AI data market
Business archive data preservation economics: three clocks, four futures
By SourceX Editorial · Reviewed by Noah Loul ·
9 sections
Short answer
Business archive data preservation economics turn on three clocks, not storage prices. Context fades as change history is trimmed and staff move on, access ends when subscriptions lapse, and rights evidence can be documented but never upgraded later. Preservation is cheap while all three still run; once one stops, no budget buys the archive back. Investors should ask which clock binds first and how a supplier beats it.
Value drivers
Business archive data
Key takeaways
- An archive is useful for AI only when the records, the context to read them and evidence that someone may license them survive together.
- Documented timers are short: Salesforce keeps field history for up to 18 months in the org and 24 through the API unless Field Audit Trail is on.
- Rights evidence runs one way: FTC staff warned in February 2024 that quietly applying more permissive data uses retroactively may be unfair or deceptive.
- The investment case lives in one of four futures, the scarcity premium, where demand for non-public records holds while most archives decay unpreserved.
- Synthetic data, better context reconstruction and concentrated buyers could each shrink that premium, so the scorecard sets a dated signpost for each.
Does archive decay change the economics of business data for AI?
Yes, because an archive's value for AI depends on three things surviving together, and each fails on its own timetable. The records must still exist and be reachable, the context that explains them must still be recoverable, and someone must still be able to show that the company may license them. Lose any one and the other two are worth far less to a model developer.
That makes business archive data preservation economics a timing problem rather than a storage problem. Keeping bytes is cheap. Keeping a ticket history with its status changes, the person who knows what a custom field meant and the customer terms in force when the ticket was opened is not, and the cost of each rises the longer it waits.
Demand is the other half. Epoch AI's 2024 study put the effective stock of quality public human text at about 300 trillion tokens and expected it to be fully used at some point between 2026 and 2032 on current trends. OpenAI says it relies mostly on public data and adds proprietary data through partnerships to reach non-public archives and metadata, and that it does not pursue paid deals for purely public information. If developers keep looking past the public web, archives that survive intact become scarcer than archives that merely exist.
What follows sets out the three decay clocks, the asymmetry that makes early preservation cheap and late preservation impossible, four futures for the market, a scorecard of dated calls an investor can grade, and the evidence that would make the thesis wrong.
Three clocks that run down a business archive
Three clocks set an archive's deadline, context, access and rights, and the shortest one sets the deadline for everything else. Each has a documented timer in ordinary business systems.
The context clock runs inside the systems themselves. Salesforce documents that, with Field Audit Trail off, it retains field history for up to 18 months in the org and up to 24 months through the API, while Field Audit Trail keeps history until someone deletes it. A record whose current values survive but whose change history does not loses the sequence of decisions a workflow model learns from. People carry the rest of the context: the Bureau of Labor Statistics reported median tenure with a current employer of 4.1 years in January 2026, so the staff who know why a field was repurposed are often gone before a buyer asks.
The access clock starts when a subscription or a plan changes. Microsoft's documentation for Microsoft 365 business subscriptions describes an expired stage, usually 30 days, and a disabled stage, usually 90 days, after which the data is deleted and buying a new subscription does not bring it back. Slack's free workspaces have deleted messages and files older than one year since August 2024, so a plan limit can erase history while the account is still active.
The rights clock behaves differently: it never resets in the company's favor. FTC technology staff wrote in February 2024 that adopting more permissive data practices, such as using data for AI training, and disclosing the change only through a surreptitious, retroactive amendment to terms or a privacy policy may be unfair or deceptive. Records keep the promises made when they were collected, so the notices, contracts and consents in force at the time have to be preserved with the records, or the rights question can never be closed.
The ledger below gives one row per clock, with a dated timer, what disappears when it stops and what a preserved archive has to carry:
| Clock | Documented timer, dated | What disappears | What a preserved archive must carry |
|---|---|---|---|
| Context: change history | Salesforce field history: up to 18 months in the org and 24 via the API unless Field Audit Trail is on (help pages, checked October 2026) | The order of status, owner and value changes behind each outcome | Field history or audit-trail exports, plus a data dictionary for custom fields |
| Context: people | Median employee tenure of 4.1 years in January 2026 (BLS, September 2026 release) | Knowledge of why fields, codes and workflows changed | Dated notes from system owners, tied to schema versions |
| Access: subscriptions | Microsoft 365 business data deleted after an expired stage (usually 30 days) and a disabled stage (usually 90 days) | The records themselves, with attachments and permissions | A verified export with counts reconciled against the live system |
| Access: plan limits | Slack free workspaces delete messages and files older than one year, since August 2024 | Older conversation history and the files shared in it | Exports taken before limits apply, with the plan history recorded |
| Rights: collection terms | FTC staff, February 2024: quiet retroactive expansion of data use may be unfair or deceptive | The ability to show what customers and staff were told at collection | Dated copies of notices, customer contracts and consents, mapped to record date ranges |
Why early preservation is cheap and late preservation is impossible
Preservation costs are small, known and paid up front, while the loss from waiting is uncertain, back-loaded and, once a clock stops, permanent. That asymmetry, not the storage bill, is the core of the economics.
At a glance
Why early preservation is cheap and late preservation is impossible
Before a clock stops, preserving an archive means an export, a completeness check, storage, a data dictionary and a folder of dated rights documents, all ordinary work an IT team and counsel can scope. After the access clock stops, the records are gone and no price recovers them. After the context clock stops, the records survive but need reconstruction through interviews and inference whose accuracy nobody can check. After the rights clock stops, the records may be readable and still unlicensable.
Public web data shows what decay looks like when nobody owns preservation. Pew Research Center sampled pages from the Common Crawl archive and found that 38 percent of webpages that existed in 2013 were no longer accessible by October 2023, and about a quarter of all pages seen between 2013 and 2023 were gone. Business systems are not the web, but steady loss unless someone acts is the default for any digital record without an owner.
The choice a company, or its investor, faces works like an option. Preserve when the expected value of a future use, scaled by the share of the archive that would still be usable, exceeds the cost of keeping it, including any duty to delete. Each input below is an editable assumption for the company to fill with its own figures, not a measured value:
- Probability of a qualifying use within the review horizon, estimated separately for licensing, internal model building, audit and litigation.
- Usable share if preserved now: the fraction of records that keep their history, context notes and rights documents through the export.
- Usable share if preserved later: the same fraction after the next clock stops, which is zero for anything past an access deadline.
- Carrying cost per year: storage, access licenses, periodic reviews and the staff time to keep the data dictionary current.
- Deletion duty: any date after which privacy law, a contract or the retention schedule requires records to go, which caps the horizon.
Four futures for preserved business archives
Two uncertainties decide which future arrives: whether companies preserve archives with context and rights evidence before the clocks stop, and whether model developers keep paying for non-public business records. Crossing them gives four futures, and only one rewards early preservation with a premium.
The scarcity premium is the thesis case: demand for non-public records holds, most companies let their clocks run out, and the archives that survive intact are few and valuable. The abundant archive is its mirror, where vendors, regulators or cheap tooling make preservation the default, supply is ample, and value moves to whoever resolves rights and prepares records fastest. Quiet loss and stranded storage are the weak-demand futures; in one, archives vanish and nobody misses them, and in the other, companies pay to keep records no buyer wants.
One regulation already pushes cloud data in the EU toward the abundant archive. The European Commission says the Data Act has applied since 12 September 2025 and that switching charges, including data egress charges, will be removed entirely from 12 January 2027. Cheaper exits slow the access clock for companies leaving a provider; they do nothing for the context or rights clocks. The matrix sets out each future with the evidence that would show it arriving:
| Future | Preservation | Demand for non-public records | Who captures value | Evidence it is arriving |
|---|---|---|---|---|
| Scarcity premium | Low: most clocks run out | Strong | Companies that preserved early, and suppliers that can prove intact context and rights | Developers keep naming non-public archives as a source; few suppliers can show change history |
| Abundant archive | High: preservation becomes the default | Strong | Rights resolution and preparation rather than possession | Longer default retention, cheaper exits and many suppliers with complete exports |
| Quiet loss | Low | Weak | No one; archives vanish unnoticed | Synthetic data covers workflow tasks; developer disclosures stop mentioning archives |
| Stranded storage | High | Weak | Storage and compliance vendors | Companies hold complete archives but report no outside use |
The decay scorecard: six dated calls an investor can grade
These are falsifiable calls, each tied to a public signpost and a date by which it can be graded. They are editorial forecasts made on 26 October 2026, not measurements, and together they show which future the evidence is moving toward.
Read them as a set. If most hold, the market is drifting toward the scarcity premium. If the synthetic data and public-text calls go the other way, quiet loss or the abundant archive becomes more likely, and paying extra for preserved archives looks weaker. The scorecard, open on every row as of publication:
| Call, made 26 October 2026 | Signpost to check | Grade by | What would count against it |
|---|---|---|---|
| EU switching and egress charges end on schedule under the Data Act | European Commission Data Act pages and any amendment to the switching rules | 12 January 2027 | A delay, or an exemption that keeps exit charges for existing contracts |
| Major workplace vendors keep default history and post-termination windows short | Salesforce field history, Microsoft 365 lifecycle and Slack free-plan help pages | 31 December 2027 | A vendor extending default history retention or the export window after termination |
| Model developers keep citing non-public archives as a training source | Developers' data pages, model cards and system cards | 31 December 2027 | Disclosures that drop partnership or archive data in favor of public and synthetic sources |
| Synthetic data does not replace real workflow records in post-training | Open model documentation such as NVIDIA's Nemotron FAQ and technical reports | 31 December 2027 | Published ablations where synthetic workflow data matches real records on business tasks |
| Regulators act on retroactive data-use changes made for AI training | FTC enforcement announcements and staff guidance | 31 December 2028 | No action, or guidance that lets companies apply new AI uses to records collected under old terms |
| Estimates of public-text exhaustion do not move later | Epoch AI's data stock projections | 31 December 2027 | A revised estimate that pushes full use of public text well past 2032 |
Could synthetic data, better models or a few buyers make decay irrelevant?
Each could shrink the premium, and each has public evidence behind it, so the thesis has to beat them rather than assume them away.
Synthetic data is the strongest substitute. NVIDIA's Nemotron training FAQ says the models use synthetic data in post-training, that the pre-training corpus also includes synthetic and rephrased sources, and that synthetic data is validated through small-scale ablations. If generated workflows pass a buyer's acceptance tests for tasks that once needed real records, decayed archives stop mattering. The same FAQ cuts the other way on quality: asked whether dataset, architecture or method matters most, it answers the dataset, and says a strong dataset can overcome limits in both.
Better models could rebuild lost context, inferring what a repurposed field meant or reconstructing a status sequence from surrounding text. That helps only with the context clock. It cannot restore records past an access deadline, and an inferred history is a weaker training signal than a logged one because nobody can check it against what happened.
Concentrated demand is the third alternative. Reddit disclosed in its quarterly report for the period ended June 30, 2025 that substantially all of the contract value associated with its licensing revenue came from two partners. If a handful of developers set the price for non-public content, a preserved archive's value depends on their appetite more than on its scarcity.
None of the three makes preservation pointless. They change who should pay for it: if any of them wins, a company preserves for audit, legal and internal AI uses, and outside licensing becomes an upside rather than the reason.
Risks that would break the preservation thesis
Five developments could make preserved archives worth less than the thesis assumes. A company should be able to name the one it is most exposed to before describing its upside.
- Deletion duties cap the horizon. The GDPR's storage limitation principle requires personal data to be kept in identifiable form no longer than its purposes need, with longer storage for archiving, research or statistics only under safeguards, so preserving everything indefinitely is not a lawful default.
- Rights can tighten as well as fade. Customer contracts renewed with AI-use restrictions can narrow what a company may license from records created under them, which is why rights documents must be mapped to record date ranges.
- Synthetic substitution. If generated data matches real records on business acceptance tests, the premium falls for every task it covers.
- Concentrated buyers. A few large developers can hold prices down however scarce intact archives become, as the two-partner concentration in Reddit's licensing revenue shows for public content.
- Cheaper preservation. Rules and tools that make complete exports routine, such as the Data Act's end of switching charges in the EU, would move the market toward the abundant archive and shift value from possession to preparation.
Diligence questions on archive survival, for any supplier including SourceX
Ask which clock binds for the archives a company counts as supply, and ask for evidence measured against that clock rather than record counts; hold SourceX to exactly the same standard.
On the supply side, SourceX takes companies through qualification, rights review, contracting, data preparation, delivery and payment when they license selected operational records to AI developers, and the company approves scope and terms and signs before any data reaches SourceX. That order puts the rights clock first: rights documents are reviewed before records move.
SourceX has not published decay or survival measurements; company and commercial metrics go to investors directly rather than into articles. Its current raise, under Rule 506(c), is open only to accredited investors, and the investor page takes requests for materials, including any of the questions below.
- For each archive counted as supply, which clock stops first, and on what date?
- What share of offered archives include field or audit history, and how far back does it run?
- Who wrote the data dictionary for each system, and are they still with the company?
- Which notices, customer contracts and consents applied to each record date range, and where are the dated copies?
- How many archives were lost to an access deadline before review finished, and what changed afterward?
- Which call in the scorecard would hurt the business most if it went the other way?
How the clocks, futures and calls were assembled
The three-clock framework, the option inputs, the scenario matrix and the six calls are editorial tools built for this article on 26 October 2026. Timers come from vendor documentation checked that day (Salesforce, Microsoft and Slack help pages), the BLS tenure release of September 2026 and the FTC's February 2024 staff post. Context comes from Pew Research Center's May 2024 study, the European Commission's Data Act pages, Epoch AI's 2024 data study, OpenAI's data approach page, NVIDIA's Nemotron FAQ, Reddit's 2025 quarterly report and the GDPR text.
Not measured: how fast real business archives lose context, access or rights evidence, how often archives are lost before review, what share of companies preserve history, and whether preserved archives sell at any premium. Each needs company records that are not public, and vendor timers change, so repeat the dated checks before relying on them.
Frequently asked questions
How long do software vendors keep business data after a subscription ends?
It varies by vendor and plan, so read the current documentation and contract for each system. Microsoft, for example, describes an expired stage of usually 30 days and a disabled stage of usually 90 days for Microsoft 365 business subscriptions before data is deleted. Plan limits can delete data even while an account is active, as Slack's free workspaces do for content older than one year.
Does an old business archive still have value for AI?
It can, if the records arrive with their history, the context needed to read them and evidence that the company may license them. Age alone is not the problem: an old archive with complete change history and dated rights documents can be more useful than a recent one without them. An archive missing all three is mostly storage.
Can AI rebuild the context an archive has lost?
Partly. A model can infer what a field probably meant or reconstruct a sequence from surrounding text, but nobody can check an inferred history against what actually happened. Reconstruction never recovers records deleted after an access deadline, and it does not create rights.
Can a company fix rights for old records by updating its privacy policy?
Not by quietly changing it. FTC staff wrote in February 2024 that adopting more permissive practices, such as using data for AI training, and disclosing the change only through a surreptitious, retroactive amendment may be unfair or deceptive. Old records generally carry the commitments made when they were collected, so counsel should review the dated notices and contracts that applied.
Has SourceX published measurements of archive decay?
No. This article publishes a framework and editorial forecast calls, not results. Company and commercial metrics are shared with investors directly, and investor materials are available on request through the investor page.
Sources
- Epoch AI estimated the effective stock of quality public human text at about 300 trillion tokens and projected full use at some point between 2026 and 2032 on current trends. Source
- OpenAI says it relies mostly on public data, adds proprietary data from partnerships to reach non-public archives and metadata, and does not pursue paid deals for purely public information. Source
- With Field Audit Trail off, Salesforce retains field history for up to 18 months in the org and up to 24 months via the API; with Field Audit Trail on, history is kept until deleted. Source
- The Bureau of Labor Statistics reported median tenure with a current employer of 4.1 years in January 2026. Source
- Microsoft 365 business subscriptions pass through an expired stage (usually 30 days) and a disabled stage (usually 90 days) before data is deleted, and a new subscription does not restore deleted data. Source
- Slack free workspaces delete messages and files older than one year, a policy in place since August 2024. Source
- FTC staff wrote in February 2024 that adopting more permissive data practices, such as AI training, and disclosing them only through a surreptitious, retroactive amendment to terms or a privacy policy may be unfair or deceptive. Source
- Pew Research Center found that 38 percent of webpages that existed in 2013 were no longer accessible by October 2023, and about a quarter of pages seen between 2013 and 2023 were gone. Source
- The European Commission says the Data Act has applied since 12 September 2025 and that switching charges, including data egress charges, will be removed entirely from 12 January 2027. Source
- NVIDIA's Nemotron FAQ says synthetic data is used in post-training and pre-training includes synthetic and rephrased sources, validated by small-scale ablations, and names the dataset as what matters most. Source
- Reddit's quarterly report for the period ended June 30, 2025 says substantially all of the contract value associated with its licensing revenue is derived from two partners. Source
- The GDPR's storage limitation principle requires personal data to be kept in identifiable form no longer than necessary for its purposes, with longer storage for archiving, research or statistics subject to safeguards. Source
Request SourceX investor materials
SourceX shares company and commercial metrics with investors directly. Any offering is made only to verified accredited investors.