Many Workday consultants still open a personal ChatGPT or Claude tab, paste a tenant screenshot, and ask for a calculated-field fix. The model answers. The firm often has no written rule about whose AI account just saw the client’s data. OpenAI and Anthropic publish different defaults for personal plans than for business workspaces. This post sticks to those published controls and what a firm should decide before anyone pastes Workday content.
What OpenAI publishes for ChatGPT
OpenAI’s Data controls in ChatGPT (updated within the last week as of this writing) states:
- On personal workspaces with ChatGPT Free, Plus, or Pro, data sharing is enabled by default. Users can turn off Improve the model for everyone and keep chat history.
- Turning that setting off does not delete saved chats.
- By default, OpenAI does not use content from ChatGPT Business, Enterprise, Edu, or ChatGPT for Healthcare workspaces to train its models.
- By default, OpenAI also does not use API Platform inputs and outputs to train its models.
- Codex has a separate “Include environments” setting for whether Codex environment context can be used for training; changing the ChatGPT setting alone does not change it.
- Temporary chats are not used to improve models and may be retained up to 30 days for safety while they remain temporary.
If a consultant uses a personal Free/Plus/Pro account with the default on, and pastes worker rows or security group names, that content is in scope for model improvement unless they opted out.
What Anthropic publishes for Claude Code
Anthropic’s Claude Code data usage page states:
- Consumer users (Free, Pro, Max, Ultra): Anthropic trains new models using data from those accounts when the training setting is on, including Claude Code used from those accounts.
- Retention for consumers: 5 years if training is allowed; 30 days if it is not.
- Commercial users (Team, Enterprise, API): Anthropic does not train generative models on Claude Code prompts/code under commercial terms unless the customer opts into something like the Development Partner Program.
- Commercial standard retention is 30 days; zero data retention is available to qualified Enterprise accounts on a per-organization basis and is not included in the standard Enterprise plan by default.
- Local Claude Code session transcripts are stored in plaintext under
~/.claude/projects/for 30 days by default for session resumption.
A personal Pro/Max Claude Code session pointed at an Extend project folder is not the same control surface as a firm Claude Team/Enterprise seat.
What Workday publishes about its own agents
Workday’s own AI (Custom Agents, Developer Agent, Agent-Ready tools) runs under Workday agreements, Agent System of Record, and Flex Credits for UMSA customers. That is a different product boundary from pasting tenant data into a consumer chatbot. Do not treat “we turned on Workday AI” as covering personal ChatGPT use.
The firm policy gap
Most AMS and implementation contracts talk about client data, laptops, and subcontractors. Few name:
- Whether personal AI accounts are allowed at all for client Workday content.
- Whether Consumer training toggles must be off (and how the firm verifies that).
- Whether only Business/Enterprise/Team seats (or a zero-retention path such as Kiweely Credits) are allowed.
- Whether Codex / Claude Code local project folders that contain tenant exports are in scope for offboarding wipes.
Without those four answers, every personal account is a shadow processor.
Practical minimum
- Ban pasting Production worker, payroll, or security configuration into personal Free/Plus/Pro ChatGPT or consumer Claude until the firm has a written exception.
- Prefer a managed ChatGPT Business/Enterprise or Claude Team/Enterprise workspace, or a path that publishes zero retention for the work (Kiweely documents zero retention for Kiweely Credits).
- For Codex, check the Include environments training setting separately from ChatGPT’s Improve-the-model control.
- On offboarding, delete local Claude Code project transcripts and ChatGPT/Codex histories that held client material, under the firm’s data-handling procedure—not only revoke Workday access.
What this post does not do
It is not legal advice. Provider policies change; re-check OpenAI and Anthropic pages before you brief a client. It does not claim every personal AI use is unlawful—only that defaults on personal plans are not the same as commercial workspace defaults, and firms that ignore the gap are flying blind.