years of experience across 44 occupations, and released a 220-task gold subset.¹³ These benchmarks illustrate why developers value data that reflects real work: realistic tasks can be used both to measure models and, once the benchmark is public, as a template for training tasks.

The stages differ in the property they reward. Pretraining rewards breadth and volume. SFT and RLHF reward the quality and consistency of a comparatively small number of examples. Reinforcement learning against verifiable tasks rewards realism and a reliable way to score the outcome. Enterprise data is most often discussed in connection with the second and third of these.

3.2 Why proprietary data is discussed now#

The case for proprietary data rests on forecasts of public data supply, on evidence that web access is narrowing, and on concern about synthetic substitutes.

Supply. Epoch AI estimated, in a paper published in June 2024, that the stock of public human-generated text of usable quality is about 300 trillion tokens, with a 90 percent confidence interval of 100 trillion to 1,000 trillion, and that this stock would be fully used at some point between 2026 and 2032 depending on training choices.¹ A later Epoch analysis states that the largest known training sets are about 15 trillion tokens, compared with roughly 500 trillion tokens of de-duplicated indexed web text, and that at 4x annual growth in compute text data could run into a "data wall" in about five years. That later analysis also estimates that images and video could add substantial additional tokens.¹⁴ The estimates are uncertain and depend on assumptions about quality and repeated epochs. A controlled study of data-constrained scaling found that repeating data for up to four epochs yields negligible changes in loss compared with unique data, which stretches the supply but with diminishing returns beyond that point.¹⁵

Importantly for this review, Epoch's original estimate excludes private data such as instant messages. The authors state that such data seems unlikely to be used at scale because of legal issues and because it is fragmented across platforms.¹ The present interest in enterprise data can be read as a test of that assumption, now that companies are offering to license it.

Access. The Data Provenance Initiative audited 14,000 web domains and reported that more than 5 percent of all tokens in the C4 corpus, and more than 28 percent of the most actively maintained critical sources, became fully restricted through robots.txt in a single year, while 45 percent of C4 is subject to terms-of-service restrictions on crawling.² Whether or not such restrictions are honored in every case, they reduce the share of the open web that developers can rely on.

Synthetic substitutes. A natural response to scarcity is to generate training data with models. The evidence is mixed. A widely cited study reported that training on model-generated content causes "model collapse," in which the tails of the original distribution disappear, and argued that data from genuine human interaction with systems becomes more valuable as generated text spreads across the web.¹⁶ Epoch's projections exclude synthetic data, describing the evidence as nascent and mixed, while noting that verification is often easier than generation, which could make large-scale synthetic data possible in some domains.¹⁴ Self-Instruct, which bootstraps instruction data from a model itself, improved GPT-3 by 33 percent absolute on one benchmark and came within 5 percent absolute of a model tuned with private user data and human annotation in a human evaluation, which shows both the promise and the remaining gap.¹⁷

These sources do not show that enterprise data improves models. They show why developers have an incentive to look for new data sources and why they emphasize authenticity.

3.3 What buyers say they want#

Five AI data companies have published pages describing the company data they seek. Table 1 summarizes the data types each names. The table records stated interest only.

Table 1. Enterprise data types named on the public data-acquisition pages of five AI data companies, as of October 2026.

The lists overlap heavily, which is the main finding of this table. Messaging, email, documents, project tracking, CRM and code appear on at least three of the five pages. Stated criteria also overlap. Scale and Handshake set minimum company size or age. Lazarus ties value to the length and completeness of the record. micro1 and Mercor tie compensation to quality, uniqueness, volume and depth.

Two further points from these pages matter for interpretation. First, the pages describe de-identification or anonymization as part of the process and give the data owner control over what leaves its systems.³⁵⁶ Second, none of the pages states which AI developers ultimately use the data or for which training stage. Mercor's chief executive has said publicly that customers do not want to hand over data that would automate large parts of their value chains, which is why labs hire former practitioners to describe workflows.¹⁹ That statement suggests that the supply of authentic enterprise records is limited by company reluctance, and it frames licensing programs as a way to obtain what experts alone cannot reproduce.

micro1 announced in October 2026 that it had committed $1 billion over twelve months to its Company Data Partnerships program, with financing from Citi and Hercules Capital, and said the data is used to build reinforcement learning environments that model workflows, decisions and exceptions.²⁰¹⁸ That is a stated budget, not a record of purchases.

3.4 Data types and their model uses#

Table 2 maps each data type to the model uses that sources associate with it and to the value drivers they cite. It combines the buyer pages above with the research literature and with transaction reports. Where a source does not link a data type to a use, we say that we found no public evidence and do not infer one.

Table 2. Enterprise data types, associated model uses and stated value drivers, with the strength of public evidence.

The strongest public evidence of demand concerns messaging, documents, project data, CRM records and code, which are exactly the categories that describe how white-collar work is done. The weakest evidence concerns categories that are valuable to the business that owns them but are not obviously a record of reasoning or decisions, such as marketing performance metrics. We found no source that supports or rules out interest in that category, and we flag it as unverifiable from public sources rather than as unwanted.

3.5 Attributes associated with higher value#

Across buyer pages, press reports and the research literature, six attributes recur.

History length and depth. Lazarus states that a long operating history is worth more than a short one.⁷ Scale asks for years of records.⁵ Mercor ties compensation to the depth of data.³ A company that sells wind-down archives, Sunset, was reported to consider headcount, years in operation, data volume and how the team used its tools.²³ The common thread is that longer histories contain more decisions, more edge cases and more cases in which the first approach failed.

Outcomes and labels. Training and evaluation both need a way to say whether work was done well. The Epoch review of RL environments quotes practitioners who name soundness as the most important quality, meaning that a high reward must reflect that the task was actually solved and not exploited, and who say enterprise workflows are attractive because their outcomes are measurable.²⁷ SWE-Gym, a training environment for software-engineering agents, packages each of its 2,438 task instances with a codebase, an executable runtime and unit tests, and reported gains of up to 19 percent absolute in resolve rate on SWE-bench variants for agents tuned on its trajectories.²⁴ In customer service, the tau-bench benchmark scores an agent by comparing the database state at the end of a conversation with an annotated goal state, and reported that even strong agents succeeded on under half of tasks.²¹ These examples suggest that records with an attached outcome (a test result, a closed ticket, a changed database state) are more usable for verifiable training than records without one. This is an inference from the design of these systems, not a finding about enterprise datasets specifically.

Linkage across systems. Scale describes its target as interconnected data that mirrors how the business runs, and it preserves thread structure and participants in chat and email.⁵ In the Spirit Airlines sale, the purchase agreement reportedly required that the anonymization preserve referential integrity so that records remained linked to individual people, which a union argued could allow reconstruction of members' information.²⁸²⁹ The episode shows both sides of linkage: it is part of what makes data valuable, and it is also what makes de-identification harder.

Domain specificity and rarity. Scale asks for domain expertise that makes data hard to replicate.⁵ Commentary on Turing's enterprise code strategy notes interest in niche languages and complex system architectures, with COBOL as an example, and flags intellectual property ownership as the main open issue.³⁰ The logic is that common material is already abundant in public corpora.

Quality. The research literature is consistent on quality. LIMA fine-tuned a 65-billion-parameter model on only 1,000 curated prompt and response pairs, with no reinforcement learning, and its authors concluded that almost all knowledge is learned in pretraining and that a small amount of instruction tuning suffices to shape outputs.³¹ The phi-1 study trained a 1.3-billion-parameter code model on about 6 billion tokens of web data filtered for "textbook quality" plus about 1 billion tokens of synthetic textbooks and exercises, and reported 50.6 percent pass@1 on HumanEval.³² DataComp-LM found that model-based filtering is key to assembling a high-quality training set, and a 7-billion-parameter model trained on its resulting 2.6-trillion-token baseline reached 64 percent on 5-shot MMLU.³³ De-duplication also matters. One study found a single 61-word sentence repeated over 60,000 times in C4, that over 1 percent of unprompted model output was copied verbatim from training data, and that de-duplication reduced such emission by about a factor of ten.³⁴ The Stack, a corpus of permissively licensed source code, found that near-de-duplication significantly improved performance.³⁵ For enterprise data, these studies imply that cleaning, de-duplication and curation raise the usefulness of a dataset, and that much of the value in a licensed archive may sit in a small curated subset.

Rights clarity. Rights appear in the buyer pages as terms: Lazarus offers a "clean rights transfer," Handshake requires acceptance of licensing terms, and Mercor and Scale give the data owner approval over each package.⁷⁶³⁵ The Spirit bankruptcy sale, which was objected to by former employees, and the report that founders selling startup archives have little visibility into buyers, illustrate that unclear rights and privacy constraints can complicate otherwise attractive data.²⁹²³

3.6 Evidence from reported transactions#

Reported transactions provide the only public indication of what buyers did, as distinct from what they said.

In April 2026 Forbes, as relayed by Gizmodo, reported that failed startups were selling old Slack and email archives to AI developers through an intermediary, SimpleClosure, whose service also licensed source code, documents and workflows.²² In October 2026 MIT Technology Review reported that foundation labs, vendors of simulated workspaces and mature technology companies were buying code bases and workspace archives (emails, chats, documents and data from tools such as Notion, Figma and Slack), and that a vendor of reinforcement learning environments, Surge AI, looked for conversations that include disagreements and difficult judgment calls.²³ The same report said buyers' valuations rely on "research theses" estimating how many useful tasks the data could yield, which is consistent with the task-generation framing in Section 3.1.²³

In August 2026 Google reportedly won a bankruptcy auction for Spirit Airlines' corporate data with a $10 million bid. Reports describe the archive as including about 100 million emails, 500 million Teams messages and around 30 million lines of code, and a second report describes more than 30 million recorded customer service calls, 17 million OneDrive files and 20.5 million SharePoint items.²⁹²⁵ A Surge AI representative quoted by Time said the purchase makes sense only if one believes an agent will generalize across the economy, and the article noted that office tasks are harder to grade than code, so it is unclear whether such data will speed progress as much as coding data did.²⁹ That caveat is important: the commercial logic is a thesis about generalization, not a demonstrated result.

A last data point concerns synthetic substitutes for support data. The Bitext customer-support dataset on Hugging Face has 26,872 question and answer pairs across 27 intents, is described as a hybrid synthetic dataset, and is presented as a first step before a company fine-tunes on a small amount of its own data.³⁶ This is consistent with a view that public or synthetic support data teaches the format of a task while real records supply a company's policies, edge cases and vocabulary. The dataset's own description supports the first half of that statement; the second half is our interpretation.

4. Discussion#

Data type matters less than what the records capture. The buyer lists in Table 1 are long and similar, which suggests that no single category is the target. What recurs is the idea that the data records real decisions made by experts over time, in context, with enough structure to turn into tasks and, where possible, scored outcomes. A chat archive in which people disagree and resolve problems fits that description. A flat export of counts does not.

Code has the clearest line to model improvement. Code is the one category in which the literature links the properties buyers cite (executable tests, issue and review history) to measured gains through benchmarks and training environments.¹²²⁴ Evidence for white-collar workflow data is, by comparison, at the thesis stage. Time's reporting notes that outcomes are harder to grade in office work.²⁹ Developers may be willing to pay for options on uncertain value, but a data owner should not read interest in code as evidence of equivalent demand for other categories.

Quality arguments cut in both directions. Small curated datasets have improved models in the LIMA and phi-1 studies, which suggests that a licensed archive does not need to be large to be useful.³¹³² The same logic implies that most of a large archive may be redundant, and that deduplication and filtering would take place before or after sale.³⁴³³ The buyer pages place the work of anonymization and extraction on the buyer or broker, which shifts the cost but not the need.³⁵

Linkage and privacy pull against each other. The attributes that make records valuable, namely identifiable threads, participants and linked tables, are the ones that make removal of personal data hard. The Spirit sale shows the tension in a contract term that requires referential integrity while a union argues that linkage permits re-identification.²⁸²⁹

Intent is not volume. The $1 billion micro1 commitment and the dollar ranges on buyer pages indicate willingness to spend, but they are stated programs and do not show how many companies have been paid or how much.²⁰ Reported actual purchases are few, are often from anonymous sources, and involve failed or bankrupt companies, whose owners have different incentives from operating businesses.²²²³

5. Limitations#

This is a narrative review, not a systematic one, and sources were selected by us. Buyer pages are marketing documents and may overstate demand, and they change without notice. Most reports of purchases rely on anonymous sources, and we could not verify them independently. Several of the research papers are preprints and were read at the abstract level for headline findings. They concern general language models and are not tests of enterprise datasets, so applying them to enterprise records is an inference. Epoch's forecasts are uncertain by design, and the paper we cite for the original estimate dates from June 2024.¹ We reviewed only English-language public material and only a subset of AI data companies. We found no public source on marketing performance data. Absence of evidence in public sources does not show absence of demand.

6. Conclusion#

AI developers and their suppliers publicly describe a consistent set of enterprise data types they want: communications, documents and procedures, project and ticket records, CRM and business-platform records, and code with its review history. They justify the interest by pointing to realism, history, linkage and expertise that public data lacks, and the research literature supports the broader claim that data quality and diversity can matter more than volume in post-training. The public record of actual purchases is thin and concentrated in distressed sales, and the link between workplace data and measured model gains is stated as a thesis, with code as the main category supported by benchmark evidence. A reader assessing whether a given dataset is of interest should therefore ask whether it records real decisions with checkable outcomes, whether it can be linked and de-identified, and whether the owner can license it. Those questions are drawn from the sources reviewed and do not constitute a valuation.