Deal economics
Historical dataset scope extension: three checks per added year
By SourceX Editorial · Reviewed by Noah Loul ·
10 sections
Short answer
Before agreeing to a historical dataset scope extension, test every year the buyer wants added on three separate questions: do the records still exist with their links intact, did the notices and contracts in force that year allow this use, and do fields, codes and policies mean what they mean today. Add only the years that pass, size the extra effort, and get the agreed window into the signed terms before anything is delivered.
License terms
Historical dataset scope
Key takeaways
- A request for more years is a scope change, and each added year needs its own evidence rather than inheriting the approval of the current window.
- Retention settings decide what survived: Salesforce keeps field history for up to 24 months through the API unless Field Audit Trail is on, and Slack's free plan deletes data older than a year.
- Older records are governed by the promises made when they were collected; FTC staff have warned that retroactive terms changes to allow AI training may be unfair or deceptive.
- A code list, workflow or policy change can make old records mean something different from new ones even when the field names match.
- The decision memo has four outcomes: extend fully, extend in bands, offer the older years later, or decline.
What should a company verify before a historical dataset scope extension?
Verify three things for each year the buyer wants to add: that the records survived with their links and history, that the notices and contracts in force when they were created allow an outside license, and that the records mean the same thing as the years already in scope. A year that fails any one of the three belongs outside the extension, or in a separate tranche with its gap stated.
The request usually arrives after a buyer has seen a sample or a dataset summary, and it sounds simple: instead of 2023 to 2026, could the window start in 2018? In practice the added years often sit on the other side of a software migration, a retention purge, a new privacy notice or a pricing policy change, and each of those changes what the company can honestly offer.
Useful evidence is dated and countable: record counts by month, a traced sample of linked records, notice and contract versions by year, and a change log of fields and codes. Memory of how things used to work is a starting point, not evidence. The worksheet later in this guide turns those checks into one page a decision owner can sign.
What are the extra years for? Questions to put to the buyer first
Ask what the older years will be used for before counting them, because the purpose decides which years and fields matter. A buyer studying seasonality needs complete annual cycles; one hunting rare failures needs the cases, not the calendar; one building a time-based test set may want the oldest years only for training.
Age has a cost for some uses. Lazaridou and colleagues, in their paper on temporal generalization in neural language models, argue that the common practice of splitting data into training and test sets at random, without regard to time, does not test how a model handles data from after its training period. A buyer asking for older records should be able to say how it will treat the distance between those years and today.
Send these questions in writing and keep the answers with the decision memo:
- Which use needs the older years: seasonality, rare cases, trend analysis, a time-based holdout or simply more volume?
- Which start date, systems and record types does the extension cover, and which fields must be present in the older years?
- Will older records face the same acceptance test as current ones, or a stated one that allows for known gaps?
- Do exclusivity, term and permitted-use clauses reach the added years in the same way as the current window?
- Would the buyer take the older years as a separate tranche, so that one weak year does not hold up the rest?
- What happens to the current scope if some of the added years fail verification?
Retained records: did the older years survive, and with their links?
Often only in part. Records outlive their context: the row that says a job happened may survive while its status history, attachments or messages were removed by a retention setting, a plan limit or a migration that carried only current records across.
Platform defaults are a common cause. Salesforce says that when Field Audit Trail is turned off, it keeps field history for up to 18 months, or up to 24 months through the API, so a CRM can hold a decade of opportunities and only two years of the changes that show how each one moved. Slack's free plan shows the last 90 days of messages, and since August 26, 2024 it has deleted messages and files more than a year old.
Older records can also move to a different store with different rules. Zendesk archives most tickets 120 days after they close; archived tickets drop out of views and reports, although search and the incremental export API still reach them. An export built from a report can therefore look complete for recent months and thin for older ones.
Finance records tend to survive longest, which can mislead. The IRS ties record keeping to the period of limitations: generally 3 years, at least 4 years for employment tax records, and 7 years when a company claims a loss from worthless securities or a bad debt deduction. Invoices from an early year may still exist because finance kept them, while the job notes that explain them were never retained.
Count records by month for every added year in the live system and in any archive, then trace a small sample end to end. A month with a sharp drop, or a sample record whose links stop at a migration date, marks the boundary of what can honestly be offered.
Which notices and contracts governed the records in each added year?
The ones in force when the records were created, not today's. Every added year needs its own answer: the privacy notice published then, the customer contract templates signed then, and any vendor or partner terms that covered the data at the time.
Updating the terms now does not reach back. In a February 2024 post, FTC staff warned that quietly amending terms or a privacy policy after the fact, so that consumer data can be put to more permissive uses such as AI training, may be an unfair or deceptive practice. Business contracts follow the same logic: a customer agreement from 2019 governs the 2019 records it covers.
Retention promises matter as well. Under the GDPR's storage limitation principle, the European Commission says personal data should be stored for the shortest time possible and that organizations should set time limits to erase or review what they hold. If a year survived only because a deletion schedule was never run, keeping it longer in order to license it is a question for counsel, not a formality.
Gather dated versions rather than current copies: archived notice pages, contract templates by effective date, and the retention schedule each year was supposed to follow. Where no dated version can be found, record the gap instead of assuming the old wording matched the new.
Context quality: do older records mean what current ones mean?
Not always, even when the field names match. A status code, category list or required field can change meaning over time, and a model trained across the change learns two definitions under one label.
The usual breaks are a migration that remapped values, a new code list that split one category into three, a workflow change that made a field mandatory, and a policy change that altered which cases reached the records at all, such as a new approval threshold or a longer returns window. Each produces records that look consistent in a spreadsheet and behave differently in training.
Ask the operations lead for a dated change log of fields, codes and policies across the whole requested window, and add a flag column marking each break. A buyer can work with a documented break. An undocumented one tends to surface later as a quality dispute, usually after delivery.
The added-year worksheet: eight checks, an owner and a flag rule
Run the worksheet once per added year, or once per band of years that share a system and a notice version. Each row has an owner who can answer from documents and a flag rule that stops a year from passing silently, and effort sits on the last row so the hours are visible before anyone commits them.
Every flag needs a written resolution: fixed, carved out, disclosed to the buyer or sent to counsel. Fill one copy of the worksheet per band:
| Check | Ask of each added year | Evidence to pull | Who answers | Flag when |
|---|---|---|---|---|
| Records exist | Do records for this year sit in a system we can still query? | Record counts by month from the live system and any archive | Systems administrator | Months with zero or sharply lower counts |
| Links and history intact | Do the records still link to customers, assets, invoices and status changes? | Twenty sample records traced end to end | Systems administrator with the operations lead | Status history, attachments or messages missing for the year |
| Retention status | Should this year still exist under our retention schedule, holds and plan settings? | Retention schedule versions, purge logs and plan settings | Records owner, with counsel on exceptions | Records kept past a stated deletion date |
| Promises in force | Which privacy notice and customer contract templates applied when these records were created? | Dated notice pages and contract templates by effective date | Legal or contracts lead | No dated version, or wording that restricts outside use |
| Third-party terms | Did a vendor or partner supply content under terms that applied that year? | Vendor agreements and data terms in force that year | Procurement with the contracts lead | Terms that bar disclosure or reuse |
| Meaning of fields and codes | Do statuses, categories and required fields mean what they mean today? | Change log of field definitions and code lists | Operations lead | A code or workflow change with no mapping |
| Policy breaks | Did a pricing, approval or service policy change which cases entered the records? | Dated policy register | Operations or finance lead | A break with no flag column in the data |
| Effort | How many internal hours does this year add, and who carries them? | Hours by role, split before and after signing | Decision owner | Hours exceed what the decision owner approved |
Illustrative: a field services company asked to start in 2017 instead of 2023
Illustrative example; the company and figures are invented. A field services company has offered work orders from 2023 to 2026 from its current dispatch system. After reviewing a sample, the buyer asks whether the window can start in 2017, which would more than double the years on offer.
The company runs the worksheet in three bands. 2021 and 2022 sit in the current system with one mapped code-list change. 2019 and 2020 were migrated in 2020 without job status history, and a pricing policy changed in 2019. For 2017 and 2018 only invoices survive, and no dated privacy notice from that period can be found.
The memo offers 2021 to 2026 in full, offers 2019 and 2020 as a separate tranche of work orders and invoices with the missing status history stated, and declines 2017 and 2018. The hours are illustrative for this example, not benchmarks. The figure maps the same bands lane by lane.
| Band | Records retained | Rights evidence | Context quality | Illustrative effort and decision |
|---|---|---|---|---|
| 2023 to 2026 (current scope) | Complete in the current system | Current notice and contract templates on file | Current codes and workflow | Already counted; keep |
| 2021 to 2022 | Complete in the current system | Dated notice and templates on file; counsel confirms | One code-list change, mapped | About 30 hours; add |
| 2019 to 2020 | Migrated in 2020 without status history | Dated notice on file; counsel confirms | Pricing policy changed in 2019; flag column added | About 70 hours; separate tranche with the gap stated |
| 2017 to 2018 | Invoices only; job notes gone | No dated notice located | Too little context to test | About 25 hours to confirm; decline |
Does a longer window mean a better offer?
No. A longer window can make a dataset more useful for some tasks and add nothing for others, and a request for more years is interest, not a commitment. Eligibility, an assessment and an indicative offer do not ensure selection, closing or any particular payment.
The strongest case against extending is that older years can drag down the whole delivery. If the 2019 records lack status history, a buyer's acceptance test may reject them, and a combined scope can fail on its weakest band. Keeping weaker years in a separate tranche protects the years that pass.
The strongest case for extending is rare work. Some tasks depend on a handful of unusual cases a year, and only a long window collects enough of them. If that is the buyer's purpose, verify those cases first rather than whole years; a narrow, verified slice of old exceptions can be more useful than complete but ordinary years. A buyer can also measure what each extra year adds with a history-depth experiment before anyone exports more.
Put the extension into the terms you would sign
Write the extension into the same terms as the original scope, line by line, before any added year is delivered. A verbal yes to more years, or a revised sample, is not a scope.
At a glance
Put the extension into the terms you would sign
When the license runs through SourceX, delivery follows the company's approval of scope and terms and its signature, so an extended window is fixed in writing before the added years move. On depth, SourceX's general guide is that buyers seek tens of thousands of records or more from one system and a history that holds steady for years, typically two or more, assessed deal by deal. For added years, consistent is the word that matters. A company that has not yet checked whether it qualifies can start with a short assessment that needs no data upload.
- Window: exact start and end dates for each system and record type, with any excluded band named.
- Fields: which fields are present in each band, and which are missing by design.
- Disclosed gaps: migrations, purges and policy breaks stated in the dataset description.
- Exclusivity and term: whether they reach the added years, and on what basis.
- Conditions and payment triggers: whether acceptance of the added years is separate from acceptance of the current scope.
- Effort: who carries the hours for each band, and which work waits until the terms are signed.
The decision memo: extend, extend in bands, offer later or decline
End with one page that a decision owner signs. It records which of four outcomes the company chose, the evidence behind it, and what would change the answer.
- Extend fully: every added year passed all eight checks, and the hours fit the approved budget.
- Extend in bands: passing years join the scope, and weaker years go in a separate tranche with their gaps stated.
- Offer later: the years may pass, but evidence is missing; name the document, its owner and a date.
- Decline: the records are gone, the rights cannot be documented, or the context is too thin to test.
- Limits: what was not checked, such as attachments in a legacy archive, and which questions went to counsel.
- Next evidence: the one count or document that would move a deferred or declined band into scope.
Frequently asked questions
How many years of history do AI data buyers usually want?
There is no fixed number. As a general guide from SourceX, buyers expect a single system to show consistent records over a run of years, typically at least two, and to hold tens of thousands of records or more, while smaller sets can qualify when the work is rare. Consistency matters more than length: four clean years can be more useful than nine years with a migration gap in the middle.
Can old records be licensed after the privacy notice has changed?
Older records are generally governed by the notice and contracts in force when they were collected, not by today's versions. FTC staff warned in 2024 that adopting more permissive practices such as AI training through a retroactive change to terms may be unfair or deceptive. Collect the dated versions for each year and have counsel review whether they allow the proposed use.
Why does a CRM keep old records but not their change history?
Many systems keep records and their history under different rules. Salesforce, for example, keeps field history for up to 18 months, or 24 months through the API, unless Field Audit Trail is turned on, so older opportunities may carry no record of how they changed. Check history and attachments separately from the main records.
Should weaker older years go in a separate tranche?
Often, yes. A separate tranche lets the buyer accept or reject the older years on their own evidence, so a gap in one band does not hold up the years that pass. State the gap in the dataset description and agree in writing how acceptance works for each tranche.
Is a buyer's request for more years an offer?
No. It is a request to change scope, and nothing binds until the company approves the scope and terms and signs. Treat the extra verification hours as effort at risk and approve them in stages, starting with the counts and dated documents that are cheapest to produce.
Sources
- Salesforce keeps field history for up to 18 months, or up to 24 months through the API, when Field Audit Trail is turned off. Source
- Free Slack workspaces can view and search the last 90 days of messages, and since August 26, 2024 messages and files older than one year are deleted. Source
- Zendesk archives tickets about 120 days after they close; archived tickets leave views and reports but remain reachable through search and the incremental export API. Source
- The IRS ties record retention to the period of limitations: generally 3 years, at least 4 years for employment tax records, and 7 years for worthless securities or bad debt claims. Source
- FTC staff warned in February 2024 that adopting more permissive data practices such as AI training through a surreptitious, retroactive change to terms or a privacy policy may be unfair or deceptive. Source
- The European Commission's GDPR guidance says personal data should be stored for the shortest time possible, with time limits set to erase or review it. Source
- Lazaridou et al. argue that time-agnostic random splits of training and test data do not test how language models handle data from after their training period. Source
See if your company qualifies
A short company assessment. No data uploads are needed.