Skip to content

Leadership and readiness

Dataset supplier handover process: when the data owner changes roles

By SourceX Editorial · Updated

9 sections

Short answer

When a dataset program's data owner changes roles, hand over four things with evidence: source access tied to a role rather than a person, the decision log behind every exclusion and de-identification choice, open issues traced to the releases and recipients they affect, and the calendar of refreshes and notices. The successor signs only after reproducing the latest release from the sources.

Checklist

Dataset supplier handover

This page covers dataset supplier handover process: when the data owner changes roles.Illustrative

Key takeaways

  • A data owner's role change is an access event and a knowledge event; plan both with a named successor and an overlap.
  • Move export jobs and approvals from personal accounts to role accounts, and rotate shared credentials.
  • Hand over the reasons for exclusions, redactions and rights decisions, not only the release files.
  • Trace every open issue from source records to derived examples, releases and recipients before handover.
  • Update every documentation field that names a maintainer or contact.
  • The handover is done when the successor rebuilds the latest release from sources and it matches.

How do you keep a dataset program running when the data owner changes roles?

Treat the role change as a planned transfer with a named successor, an overlap period and a test: the successor reproduces the most recent release from the source systems, using only what was handed over. If that works, the program has survived the move; if it does not, the gaps it exposes are exactly what still lives in one person's head.

In a company that licenses operational records, the data owner is often an operations or IT lead who built the first exports, decided which fields to leave out and answered the buyer's questions. When that person is promoted or moves teams, the dataset rarely breaks at once. It breaks at the next refresh, the next correction request or the next question about why a field was dropped, months later, when nobody can explain the original decision.

The checklist on this page covers the four things that must move: access, decisions, open issues and commitments. Each item names the evidence to hand over and the condition that marks it done, so the outgoing owner, the successor and whoever approved the program can all sign the same list.

What breaks first when the person who ran the exports moves on?

Access breaks first, because export jobs and API tokens are often created under the owner's personal account. Decisions break next, because the reasons behind exclusions and redactions were never written down outside email threads and chat.

Security frameworks already treat a role change as an access event, not only a departure. NIST SP 800-53 control PS-5, Personnel Transfer, calls for reviewing and confirming the ongoing operational need for current access when a person is reassigned, modifying access to match, and, among possible actions, providing access to the official records the person had access to in previous accounts. Control AC-2, Account Management, asks organizations to notify account managers when users are transferred, to change shared or group account authenticators when individuals are removed from the group, and to align account management with personnel transfer processes. A dataset program should run the same review for every system its exports touch.

  • Scheduled exports fail when the owner's personal token is revoked under the transfer review.
  • A shared service account keeps working, but its credential is still known to someone who no longer needs it.
  • The successor cannot say why a notes field was excluded, so a well-meant refresh puts it back.
  • The list of who received which release lives in the former owner's sent mail.
  • A correction the buyer was promised is still open, and nobody else knows its scope.

The source access evidence and decision handover checklist

Work through the checklist during the overlap period, with the outgoing owner and the successor both present for each item. An item is done only when its evidence exists somewhere the successor controls and the done-when condition has been observed, not described.

The source access evidence and decision handover checklist
ItemEvidence to hand overDone when
Ownership recordProgram owner and backup named as roles, with the successor's name and start dateDataset documentation and internal register show the new owner
Source system accessList of systems, service accounts, scopes and where credentials are storedPersonal tokens replaced by role accounts; shared credentials rotated
Export definitionsQueries, job configs and filters in version control, not in a desktop folderSuccessor runs each export from the repository
Lineage mapSource tables and fields to derived files, per release versionEvery field in the latest release traces to a source
Decision logExclusions, redactions, de-identification choices and rights review outcomes, each with reason and approverSuccessor can explain the three most consequential decisions unaided
Release and recipient logVersion, date, contents hash or manifest and recipient for every deliveryLog matches what recipients confirm they hold
Open issue registerEach issue with id, severity, affected sources, releases and recipientsEvery open issue has an owner and a next date
Commitments calendarRefresh dates, correction and notice obligations as written in each agreementCalendar entries reference the agreement clause, read by the successor
Counterparty contactsWho to contact at each recipient and at any processing partner, and for whatSuccessor introduced in writing to each contact
Reproduction testNotes from rebuilding the latest release from sourcesRebuilt release matches the delivered manifest, or every difference is explained

Which records prove why a decision was made, not just what was delivered?

A decision log entry needs the decision, the reason, the evidence consulted, the approver, the date, the releases it applies to and the trigger that would reopen it. A release manifest shows what was delivered; only the log shows why a field, a date range or a customer segment is absent.

At a glance

Which records prove why a decision was made, not just what was delivered

Dataset documentation standards ask for the maintenance side of this. The Datasheets for Datasets paper's maintenance questions ask who will support, host and maintain the dataset, how its owner or curator can be contacted, whether there is an erratum, and whether it will be updated to correct errors or add or remove instances. The Croissant responsible AI vocabulary includes a release maintenance plan property covering updating timeframe, maintainers and deprecation policies, and the Hugging Face dataset card template has fields for who curated the dataset and a dataset card contact. Each of those fields names a person or role, so each one must be updated at handover, or the published documentation points recipients to someone who no longer answers.

  • Decision id and date, with the release versions it governs.
  • What was decided, in one sentence a newcomer can apply.
  • Why, citing the evidence: a rights review note, a privacy assessment, a buyer requirement.
  • Who approved it, by role and name.
  • Revisit trigger: the change in source, agreement or law that would reopen it.

Which released examples and recipients does a source change affect?

Answer that by following one issue from the source record to every derived example, every release that contains it and every recipient of those releases. If the handed-over logs cannot support that trace in a few minutes, the successor inherits a program that cannot answer the first correction request it receives.

Use the issue register below as the unit of handover for anything still open. The illustrative entries follow one issue: a batch of technician notes in which phone numbers were not masked.

Which released examples and recipients does a source change affect?
FieldWhat to recordIllustrative entry
Issue id and severityStable id and a level tied to stated criteriaDI-07, high: personal data present in delivered text
Source and derived linksSource records and the derived examples built from themNotes from one source table, one quarter; examples in files 3 and 4
RepairThe change made and where it is versionedMasking rule added to export config, commit noted
Regression checksTests that prove the repair and protect other fieldsPattern scan on all text fields; row counts unchanged
Recipient and version logEvery recipient holding an affected versionReleases 2 and 3; two recipients
Change noticeWhat was sent, to whom, when, under which agreement termNotice and replacement files sent to both recipients
ReconciliationConfirmation that recipients replaced or deleted affected filesOne confirmation received, one pending; follow-up date set
ClosureWho closed it and on what evidenceOpen until the second confirmation arrives

How should ownership be assigned so the next move is easier?

Assign the dataset to a role or team with a named lead, and give approval rights to the team rather than to one account. Then the next role change updates one membership instead of every job, approval rule and document.

Software repositories show the pattern. GitHub's code owners feature lets a CODEOWNERS file name individuals or teams responsible for paths in a repository, automatically requests their review on pull requests that modify those paths, and requires that a team listed as owner be visible and have write permission. A dataset program that keeps its export definitions and decision log in a repository can name a team as owner of both, so a change to a filter or a redaction rule cannot merge without someone currently in the role reviewing it.

The same thinking applies outside code: a shared mailbox for recipient questions instead of a personal address, a role account for scheduled exports, and documentation fields that name the role first and the current person second.

Role-based ownership also changes what the overlap period has to cover. When approvals, mailboxes and job accounts already belong to a team, the handover shrinks to the knowledge items on the checklist: decision reasons, open issues and commitments. When they belong to a person, every one of them has to be found, moved and tested, and the ones nobody remembers are found only when they fail.

Keep a short ownership register for the program itself: one line per system, job, document and recipient relationship, with the owning role, the current person and the date last confirmed. Reviewing that register each quarter, and at every role change in the team, catches drift long before a handover forces the question.

Worked example (illustrative): an operations analyst moves to a planning role

In this illustrative case, a field service company has licensed three quarterly releases of work order and technician note records. The operations analyst who built the program moves to a planning role and will remain at the company. Two recipients hold releases, one issue is open (the unmasked phone numbers above) and the next refresh is due in six weeks.

The company sets a two-week overlap. In week one the analyst and the successor, a data engineer, walk the checklist: two export jobs are moved from the analyst's personal token to a role account, a shared database credential is rotated, and the decision log is rebuilt from email for the four decisions that shaped the releases, including why customer addresses were reduced to postal code. In week two the engineer rebuilds release 3 from sources. The row counts match, but one derived file differs because a filter lived only in the analyst's local copy of a query; it is committed and the rebuild then matches.

The exception is an owner who leaves without notice. NIST control PS-4, Personnel Termination, calls for disabling access and revoking credentials while retaining access to the information and systems the person controlled. For a dataset program that means pausing new releases, reconstructing the checklist from systems and recipients rather than from the person, and resuming only after the reproduction test passes.

When should the program pause instead of handing over?

Pause new releases when the successor cannot reproduce the latest release, when the decision log is missing the reasons for any rights or de-identification decision, or when a high-severity issue is open without a confirmed repair. Shipping a refresh under those conditions converts an undocumented gap into a new delivery that repeats it.

Narrow the scope rather than pausing everything when only one source is affected: continue releases from systems that pass the checklist and hold the rest. Bring in counsel or the licensing counterparty when the commitments calendar depends on agreement terms nobody on the team can interpret, such as notice periods or deletion obligations, rather than guessing at them. This page describes operational practice, not the meaning of any specific agreement.

Handing over a program that runs with SourceX

When records are licensed through SourceX, nothing is delivered before the company has signed off on scope and terms, and SourceX carries out the de-identification through approved processing routes. Those approvals, the signed agreement and the delivery records belong on the handover checklist like any other decision evidence.

The successor should also be introduced to the SourceX contact for the program, so questions about refreshes or corrections reach someone who can answer them. Companies considering a first program can start with a short assessment that needs no data uploads.

Frequently asked questions

Who should own a dataset program after the original owner moves?

A role or team with a named lead and a named backup, not a single account. The successor should have authority over the source systems, the export definitions and the decision log, and should be named in the dataset documentation.

How long should the handover overlap be?

Long enough for the successor to rebuild the latest release from sources and walk every checklist item with the outgoing owner. The length depends on the number of sources and open issues; the reproduction test, not the calendar, decides when it ends.

What if the decisions were never written down?

Rebuild the log during the overlap from email, tickets and release notes, starting with exclusions, redactions and rights review outcomes. Mark any decision whose reason cannot be recovered and review it before the next release rather than repeating it by default.

Do we need to tell recipients when the data owner changes?

Check each agreement for contact and notice terms, and update any published dataset documentation that names the maintainer or contact. Introducing the successor in writing to each recipient is good practice even where no notice is required.

What is the single best test that a handover worked?

The successor rebuilds the most recent release from the source systems using only the handed-over materials, and the result matches the delivered manifest or every difference is explained. Anything the rebuild needed that was not handed over is a gap to close.

Sources

  • NIST SP 800-53 Rev. 5 controls PS-5 (Personnel Transfer), AC-2 (Account Management) and PS-4 (Personnel Termination) set out access review, notification, credential and record access steps when people change roles or leave. Source
  • The Datasheets for Datasets maintenance questions ask who maintains the dataset, how the owner can be contacted, whether there is an erratum and whether the dataset will be updated. Source
  • The Croissant responsible AI vocabulary defines a release maintenance plan property covering updating timeframe, maintainers and deprecation policies. Source
  • The Hugging Face dataset card template includes fields for who curated the dataset and a dataset card contact. Source
  • GitHub CODEOWNERS files name individuals or teams responsible for code, owners are automatically requested for review on changes, and owner teams must be visible with write permission. Source

See if your company qualifies

A short company assessment. No data uploads are needed.

See if your company qualifies