Even when a consultant “won’t train the model,” personal AI tools can remember. That is enough to keep Workday worker IDs, supervisory org names, or pay components on a laptop or in a consumer account after the ticket closes.
What OpenAI publishes
OpenAI’s ChatGPT data-controls article distinguishes model training (“Improve the model for everyone”) from Memory personalization. Turning training off does not delete chats, and Memory is a separate control. Temporary chats are not used to improve models and may be retained up to 30 days for safety while temporary.
If Memory is on and a past chat included employee IDs, later answers can surface that context without a new paste.
What Anthropic publishes for Claude Code
Anthropic’s Claude Code data-usage page states local clients store session transcripts in plaintext under ~/.claude/projects/ for 30 days by default for session resumption (cleanupPeriodDays adjustable). Consumer plans that allow training retain data for 5 years; without training allowance, 30 days on Anthropic’s side for those accounts as documented.
A local Extend folder that contains exported worker XML is therefore a client-data store on the laptop, independent of Workday’s tenant.
Firm response
- Ban Memory features for client Workday work on personal accounts—or ban personal accounts entirely for client content.
- Set
cleanupPeriodDays(or wipe~/.claude/projects/) in offboarding checklists. - Prefer managed Business/Enterprise/Team seats or a published zero-retention credits path for client tenants.
- Never put Production exports into project folders an assistant indexes by default.
This is not a claim that every memory feature is unlawful. It is a claim that retention exists where vendors say it does, and worker identifiers are still client data there.