Skip to content

Leadership and readiness

Business data fit without uploads: what a no-upload assessment shows

By SourceX Editorial · Reviewed by Noah Loul ·

10 sections

Short answer

Without uploading records, a company can learn whether it is eligible, which record families match AI uses, its history and volume in bands, where rights questions sit and how much internal work preparation would take. That is enough to decide whether to fund a scoped investigation. Data quality, exact de-identification effort and value still need a reviewed sample later.

Checklist

Business data fit

This page covers business data fit without uploads: what a no-upload assessment shows.Illustrative

Key takeaways

  • A no-upload assessment answers whether to fund a scoped investigation, not what the data is worth.
  • Eligibility, record families, history and volume bands, linkage and workload can all be settled from metadata.
  • Rights triage works from contracts and privacy notices and keeps access, privacy, ownership and licensing separate.
  • Quality, exact de-identification effort and value need a reviewed sample later, under agreed terms.
  • Name one decision owner, a fixed internal time budget and the conditions that end the exercise.

What can a company learn about a data deal without uploading records?

Without uploading a single record, a company can learn whether it is eligible, which of its record families match what AI developers use, roughly how much history and volume it holds, where rights questions sit, how much internal work preparation would take, and therefore whether a scoped investigation is worth funding. That is enough for leadership to make a go, narrow or stop decision on the next step.

What it cannot learn without records is the final answer on quality, the exact de-identification effort, or what a specific buyer will conclude after inspecting a sample. A no-upload assessment is a decision about whether to spend effort, not a valuation.

This page gives leaders a metadata-only discovery and decision checklist: what to collect, who owns each item, what each answer lets you decide, and where the evidence runs out. It complements guidance on which metadata is safe to share; the focus here is what leadership can decide from it.

What can leadership decide now, and what needs more evidence?

Leadership can decide now whether to open a scoped investigation and who owns it; it should not decide terms, scope of records or delivery until rights and a sample review are done. The table separates the questions by the evidence they need.

What can leadership decide now, and what needs more evidence?
QuestionAnswerable from metadata?What settles itDecision owner
Is the company eligible for the program?YesHeadcount history, country, language of recordsCEO or owner
Do our record families match AI uses?MostlySystem inventory mapped to use typesOperations or data lead
Is there enough history and volume?Yes, in bandsDate ranges and record counts per systemIT or data lead
Do we have the right to license?PartlyContract, privacy notice and ownership review by counselGeneral counsel or outside counsel
How hard is preparation?RoughlyExport method, linkage keys, share of free textIT lead
Is the data good enough?NoProfiling or a reviewed sample under agreed termsData lead with buyer input
What would it be worth?NoBuyer fit, rights, quality and terms togetherCEO with finance

The metadata-only discovery checklist

Work through the checklist in order, because early items can end the exercise cheaply. Each item is answered from system settings, reports, counts and documents, never from exporting or sharing records themselves.

The structure borrows from established dataset documentation practice. Datasheets for Datasets asks creators to document motivation, composition, collection process and recommended uses; Google's Data Cards Playbook describes Data Cards as structured summaries of essential facts about a dataset's sources, collection and intended use; and Hugging Face dataset cards carry metadata such as license, language and size. MLCommons' Croissant format puts dataset metadata, resource descriptions and structure into one machine-readable file. You do not need any of these formats, but their questions are a good test of whether your answers are complete.

Checklist · 11 items

  1. Company fit: country, peak full-time headcount, and the language of the records you would offer.
  2. System inventory: every system of record (for example Salesforce, HubSpot, NetSuite, ServiceTitan, Zendesk, Jira), with owner, years in use and whether older systems were migrated or archived.
  3. Record families: the business processes each system captures, such as quotes, orders, jobs, tickets, claims or approvals.
  4. History and volume bands: first and last record dates and order-of-magnitude counts per family, with known gaps.
  5. Linkage: the keys that connect records across systems, such as order, job, ticket or invoice numbers.
  6. Content mix: structured fields versus free text, attachments and images, and the main language of the text.
  7. Personal and sensitive data categories: customer contacts, employee data, payment data, health or other special categories, at the level of field names.
  8. Contract constraints: customer, supplier and software agreements that mention data use, confidentiality or ownership.
  9. Privacy promises: what your privacy notices and terms told customers and employees about how their data would be used.
  10. Workload: who would export, link and review, and what else those people are committed to.
  11. Decision owner and stop points: one named owner, a budget of internal time, and the conditions that end the exercise.

0 of 11 ticked. Tick each item as you go.

Who should bring what to the first discovery meeting?

Assign each checklist item to one person before the first meeting, and ask each to bring documents and counts rather than opinions. A single working session with the right people usually answers most of the checklist; a series of informal conversations rarely does.

At a glance

Who should bring what to the first discovery meeting

Keep the meeting itself free of records. Screens showing live customer data, sample exports and spreadsheets with names are not needed for any item on the checklist, and bringing them creates exactly the exposure the no-upload approach is meant to avoid. Field lists, report totals, contract excerpts and policy documents are enough.

Close the meeting with a one-page summary: the answers, the gaps, the owner, the internal time budget and the next decision date. That page is what leadership signs off, and it is also what makes a later conversation with any intermediary or buyer efficient.

  • CEO or owner: headcount history, ownership structure, any prior commitments about company data, and the decision on whether to proceed.
  • Operations lead: the business processes behind each system and which records capture real decisions and outcomes.
  • IT or data lead: system inventory, export methods, date ranges, count bands, linkage keys and the share of free text.
  • Counsel or the person who manages contracts: customer, supplier and software agreements that mention data, plus current and past privacy notices.
  • Finance: internal cost of the people involved and any timing constraints, such as an audit or a system migration.

How do you triage rights without opening records?

Rights triage works from documents, not data: who created the records, what contracts say about them, and what people were told when the data was collected. Keep four questions separate, because answers to one do not settle another: access (can you technically export it), privacy permission (may personal data be used this way), ownership (is it yours or a customer's), and contractual licensing rights (may you license it to a third party).

Privacy promises deserve particular attention. The U.S. Federal Trade Commission has warned that a company adopting more permissive data practices, such as using data for AI training, and telling consumers only through a surreptitious, retroactive change to its terms or privacy policy could be engaging in unfair or deceptive practices. In the EU, the GDPR's data minimization principle requires personal data to be adequate, relevant and limited to what is necessary for the purposes of processing. Neither question can be answered by the data team alone.

For a no-upload assessment, the output of rights triage is a short list: record families that look clear, families that need counsel's review, and families to exclude now. De-identification does not by itself create a right to license; it reduces privacy risk in data you otherwise have the right to use. This is general information, not legal advice; counsel should review your specific contracts and notices.

How do you estimate workload and opportunity cost conservatively?

Estimate workload by naming the people who would do each step and asking them for a range, then use the high end. Most of the effort sits in export, linkage and de-identification review, and it falls on people who already have full-time jobs.

Export effort depends on the system: whether it has a documented bulk export or API, whether history lives in an archive or an old system, and whether attachments are needed. Linkage effort depends on whether keys are consistent across systems. De-identification effort depends on how much free text there is and what it contains; NIST SP 800-188 describes approaches such as removing identifiers, transforming quasi-identifiers and generating synthetic data, and recommends governance such as a review board and re-identification risk assessment.

Opportunity cost is the work those people would not do. Write it down explicitly, for example a delayed system migration or a slower month-end close, so the decision to proceed is made with it in view. Conservative assumptions mean assuming the slower export, the longer review and the narrower set of licensable records.

Worked example (illustrative): a regional services company decides on a next step

This example is fictional. A U.S. commercial HVAC and plumbing contractor with a long operating history runs ServiceTitan for jobs and dispatch, an older field service system retired several years ago, QuickBooks for accounting, and a shared mailbox for customer requests.

Discovery from metadata takes a few meetings. The company is eligible on size and country. Job records in ServiceTitan span several years with technician notes, parts, invoices and callbacks joined by job number; the retired system holds older jobs but its export path is unknown. The shared mailbox has years of free-text requests with customer names and addresses, and no reliable link to jobs.

Rights triage finds standard service terms with no data-use restriction for most commercial customers, but two large facility-management clients have contracts with broad confidentiality clauses. The privacy notice says little about secondary use. Counsel is asked to review those two contracts and the notice.

The decision: proceed with a scoped investigation of ServiceTitan job records only, excluding the two restricted clients and the mailbox, with the operations director as owner and a fixed internal time budget. The retired system is parked until its export path is known. Nothing has been uploaded, and leadership has a clear next step with its limits written down.

When is metadata not enough?

Metadata is not enough when the value or the risk depends on what the content actually says. Four situations come up often, and in each the right move is to plan a controlled look at a sample rather than to guess.

A controlled look means a small, documented sample reviewed under written terms that say who may see it, for what purpose, and when it is deleted. It is a later step with its own approval, not part of the no-upload assessment, and it should only cover the record families that passed rights triage.

  • Free text carries the value: if most of the usefulness is in notes, emails or documents, field names and counts cannot show whether the text is substantive or boilerplate.
  • Language share must be measured: for companies outside the U.S., Canada and the EU, eligibility depends on the English share of the records actually offered, which has to be measured on the dataset rather than assumed from the company's working language.
  • Sensitive categories are mixed in: when health, financial or children's data may appear inside general records, only a reviewed sample under agreed terms can show how much de-identification would be needed.
  • Quality is disputed internally: if teams disagree about how complete or accurate a system is, profiling the data is the only way to settle it.

When should leadership end, shrink or escalate the assessment?

Stop when the checklist shows no eligible record family, no licensable rights, or no one able to do the work. Narrow when one system is clearly stronger than the rest. Request specialist review whenever contracts, privacy notices or sensitive data are involved.

At a glance

When should leadership end, shrink or escalate the assessment

  • Stop: the company does not meet eligibility, or the only substantial records belong to customers under restrictive contracts.
  • Stop: no one can own the work without displacing a higher priority.
  • Narrow: to one system or record family with clear linkage and history.
  • Narrow: by excluding customers, regions or periods with known restrictions.
  • Specialist review: counsel for contracts, privacy notices and cross-border questions; a privacy specialist for de-identification design; finance for internal cost and timing.
  • Revisit: when company priorities, eligibility, offer terms, rights or workload assumptions change.

How the SourceX assessment fits a no-upload approach

With SourceX, the first step is itself a no-upload exercise: a brief company assessment decides whether the company qualifies. SourceX works with U.S., Canadian and EU companies that had 50 or more full-time employees at their peak, contractors excluded; companies in the UK, Australia, New Zealand, India, Singapore, South Africa, Latin America, the UAE and Saudi Arabia can be considered when the data offered is at least 90% English, confirmed on the actual dataset.

If a company proceeds, it approves the scope and terms and signs the agreement before any data is delivered to SourceX. The checklist on this page is a way to arrive at that assessment with clear answers, named owners and known gaps.

Frequently asked questions

Can I find out if my company's data is a fit without sharing any records?

Yes, for the first decision. Eligibility, record families, history and volume bands, linkage and rights questions can all be answered from metadata and documents. Quality and final value need a reviewed sample later, under agreed terms.

What metadata should I prepare for a no-upload assessment?

A system inventory with owners and years in use, record families per system, date ranges and count bands, linkage keys, the share of free text and its language, personal data categories by field name, and the contracts and privacy notices that govern the data.

What can leadership decide before any data is reviewed?

Whether to fund a scoped investigation, which systems to include or exclude, who owns it and what conditions end it. Terms, final scope and delivery decisions should wait for rights review and a sample.

Does de-identification mean we are allowed to license the data?

No. De-identification reduces privacy risk but does not create ownership or contractual rights. Whether you may license records depends on contracts, privacy notices and applicable law, which counsel should review.

When should we stop a data licensing investigation early?

Stop when no record family is eligible, when rights are blocked by contracts you cannot clear, or when no one can own the work without displacing higher priorities. Narrowing to one strong system is often better than stopping.

Sources

  • Datasheets for Datasets proposes documenting each dataset's motivation, composition, collection process and recommended uses. Source
  • Google's Data Cards Playbook describes Data Cards as structured summaries of essential facts about ML datasets, including upstream sources, collection methods and intended use. Source
  • Hugging Face documentation says dataset card metadata describes important information about a dataset such as its license, language and size. Source
  • MLCommons describes Croissant as a format for ML datasets that combines metadata, resource file descriptions, data structure and ML semantics in a single file, building on schema.org. Source
  • FTC staff wrote in February 2024 that adopting more permissive data practices, such as using data for AI training, through a surreptitious, retroactive change to terms or a privacy policy could be unfair or deceptive. Source
  • GDPR Article 5(1)(c) requires personal data to be adequate, relevant and limited to what is necessary in relation to the purposes for which they are processed. Source
  • NIST SP 800-188 (2023) covers de-identification techniques such as removing identifiers, transforming quasi-identifiers and synthetic data, and recommends governance including a disclosure review board. Source

See if your company qualifies

A short company assessment. No data uploads are needed.

See if your company qualifies