Production Workday tenants hold real workers, pay, and security boundaries. This guide is a stop list for consultants: situations where you should refuse to run a personal AI agent against Production, based on what Workday and the major AI vendors publish about accountability and data use.
Refuse: personal consumer AI as the path into Production
Refuse when the only AI path is a personal ChatGPT Free/Plus/Pro or consumer Claude plan and the work includes Production worker data, payroll, or security configuration.
OpenAI documents that personal Free/Plus/Pro workspaces have model-improvement sharing on by default, while Business/Enterprise workspaces do not train on content by default. Anthropic documents that consumer Claude plans can allow training with multi-year retention, while commercial Team/Enterprise/API terms do not train by default. Those are not Workday features; they are the AI account sitting between you and the tenant paste.
A managed firm AI seat, a zero-retention credits path, or Workday-native agents under the customer’s agreement are the alternatives—not “I’ll turn the toggle off later.”
Refuse: shared Integration System User behind an agent
Refuse when an external agent reaches Production APIs through a shared Integration System User (ISU) that many people and prompts share.
Workday’s Agent-Ready tools documentation states that Agent-Ready MCP tools are secured with Delegate Authorization safeguards and that runtime authorization uses an Agent System User (ASU) OAuth flow so calls can be authorized against both the agent and the user. Building a chatbot on a single powerful ISU recreates security in prompt text instead of in Workday groups. That pattern is a reason to stop, not a shortcut.
Refuse: writes with no out-of-band approval
Refuse Production writes when the only “approval” is the model saying yes inside the same chat that an untrusted document can influence.
Workday’s Developer Agent product materials emphasize that agents built with Developer Agent register in Agent System of Record with auditability. Customer custom agents inherit user permissions and can pause for human approvals. If your personal agent stack cannot show a human-approved, tenant-visible change record, do not let it mutate Production.
Refuse: “Sandbox tested” without naming which agent meters where
Refuse a Production go-live that assumed every non-Production test was free of Flex Credits without checking Workday’s exceptions.
Workday’s Platform Consumption Console documentation states that most offers do not apply non-Production usage to the Flex Credit balance, but Deployment Agent and Adoption Agent can consume credits in listed non-Production environments, with Production dashboards reflecting those exceptions. A personal agent that skips that check is guessing with the customer’s money.
Refuse: Developer Agent or custom agents without entitlement clarity
Refuse to run Developer Agent–driven builds for a customer who has not confirmed Extend AI access and Flex Credits policy with their Workday account team.
Workday states Developer Agent is generally available starting 2026R2+, sold only through Extend AI, with Flex Credits-based pricing, and the Rate Card meters Developer Agent at 2 credits per invocation after 50,000 annual invocations. Guessing entitlement from a blog post is not enough for Production.
What to offer instead
- Do the analysis in a non-Production tenant the customer authorizes.
- Use Workday-governed agents (Workday-built or custom agents under ASOR) for in-tenant AI.
- For desktop build/test work, use a firm-managed AI account or a published zero-retention credits path, with client isolation.
- Keep Production promotions, first installs, and model-component changes on the customer’s own approval path.
Limits of this guide
This is practice guidance grounded in public Workday, OpenAI, and Anthropic pages. It is not a substitute for the customer’s security policy, DPA, or counsel. Re-check those pages when they change.