AI data market
Enterprise data licensing investment thesis: the 7-job missing middle
By SourceX Editorial · Reviewed by Noah Loul ·
10 sections
Short answer
The enterprise data licensing investment thesis rests on a gap. Software platforms store company records, and terms such as Slack's bar third-party AI training on them, while AI labs and data marketplaces want finished, rights-cleared datasets. Seven jobs sit between, from qualifying a holder to passing a buyer's acceptance test. The thesis is investable only if demand reaches beyond a few buyers and those jobs get cheaper each time they are done.
License terms
Enterprise data licensing
Key takeaways
- The middle is a sequence of seven jobs, and the investment question is who can do all seven repeatedly, not who holds the most records.
- Platform terms push the work toward the data owner: Slack's API terms bar apps offered outside the developer's organization from training LLMs on API data.
- Buyers ask for finished goods: OpenAI's data partnerships page wants large-scale data not readily available online and excludes third-party information.
- Marketplaces assume the jobs are done: AWS Data Exchange makes providers attest their right to distribute and keep products updated.
- The cleanest falsifier is buyer breadth: Reddit attributes substantially all of its licensing contract value to two partners.
What is the missing middle, and is it investable?
The missing middle is the work between a company's operating records and a dataset an AI developer accepts under license: finding the holder, scoping, proving authority, exporting with context, preparing, contracting and passing acceptance. It is investable only if buyers are numerous enough and the work gets cheaper each time it is done.
The phrase describes a structural gap, not a market size. On one side sit the platforms where records live, such as CRM, ticketing, ERP and messaging systems, whose terms govern what leaves them. On the other sit AI developers and data marketplaces, which buy or list datasets that are already scoped, cleared and documented. Neither side is organized to turn one mid-sized company's records into the other side's finished input.
An enterprise data licensing investment thesis therefore has to answer two questions at once. Does demand for those finished datasets reach beyond a handful of buyers, and can someone do the middle work repeatedly at a cost buyers accept? This article maps the jobs, tests why incumbents stop short of them, and sets out a falsifiable thesis map with evidence on both sides.
Seven jobs between a company's records and an accepted license
Seven jobs turn records into an accepted dataset, and each produces an output a diligence reviewer can ask to see. The map lists that proof for each job, who is positioned to do it today and where it usually breaks; the entries are editorial judgments, not survey results.
Two features of the map matter for investors. The jobs run in sequence, so a failure late in the chain, at acceptance, wastes the cost of every earlier job. And the early jobs are company-specific: qualifying one firm's ticket archive teaches little about the next firm's claims system, which is why repeatability, not record volume, is the economic question.
| Job | Proof it is done | Positioned today | Where it breaks |
|---|---|---|---|
| 1. Qualify a holder | Inventory of systems, years covered and record counts | No default owner inside the company | Short or inconsistent history, too few records |
| 2. Scope against a buyer task | Written task, fields and acceptance test agreed with a buyer | Buyer data teams, mainly for large-scale datasets | Records match no named task |
| 3. Prove authority to license | Rights review signed off by someone able to sign | The company's counsel, case by case | Customer contracts, third-party content, no signer |
| 4. Export with context | Sample export with links and history intact | The company's admins, using the platform's own export tools | Platform terms, or exports that drop context |
| 5. De-identify and prepare | Processed sample, method and residual-risk note | Specialist vendors, hired by either side | Redaction strips the decision context buyers need |
| 6. Contract the license | Signed agreement with scope, term and permitted use | Each side's legal team, deal by deal | Exclusivity, liability or term negotiations stall |
| 7. Deliver and pass acceptance | Buyer acceptance recorded against the agreed test | The buyer's evaluation team | Quality defects or missing context found late |
Why software platforms stop at the export button
Platforms hold the records on their customers' behalf, and terms like Slack's restrict AI use by third parties. That leaves the owner's own export, under the owner's authority, as the main route out, and it puts the first four jobs on the owner's side of the line.
Slack's API terms show the pattern. An application offered outside the developer's organization may not use API data to train a large language model, and may not bulk export message and file data unless an additional agreement allows it. A third-party tool therefore cannot become the licensing route, while a workspace owner exporting its own data, and deciding what to license, still can.
The thesis needs this to hold across platforms, not just one. A vendor that launched an opt-in program to license customer records to AI developers would occupy the middle directly, with distribution an intermediary cannot match. That is a falsifier in the thesis map below; this article found no such program at a major CRM, ticketing or ERP vendor to cite, though it did not survey every vendor.
Why labs and marketplaces buy finished datasets, not raw records
Buyers on the other side want goods that are already scoped, cleared and documented, which pushes the preparation, contracting and acceptance jobs back toward the supply side.
OpenAI's data partnerships page describes its target as large-scale data that reflects human society and is not already easy for the public to find online, and it excludes anything sensitive or personal as well as anything owned by a third party. Operating records are full of customers, employees and counterparties, so they reach that bar only after de-identification and a rights review that someone else has to run.
Marketplaces assume the same. AWS Data Exchange requires each provider to attest that it has the legal right to distribute the data it publishes, to keep products regularly updated and to run a defined customer support process. Unless a provider qualifies for its Extended Provider Program, products may not include information that can identify a person unless that information is publicly available. A marketplace lists finished products; it does not turn an archive into one.
The middle exists by subtraction. Platforms are not set up to license what they hold for customers, and buyers and marketplaces are not set up to build what they buy. Whether that gap is a business depends on who pays for crossing it.
Who pays for the middle work, from which budget, and how?
A thesis needs a named budget owner and a commercial mechanism, not just a gap. Three budget owners appear in public material, each buying for a different reason, and the mechanism in all three is a license to a defined scope with an acceptance test.
Data partnership and post-training teams at model developers are the clearest. OpenAI publishes the criteria its data partnerships use, and Epoch AI's 2024 study, which estimated about 300 trillion tokens of quality public human text and projected full use between 2026 and 2032 on trend, describes the pressure behind paying for non-public sources.
Evaluation teams are the second. OpenAI built GDPval's 1,320 tasks across 44 occupations with professionals averaging more than 14 years of experience, and Anthropic's January 2026 guidance tells agent builders to source realistic tasks from the failures they see. Evaluation budgets buy small, hard, held-out sets, which favors scoped records over raw volume.
Vertical AI companies and enterprise AI teams are the third, and the least documented in public sources. They are the test of breadth: if the only budgets sit at a few labs, the middle is a services market for those labs rather than a market of its own.
The mechanism is the same across all three: a license to a defined scope, with acceptance against a written test. In SourceX's version, the supplying company agrees the scope and terms and signs the license first, and only then are records delivered; from that point SourceX can use, sublicense and distribute the de-identified dataset within the limits the agreement sets. That is a description of the mechanism; it is not evidence that the mechanism has reached scale.
The falsifiable thesis map: six claims with evidence on both sides
Each claim below can be checked, and each has a stated falsifier. The status column is an editorial reading of the cited evidence on 26 October 2026, not a measurement, and a reader with the same sources can disagree with it.
Two claims carry most of the weight. If demand never reaches beyond a few buyers (claim 2) or the middle work never gets cheaper with repetition (claim 6), the other four can hold and the business still disappoints.
| Claim | Evidence for | Evidence against | Would falsify it | Status |
|---|---|---|---|---|
| 1. Developers pay for non-public records | OpenAI's partnerships page seeks data not readily available online | Over 98% of Nemotron-4 340B alignment data was generated (NVIDIA, 2024) | Disclosures showing business skills built on public and generated data alone | Supported for non-public data in general; unproven for operating records |
| 2. Demand reaches beyond a few buyers | Evaluation suites such as GDPval need real work across 44 occupations | Reddit FY2025 10-K: substantially all licensing contract value from two partners | Buyer counts that stay at a handful across filings and requests | Against, on public evidence |
| 3. Platforms will not fill the middle | Slack API terms bar third-party LLM training and bulk export | Platforms hold the records and the customer relationship | A major vendor launching opt-in licensing of customer records | Holding; one vendor's terms cited |
| 4. Labs will not source mid-sized firms directly | Partnerships criteria favor large-scale sets and exclude third-party data | Labs can take stakes in suppliers: Meta's investment valued Scale AI at over $29 billion | Lab programs aimed at mid-sized companies' operating records | Mixed |
| 5. Marketplaces will not fill the middle | AWS Data Exchange requires attested rights, updates and support | Marketplaces could add rights and preparation services | A marketplace offering rights review and de-identification as a service | Holding; one marketplace cited |
| 6. The middle work gets cheaper with repetition | None public | EU training-content summaries add documentation work | Time to acceptance and first-pass acceptance flat across successive deals | Unproven; company-level data not public |
Can a large stock of records coexist with weak demand?
Yes, and that is the strongest alternative to the thesis. Companies with 50 or more employees typically run systems full of records, but a stock is not a market: records become supply only when a buyer has a task for them and someone has done the seven jobs.
At a glance
Can a large stock of records coexist with weak demand
Public evidence shows how thin demand can be even for data with obvious value. Reddit, whose archive is large and widely used, says substantially all of the contract value tied to its licensing revenue comes from two partners. If a well-known public archive concentrates on two buyers, one mid-sized company's operating records face a harder test.
Three readings separate the cases, and a company in the middle should know which one it is in:
- Supply-bound: buyers issue written requests with named tasks that go unfilled, so the middle work is valuable and scarce.
- Demand-bound: cleared, prepared datasets wait without buyers, so the middle is a cost center however well it runs.
- Execution-bound: requests and datasets both exist but deals stall in review, export or acceptance, which is the case an operator can fix.
Five developments that could close or shrink the middle
Five developments would weaken the thesis even if the gap is real today. Each has a signal an investor can watch:
- Synthetic substitution: NVIDIA says over 98 percent of Nemotron-4 340B's alignment data was generated; if business-task training follows, real records shrink to a test-set niche.
- Buyer concentration: a two-partner pattern like Reddit's gives buyers pricing power and renewal leverage over any supplier, however good its process.
- Integration by labs: equity ties such as Meta's investment in Scale AI show labs can pull supply in-house or close to it.
- Documentation cost: EU general-purpose AI providers must publish training-content summaries on a mandatory template, enforceable by the AI Office from 2 August 2026, which rewards documented supply but raises the cost of producing it.
- Vendor capture: a platform that adds opt-in licensing of customer records would take the middle with its own distribution.
Questions for any company claiming the middle, SourceX included
Ask for proof at each job, using definitions you set. A company that claims the middle should be able to show where its deals stall and whether each deal costs less than the last.
By its own description, SourceX handles the full sequence for companies that license operational data: qualifying and sourcing supply, reviewing rights, contracting, preparing the data, delivering it and handling payment, a span that touches all seven jobs. Its deal and acceptance metrics are not public; company and commercial metrics go to investors directly. The raise it is running is open only to accredited investors, under Rule 506(c), and the investors page takes requests for materials.
Questions that map to the seven jobs:
- Of companies that passed qualification, what share reached a signed rights review, and what were the rejection reasons?
- How many distinct buyers issued written requests with a named task in the last four quarters?
- Which platform exports does each delivered dataset depend on, and whose terms were reviewed?
- What share of datasets passed the buyer's acceptance test on first delivery, with the denominator stated?
- Has time from signed agreement to accepted delivery fallen across successive deals, with open deals counted?
- Which falsifier in the thesis map would most change the business, and what is the plan if it arrives?
Methods note: how the job map and thesis map were built
Both the seven-job map and the six-claim thesis map are editorial constructions prepared for this article on 2026-10-26. They draw on Slack's API terms, OpenAI's data partnerships and GDPval pages, AWS Data Exchange documentation, Reddit's FY2025 Form 10-K, NVIDIA's Nemotron-4 340B publication page, Epoch AI's 2024 data study, Anthropic's January 2026 eval guidance, Scale AI's June 2025 announcement and the European Commission's template guidance.
Not measured: how many companies hold licensable records, buyer counts beyond public filings, prices, preparation costs and any SourceX metric. Platform evidence rests on one vendor's terms and marketplace evidence on one marketplace's documentation; the status readings are judgments a reader can redo with the same sources.
Frequently asked questions
What is enterprise data licensing?
It is the licensing of a company's own operating records, such as support tickets, claims, invoices or call transcripts, to an AI developer for training or evaluation under a written agreement. It differs from content licensing because the records were never published and usually contain information about customers and staff that must be reviewed and de-identified first.
Why don't software vendors license the data they host?
They host it on behalf of customers, and their terms restrict how others may use it for AI; Slack's API terms, for example, bar apps offered outside the developer's organization from training LLMs on API data. A vendor could build an opt-in licensing program for customer records, which is why that appears as a falsifier in the thesis map.
Is the missing middle a services business or a platform business?
Today it looks like services, because the early jobs are specific to each company. It becomes more platform-like only if the work gets cheaper with repetition and demand broadens, which is exactly what claims 2 and 6 in the thesis map test.
How does the missing middle differ from a data marketplace?
A marketplace lists datasets a provider has already scoped, cleared and packaged; AWS Data Exchange, for instance, makes providers attest their right to distribute and keep products updated. The middle is the work before listing: turning one company's records into a product a buyer or a marketplace can accept.
What evidence would show the enterprise data licensing thesis is wrong?
Buyer counts that stay at a handful, model disclosures showing business skills trained without non-public records, a major platform launching opt-in licensing, or deal times and acceptance rates that do not improve across successive deals. Any one weakens the thesis; two together would break it.
Sources
- Slack's API terms bar applications offered outside the developer's organization from training LLMs on API data or bulk exporting message and file data without an additional agreement. Source
- OpenAI's data partnerships page seeks large-scale datasets not readily available online and excludes sensitive, personal or third-party information. Source
- AWS Data Exchange requires each data provider to attest that it has the legal right to distribute the data it publishes. Source
- AWS Data Exchange providers must have a defined customer support process and a means to keep data regularly updated. Source
- Outside the Extended Provider Program, AWS Data Exchange products may not include information that identifies a person unless it is publicly available. Source
- Epoch AI (2024) estimated the stock of quality public human text at about 300 trillion tokens and projected full use between 2026 and 2032 on trend. Source
- OpenAI's GDPval spans 1,320 tasks across 44 occupations, crafted by professionals with over 14 years of experience on average. Source
- Anthropic's January 2026 eval guidance tells agent builders to source realistic tasks from the failures they see. Source
- NVIDIA reported in June 2024 that over 98% of the data used in Nemotron-4 340B's alignment process was synthetically generated. Source
- Reddit's FY2025 Form 10-K says substantially all contract value associated with its licensing revenue comes from two partners. Source
- Scale AI announced in June 2025 a Meta investment valuing the company at over $29 billion. Source
- The European Commission's training-content summary template is mandatory for general-purpose AI model providers, with AI Office enforcement from 2 August 2026. Source
Request SourceX investor materials
SourceX shares company and commercial metrics with investors directly. Any offering is made only to verified accredited investors.