AI data market
Business archives as an AI investment: five tests for which ones have buyers
By SourceX Editorial · Reviewed by Noah Loul ·
9 sections
Short answer
Business archives deserve an AI investment thesis only when they preserve what public data lacks: decisions with their context, the outcome that followed and evidence of who may license them. Model developers now seek non-public data and score models on real work, so the demand side is real. Most archives still fail on rights evidence, freshness or missing outcomes, so score all five readiness dimensions before assigning value.
System export
Business archives AI
Key takeaways
- An archive's value rests on what survives: decisions, context, outcomes and rights evidence, not volume.
- Demand for non-public records of real work exists, but it is task-specific and concentrated.
- Judgment and procedure tasks reward history; fact and price tasks punish it.
- Score five readiness dimensions and let the weakest one, not the average, set the next action.
- Time the reconstruction and review work, then test a small sample against public and synthetic baselines before assigning value.
Do business archives deserve their own AI investment thesis?
Yes, but a narrow one. Business archives earn an AI investment thesis when they preserve what public data cannot supply, such as decisions with their context, the outcome that followed and documents showing who may license them, for tasks where history stays valid. Most archives fail at least one of those tests, so the thesis is about selecting archives, not owning them.
The demand side has moved toward archives. OpenAI's data partnerships page says it is looking for large-scale data that the public cannot easily reach online today. Its GDPval evaluation, published on September 25, 2025, scores models on 1,320 tasks from 44 occupations in the top nine industries contributing to U.S. GDP, written by professionals with an average of 14 years of experience. Model developers want examples of real work, and real work is mostly stored in company systems.
The supply side is where investors get hurt. A large stock of records can coexist with weak purchasing demand, and an archive that looks impressive by size can fail a buyer's first sample. This article inventories what survives in an archive, separates tasks that reward history from tasks that punish it, and scores readiness on five dimensions before anyone assigns value.
What survives in a business archive, and what quietly disappears
Source records usually survive; the context that makes them useful often does not. An archive is the residue of operations, retention rules and system migrations, so its contents reflect what someone was required, or remembered, to keep.
Retention law explains much of what survives. The IRS's current guidance tells businesses to keep most tax records for 3 years, employment tax records for at least 4 years, and records for some situations for 6 or 7 years. Those rules preserve invoices, payroll and ledgers well; they do nothing for the ticket comments, approval notes and policy versions that show how decisions were made.
System defaults decide much of the rest. Salesforce's current documentation says that when Field Audit Trail is off, field history is retained for up to 18 months, and up to 24 months through the API. An eight-year CRM archive can therefore hold the final value of every deal but little record of how stages, amounts or owners changed along the way.
Inventory five layers before anything else, because each one answers a different buyer question:
- Source records: the tickets, orders, jobs, claims or messages themselves, with stable IDs.
- Decision context: who decided, with what information in front of them, under which approval path.
- Policy versions: the price book, rulebook, tariff or SOP in force when each decision was made.
- Outcome labels: what happened afterward, such as a reopen, a refund, a callback, or a claim paid or denied.
- Ownership evidence: contracts, notices and assignments showing the company may license each layer.
Which buyer tasks reward history, and which punish it?
Tasks built on judgment and procedure reward history; tasks built on current facts punish it. A buyer requirements matrix makes the split explicit before anyone prices an archive.
The decay is measurable. In Mind the Gap (NeurIPS 2021), language models performed up to 16 percent worse at predicting articles published up to two years after their training period ended, and a smaller model trained on more recent data outperformed a model 60 percent larger that lacked it. Archives that mostly record facts, prices or product details lose value with every year of age; archives that record how people reasoned through cases age far more slowly.
The matrix also exposes a common error. A ten-year archive is not ten years of usable data for a procedure task if the process changed three years ago. It is three years of current practice and seven years of a different process, which may still help as contrast but should not be priced as training on how the work is done now.
| Buyer task | What the archive must show | Freshness sensitivity | Archive verdict |
|---|---|---|---|
| Evaluating judgment, such as claim triage or escalation calls | The situation, the decision and a verified outcome | Low: a good call stays a good call | Strong fit when outcomes are recorded |
| Training agents on procedures, such as order exceptions | Steps, tools and approvals in sequence | Medium: valid while the process is unchanged | Fit for the period after the last process change |
| Applying policies or rules | The rule version in force for each case | Medium: depends on whether rule versions survive | Fit only when policy versions are kept |
| Product or price knowledge | Current catalog, prices and specifications | High: wrong as soon as it is superseded | Weak; stale facts teach errors |
| Forecasting and demand planning | Long, consistent time series | Mixed: history is the input, but regime shifts break it | Fit when breaks are documented |
The Business Archive Opportunity Map: five readiness dimensions
Score each archive from 0 to 3 on five dimensions and let the weakest score decide what happens next. An archive with strong records and no rights evidence is not mostly ready; it is blocked until its weakest dimension is fixed.
Gating rules: a 0 or 1 on survival or rights evidence stops commercial work until it is fixed; a 0 or 1 on decision context or outcome labels narrows the archive to the subset that scores 2 or more; task durability decides which buyer requests the archive can answer at all. The illustrative scores in the last column, explained in the next section, stop at rights evidence before anyone estimates value.
Methods note: the five dimensions, the 0 to 3 scale and the gating rules form an editorial framework assembled for this article on 2026-10-26 from the sources cited here, including IRS retention guidance, Salesforce documentation, the Mind the Gap study, SWE-bench and OpenAI's published data and evaluation pages. The scores are illustrative. Not measured: how often real archives reach each score, what buyers pay, or how scores relate to acceptance.
| Dimension | What to test | Scores 3 when | Scores 0 or 1 when | Illustrative score |
|---|---|---|---|---|
| 1. Survival | How much of the claimed period and systems still exists and exports | Continuous history with stable IDs across migrations | Gaps, purges or a migration that dropped closed records | 3 |
| 2. Decision context | Whether each record shows who decided, with what information and approval | Comments, approvals and attachments are linked to the record | Only final field values remain | 2 |
| 3. Outcome labels | Whether what happened next is recorded and checkable | Reopens, refunds or results are recorded per case | Outcomes live only in people's memory | 3 |
| 4. Rights evidence | Contracts, notices and signing authority that permit licensing | Documents reviewed and a signer identified | Customer contracts unreviewed or third-party content mixed in | 1 |
| 5. Task durability | Fit to a named buyer task in the requirements matrix | Judgment or procedure task with a stable process | Fact or price task, or a recent process change | 2 |
A worked, permissioned archive example (Illustrative)
Illustrative: a fictional equipment rental company with about 300 employees has nine years of rental contracts, return inspections and damage disputes in its rental management system, with dispute correspondence in email and invoices in its accounting system. The owner has authorized a readiness review; no data has left the company.
Survival scores 3: rental and inspection records export with stable contract IDs for all nine years, and invoices join on contract number. Decision context scores 2: inspectors' notes and photos are attached, but the reasoning behind reduced or waived damage charges sits in email threads that link to contracts only for the last four years. Outcome labels score 3: every dispute ends in a recorded charge, reduction or waiver, which gives each decision a checkable outcome.
Rights evidence scores 1: about a fifth of rentals fall under national account agreements with confidentiality clauses nobody has reviewed, and some inspection photos show customers' job sites and staff. Task durability scores 2: a developer building a damage-assessment agent would value the judgment in the disputes, but the company replaced its damage rate card five years ago, so charge amounts only fit the recent period.
Decision: stop commercial work until counsel reviews the national account agreements, exclude photos that show people or identifiable sites, and scope any offer to the four years with linked dispute reasoning. The archive is promising, not yet valuable, and its weakest dimension sets the next action.
Price the reconstruction and review effort before the archive
Estimate the work needed to make an archive usable before estimating what it is worth, because reconstruction and review effort, not record count, decides whether the thesis survives. The estimate should come from a timed sample, not a rule of thumb.
Public benchmarks show how hard the filter can be. SWE-bench (ICLR 2024) started from about 90,000 pull requests across 12 popular Python repositories and kept 2,294 task instances, roughly 2.5 percent, after requiring each to resolve a linked issue, change test files and survive further filtering. That archive was public, well linked and written by engineers; a business archive with partial context should not assume less attrition for demanding tasks.
Run the estimate in four steps, then treat the result as the archive's cost of goods. If the timed effort per usable record exceeds what any named buyer task could justify, the archive fails the thesis whatever its size.
- Draw a random sample of 50 to 100 records by a written rule across the whole date range, not hand-picked cases.
- Time the context reconstruction for each record: finding linked approvals, attaching the policy version and confirming the outcome.
- Time the review: privacy redaction, third-party content checks and a second reviewer on a subset.
- Scale the per-record times to the in-scope population, report the range across reviewers, and add rights review as a fixed block rather than a per-record cost.
Test a small sample against a public or synthetic baseline
Before committing to a whole archive, test whether a small prepared sample beats what a buyer could get without it on one defined business task. If it does not, the archive's premium is not real.
At a glance
Test a small sample against a public or synthetic baseline
The baseline must be credible, because substitutes can be strong. Microsoft researchers trained phi-1, a 1.3 billion parameter code model, on filtered web data plus about 1 billion tokens of textbooks and exercises generated by GPT-3.5, and reported 50.6 percent pass@1 on HumanEval in 2023. Benchmark builders also simulate business settings directly: tau-bench emulates conversations between a user simulated by language models and an agent given domain-specific API tools and policy guidelines.
Fix the design before any result is seen. SourceX has not published results from a test like this; the design below is how an investor should ask any archive-backed company to show its evidence.
- Task: one named task with a written acceptance test, such as deciding whether a disputed damage charge should stand.
- Arms: the archive-derived sample, a public-data baseline and a synthetic baseline built to the same specification.
- Metric: the buyer's own acceptance metric, scored on held-out cases drawn from the archive's most recent period.
- Uncertainty: enough held-out cases to report a confidence interval, published beside the point estimate.
- Decision rule: proceed only if the archive arm beats both baselines by a margin set before results are seen.
When the archive thesis fails
The thesis fails when buyers do not need the history, cannot lawfully take it, or can get a substitute more cheaply. Each failure mode has public evidence an investor can monitor.
- Concentrated demand. Two partners account for substantially all of the contract value in Reddit's licensing revenue, according to its 2024 annual report; a market with few buyers can leave good archives unsold.
- Limits on keeping and reusing personal data. In Digi (C-77/21, 20 October 2022), the EU Court of Justice held that copying customer data into a test database can be compatible with the original purpose, but that storage limitation bars keeping it longer than the new purpose requires. Archives holding personal data cannot be stockpiled on speculation.
- Freshness decay. Mind the Gap showed models degrading on text from after their training period, so archives of facts and prices lose value faster than archives of judgment.
- Synthetic substitutes. phi-1 and tau-bench show that generated data and simulated environments can stand in for real records on some tasks; where they do, the archive premium disappears.
- Missing rights evidence. If the ownership layer cannot be documented, the other four dimensions do not matter, and the archive is storage.
Diligence questions for archive-backed data companies, SourceX included
Ask every archive-backed company to show its scoring and its evidence, not its record counts. SourceX should answer the same questions.
SourceX's own role is to help companies turn the business records they select into licensed AI datasets; it takes on sourcing and qualification, the rights review, contracting, data preparation, delivery and payment. Its general guide on size lines up with the survival test: buyers tend to want years of consistent history, usually two or more, and at least tens of thousands of records held in a single system. A smaller archive can still qualify when the work it records is rare, and every case is judged deal by deal.
Readiness scores, acceptance results and commercial metrics for SourceX are not public; the company discloses company and commercial metrics to investors directly, and its current raise, under Rule 506(c), is open to accredited investors. The investor overview is available on request, and the questions below are a fair place to start.
- How many archives did you score in the last year, and what share stopped at survival or rights evidence?
- For accepted datasets, which buyer tasks did they serve, and how recent did the records need to be?
- How many reviewer hours did reconstruction and review take per usable record, as a range across reviewers?
- Which archive-derived samples beat a public or synthetic baseline on a buyer's own metric, by what margin and with what interval?
- How many distinct buyers have issued written requests for archive-derived data?
Frequently asked questions
Can old business records be valuable for AI training?
They can, when they record decisions with their context and outcomes for a task where history stays valid, and when the company can document its right to license them. Records of facts and prices age quickly, while records of judgment and procedure age slowly. Volume alone does not make an archive valuable.
How old is too old for AI training data from an archive?
It depends on the task. Research on temporal generalization found that models degrade on text from after their training period, so fact-heavy archives lose value quickly while judgment tasks age far more slowly. For procedure tasks, the useful cut-off is usually the last process change rather than a fixed number of years.
Does holding an archive mean a company may license it for AI?
No. Customer contracts, privacy notices, platform terms, employee and contractor agreements and third-party content all affect what may be licensed, so a rights review should come before any valuation. This is general information; counsel should review the specific documents.
What should an investor ask an archive-backed data company?
Ask how archives are scored, what share stop at survival or rights evidence, how many reviewer hours reconstruction and review took per usable record, and whether archive samples beat public or synthetic baselines on a buyer's own metric. Ask how many distinct buyers have issued written requests.
Does SourceX publish archive study results?
No. Archive readiness scores and acceptance results are unpublished, and SourceX reserves company and commercial metrics for direct conversations with investors. Investors who want to test an archive claim can ask for materials at sourcex.si/investors; SourceX makes any offering only to verified accredited investors.
Sources
- OpenAI's data partnerships page seeks large-scale datasets that reflect human society and are not already easily accessible online to the public. Source
- OpenAI's GDPval (September 25, 2025) covers 44 occupations in the top nine industries contributing to U.S. GDP, with 1,320 tasks crafted by professionals averaging over 14 years of experience. Source
- The IRS advises keeping most tax records for 3 years, employment tax records for at least 4 years, and records for certain situations for 6 or 7 years. Source
- Salesforce says that when Field Audit Trail is turned off, field history data is retained for up to 18 months, and up to 24 months via the API. Source
- Mind the Gap (NeurIPS 2021) found models up to 16 percent worse on articles published up to two years after the training period, and a smaller model with recent data outperforming a 60 percent larger one. Source
- SWE-bench filtered about 90,000 pull requests from 12 popular Python repositories down to 2,294 task instances, each tied to an issue and a change to test files. Source
- phi-1, a 1.3 billion parameter code model trained on filtered web data and about 1 billion tokens of GPT-3.5-generated textbooks and exercises, reached 50.6 percent pass@1 on HumanEval. Source
- tau-bench emulates dynamic conversations between a user simulated by language models and a language agent given domain-specific API tools and policy guidelines. Source
- Reddit's annual report for 2024 says substantially all contract value associated with its licensing revenue is derived from two partners. Source
- In Digi (C-77/21, 20 October 2022), the Court of Justice held that storing customer data in a test database can be compatible with the original purpose, but storage limitation bars keeping it longer than the tests require. Source
Request SourceX investor materials
SourceX shares company and commercial metrics with investors directly. Any offering is made only to verified accredited investors.