To connect ChatGPT Enterprise or ChatGPT Team to live Workday data safely for your whole workforce, register a user-delegated Workday OAuth client (Register API Client with Authorization Code Grant and PKCE) behind a remote Model Context Protocol (MCP) server—never a shared Integration System User (Register API Client for Integrations). With user-delegated OAuth, every employee signs in once on Workday’s own SSO screen, and Workday evaluates every REST, WQL, SOAP, and custom report (RaaS) call against that specific employee’s security groups, constrained organizations, and domain policies.
When employees cannot query Workday directly from ChatGPT, they export CSV reports, copy compensation tables, and paste sensitive worker rows into chat threads where Workday’s security policies no longer apply—and where balances and org charts go stale within hours. Yet the quickest shortcut many IT teams prototype—wiring a single Integration System User credential into an internal MCP wrapper—creates a far worse governance failure.
Why Shared Service Accounts Fail Security Review for ChatGPT
Workday provides two completely distinct OAuth registration tasks, and confusing them is the single biggest architectural mistake in enterprise AI rollouts:
| Security Dimension | Shared ISU (Register API Client for Integrations) |
User-Delegated OAuth (Register API Client + PKCE) |
|---|---|---|
| Security DimensionWho Workday sees on every call | Shared ISU ( Register API Client for Integrations )One static ISU_* account for every ChatGPT user |
User-Delegated OAuth ( Register API Client + PKCE)The individual signed-in worker (Employee_ID / Workday Account) |
| Security DimensionRow-level & org constraints | Shared ISU ( Register API Client for Integrations )Typically Integration System Security Group (Unconstrained) seeing the whole company |
User-Delegated OAuth ( Register API Client + PKCE)Role-based constrained groups (Employee As Self, Manager for own Supervisory Organization only) |
| Security DimensionCompensation & PII blast radius | Shared ISU ( Register API Client for Integrations )One bug in middleware prompt filtering leaks executive or peer pay | User-Delegated OAuth ( Register API Client + PKCE)Enforced inside Workday’s kernel before a single byte leaves the tenant |
| Security DimensionSSO and MFA enforcement | Shared ISU ( Register API Client for Integrations )Bypassed (Do Not Allow UI Sessions, static refresh token) |
User-Delegated OAuth ( Register API Client + PKCE)Enforced on Workday’s sign-in page via corporate SAML/OIDC and MFA |
| Security DimensionWorkday Audit Trail | Shared ISU ( Register API Client for Integrations )Every read and business process logs under ISU_ChatGPT |
User-Delegated OAuth ( Register API Client + PKCE)Signons and Attempted Signons and BP history log the exact employee |
| Security DimensionTermination revocation | Shared ISU ( Register API Client for Integrations )Access persists unless IT manually strips ChatGPT workspace membership | User-Delegated OAuth ( Register API Client + PKCE)Deactivating or terminating the worker in Workday revokes OAuth refresh tokens immediately |
An ISU belongs in overnight payroll feeds and point-to-point background syncs (covered in our ISU and API client setup guide). For an interactive assistant opened by individual contributors, managers, and HR business partners on laptops and phones, only Register API Client with per-user OAuth preserves Workday’s configurable security model.
Why Workday’s Native Connectors Leave Out ChatGPT
IT leaders evaluating first-party options quickly hit two roadblocks when their company standardizes on OpenAI:
- No first-party ChatGPT app: Workday’s native conversational connectors center on the Sana from Workday beta in Claude and Sana for Gemini Enterprise, leaving ChatGPT Enterprise and Team workspaces without a first-party Workday connector.
- Commercial and licensing gates on Agent Gateway: Routing custom MCP clients through Workday’s own Agent Gateway (
us.agent.workday.com/{tenant}/mcp) requires signing the Universal Main Subscription Agreement (UMSA) or Innovation Services opt-in, metering usage against Workday Flex Credits, holding Workday Extend Professional (Extend AI), and registering individual Agent-Ready Tools in Agent System of Record (ASOR)—which still do not expose ad-hoc WQL or arbitrary existing RaaS reports.
Connecting through standard Workday APIs (Register API Client) sidesteps every one of those prerequisites because every core Workday tenant already ships with OAuth 2.0 Authorization Code + PKCE, REST, WQL, SOAP Web Services, and RaaS.
How Per-Employee OAuth + Remote MCP Works Step by Step
ChatGPT Enterprise and Team connect to external company tools over HTTPS using remote MCP servers paired with OAuth 2.1 / PKCE dynamic authorization. Here is the exact request lifecycle when user-delegated Workday OAuth is wired cleanly:
1. Tenant Registration in Workday (Register API Client)
A Workday Security Administrator (holding Security Administration in the System functional area) runs Register API Client:
Task: Register API Client
Client Grant Type: Authorization Code Grant
Support Proof Key for Code Exchange (PKCE): Selected
Access Token Type: Bearer
Redirection URI: https://<connector-host>/oauth/callback
Scope (Functional Areas):
- Core Staffing
- Personal Data
- Absence
- Time Tracking
- System (plus Include Workday Owned Scope for WQL metadata discovery)For SSO-only tenants where employees have no native Workday password, see how PKCE OAuth works with corporate SSO.
2. One-Time Workspace Connection and Employee Sign-In
- A ChatGPT Enterprise or Team workspace admin adds the remote MCP connector URL once for the organization—zero Python runtimes, Node scripts, or MDM laptop packages to push, meaning ChatGPT works identically across macOS, Windows, web browsers, iPhone, and Android.
- The first time an employee asks ChatGPT “How many vacation days will I have accrued by December 20?” or “Who on my team has pending time-off requests?”, ChatGPT initiates the OAuth redirect.
- The employee completes corporate SAML SSO and MFA directly on Workday’s own sign-in screen. The MCP layer never sees the employee’s password or MFA factor.
3. Live Multi-Surface Execution Under Employee Scope
Once signed in, a single question in ChatGPT routes to the right Workday API surface—covered in comparing RaaS, WQL, and REST—always carrying Authorization: Bearer <employee_access_token>:
- Workday REST APIs for self-service tasks, inbox items, and worker profiles (
GET /common/v1/workers/me,GET /absenceManagement/v1/...). - Live WQL (
/wql/v1/data) for ad-hoc analytical joins and org rollups (Supervisory Organizationheadcount, open requisitions, time-off balances) discovered dynamically against your tenant’s schema. - Custom RaaS reports so HRIS teams can expose governed calculated fields that already power internal dashboards without rebuilding logic in code.
- SOAP Web Services (authenticated via HTTP
Bearertoken) for legacy benefits, staffing, and financial reads that Workday has not yet exposed in REST.
If an individual contributor asks ChatGPT “What is Maria’s base salary?”, Workday’s own domain security policy (Worker Data: Compensation) rejects or omits the fields—just as if that employee typed Maria’s name into the Workday search bar (see why a Workday user cannot see a field or task).
Eliminating the Middleware Data Honeypot and Chat Write Risks
Solving the Workday ISU problem leaves two security questions that CISO and InfoSec teams ask before approving any ChatGPT connector:
- Where do employee OAuth tokens and Workday HR payloads live inside the MCP gateway? Traditional integration gateways store long-lived OAuth refresh tokens in a vendor database and write request/response payloads to observability logs.
- What stops a prompt injection in an email or document pasted into ChatGPT from triggering a Workday write? In-chat “Type YES to confirm” prompts can be socially engineered or bypassed when autonomous tool calling is active.
Kiweely Bridge—coming next for Claude, ChatGPT, and Gemini—was designed specifically around those two constraints (detailed on Bridge security and Bridge comparison):
- Zero stored credentials and zero stored payloads: Bridge keeps no database table of employee Workday tokens. Each employee’s OAuth grant is encrypted with AES-256-GCM and carried inside the client session token, decrypted only in RAM for milliseconds during a live Workday call. Schema-locked telemetry logs only timestamp, tenant ID, tool name, HTTP status, and latency—never WQL rows, worker names, or pay figures.
- Read-only by default, out-of-band browser approval for writes: Disabled write tools are absent from the tool list ChatGPT sees. When an admin enables a write action (such as
Request Time Offor updating contact details), Bridge pauses execution and opens an out-of-band approval card in the employee’s browser displaying Now vs. After. ChatGPT stays blocked on a short-lived action hash until the human clicks Allow in the browser. - AI admin agent setup and persona testing: Instead of hand-crafting JSON policies, an AI admin agent helps HRIS and IT set up
Register API Clientin Sandbox, apply RAM-level field redaction rules, and test prompts asEmployee,Manager, andHR Partnerbefore publishing to Production. - Flexible perimeter & predictable pricing: Run on zero-storage Kiweely Cloud, a Dedicated single-tenant pool with a static outbound IP (for Workday
Maintain IP RangesandManage Authentication Policies), or a self-hosted stateless container inside your own AWS, GCP, Azure, or Kubernetes VPC. Pricing is flat per company by licensed Workday employee count with unlimited calls covering Production plus all Sandbox and implementation tenants: $499/mo (up to 1,000 employees), $1,499/mo (1,001–5,000), $3,999/mo (5,001–20,000), Enterprise by quote above 20,000 or dedicated/self-hosted, and free for one person on their own Workday developer tenant.
(Note: If your team wants visual self-service portals, manager dashboards, or chat approval cards outside Workday rather than conversational AI inside ChatGPT, Kiweely Leaf—also coming next—uses the same stateless user-delegated Workday OAuth pass-through for custom browser, Teams/Slack, and mobile screens. Consultants building Extend apps and Studio integrations today can start immediately in Kiweely Vine.)
Pre-Flight Checklist for IT and Workday Admins
Before rolling out a Workday MCP connection to your ChatGPT Enterprise or Team workspace:
- Reject any architecture backed by
Register API Client for Integrationsor a sharedISU_*account for multi-employee chat. - Confirm
OAuth 2.0 Clients Enabledis checked in Edit Tenant Setup - Security and register a dedicatedAuthorization Code Grant+PKCEclient in Sandbox first. - Scope Functional Areas tightly to only the domains employees should query, and remember to mint fresh tokens whenever scopes change.
- Test three real personas in Sandbox—an individual contributor (
Self-Serviceonly), a line manager (Role-Based Constrained), and an HR partner—to verify Workday’s domain boundaries hold identically in ChatGPT and the Workday UI. - Require out-of-band browser confirmation before enabling any mutating business process or REST
POST/PATCHendpoint.
Frequently asked questions
Does Workday offer an official ChatGPT connector today?
No. Workday's first-party external AI partnerships cover the Sana from Workday connector in Claude (beta) and Gemini Enterprise on Google Cloud; Workday does not ship a first-party ChatGPT connector.
Why should IT reject an Integration System User (ISU) for a company-wide ChatGPT Workday connector?
An Integration System User registered via Register API Client for Integrations runs every API query under one shared security profile. If you attach an unconstrained or broad ISU to a multi-user ChatGPT workspace, any prompt mistake or misconfigured filter risks exposing other employees' compensation, performance, or personal records—and every transaction in Workday's audit trail attributes to the service account instead of the human employee.
Do we need Workday Agent System of Record, Flex Credits, or Workday Extend to connect ChatGPT?
No. User-delegated MCP connectors such as Kiweely Bridge (coming next) use Workday's standard Register API Client task (Authorization Code Grant with PKCE) available on every core Workday HCM or Financials tenant, with zero Workday AI agreement, Flex Credits, or Extend subscription required.
How are write actions like Request Time Off protected against prompt injection in ChatGPT?
In Kiweely Bridge (coming next), connections are read-only by default. When an administrator enables a write action, Bridge pauses the call and opens an out-of-band approval card in the employee's browser showing the exact Now vs. After values. A prompt injection inside the ChatGPT thread cannot click Allow in that separate browser session.