Enforcing per-employee Workday OAuth guarantees an employee cannot pull records Workday denies them, yet it leaves a critical second gap wide open: privileged users accidentally pulling sensitive fields they are allowed to see into an external LLM context window. Deterministic in-memory field stripping closes that gap by dropping high-risk Workday attributes—such as nationalIdentifiers, bank paymentElectionsCurrent, and peer compensation fields—in RAM before the payload ever reaches Claude, ChatGPT, or Gemini.
Think about what happens when an HR Partner, Payroll Administrator, or Benefits Partner asks an AI assistant a routine question like “Summarize the active staffing roster for Supervisory Organization Warehouse 4.” If the assistant calls a shared custom RaaS report, SOAP Get_Workers, or broad WQL query, Workday faithfully checks that HR Partner’s security groups, sees legitimate View / Get permissions on Person Data: ID Information and Worker Data: Compensation, and returns every column in the payload. Without field-level redaction at the connector boundary, Social Security Numbers, IBANs, and base salaries land directly inside a chat transcript for a question that only needed names and job titles.
Why Workday Domain Security Alone Cannot Solve Context Over-Collection
In Workday’s domain security model, a user’s permission to a report field or REST subresource is binary per target constraint: if your security group has View / Get access on the securing domain, Workday returns the value whenever an API call or shared RaaS report requests it.
You cannot solve AI over-collection simply by stripping permissions inside Workday, for three operational reasons:
- Breaking day-to-day Workday UI work: Revoking
Person Data: ID Information(Personal Data functional area) orWorker Data: Payroll (Payment Elections)(Core Payroll functional area) from your Payroll and HR administrators prevents them from fixing tax IDs (Edit IDs) or resolving direct-deposit errors inside the Workday browser UI. - Coarse-grained SOAP and RaaS payloads: Legacy Workday Web Services (
Human_ResourcesGet_WorkerswithInclude_Personal_InformationorInclude_Compensation) and multi-purpose shared custom RaaS reports bundle dozens of fields into a single response document. An assistant calling an existing RaaS endpoint receives every column that the caller’s security groups permit. - LLM exploration in ad-hoc WQL: When an AI assistant dynamically composes WQL against
allWorkersto answer an analytical question, it may select extra columns (totalBasePay_Amount,disciplinaryActionsInProgress) “just in case” unless the gateway blocks those fields both in the query AST and in the returned JSON buffer.
High-Risk Workday Domains and Fields That Rarely Belong in Chat
Before connecting any AI assistant to a tenant—even after auditing over-permissive security groups—map the exact Workday domains and WQL/RaaS fields that should be blocked at the AI connector layer across all personas:
| Data Category | Workday Securing Domain (Functional Area) | Example allWorkers WQL / API Fields |
Recommended AI Connector Rule |
|---|---|---|---|
| Data CategoryNational & Tax IDs | Workday Securing Domain (Functional Area)Person Data: ID Information (Personal Data) |
Example allWorkers WQL / API FieldsnationalIdentifiers, nationalIdentifiersDisplayedWithIDOnly |
Recommended AI Connector RuleBlock globally in RAM for all personas |
| Data CategoryBank & Direct Deposit | Workday Securing Domain (Functional Area)Worker Data: Payroll (Payment Elections) (Core Payroll) |
Example allWorkers WQL / API FieldspaymentElectionsCurrent, paymentElectionOptionLines |
Recommended AI Connector RuleBlock globally in RAM (keep boolean status only if needed) |
| Data CategoryDisciplinary & PIPs | Workday Securing Domain (Functional Area)Self-Service: Disciplinary Action / Talent review domains |
Example allWorkers WQL / API FieldsdisciplinaryActionsInProgress, cf_DisciplinaryActions_Completed |
Recommended AI Connector RuleBlock globally or restrict strictly outside general chat |
| Data CategoryBase & Bonus Pay | Workday Securing Domain (Functional Area)Worker Data: Compensation / Worker Data: Compensation by Organization (Core Compensation) |
Example allWorkers WQL / API FieldstotalBasePay_Amount, totalBasePayAnnualized_Amount, bonusAndOne_TimePayments |
Recommended AI Connector RuleAllow self-only or block unless explicitly enabled for comp workflows |
| Data CategoryHome Contact Details | Workday Securing Domain (Functional Area)Person Data: Home Contact Information (Personal Data) |
Example allWorkers WQL / API FieldsHome street address, personal phone, personal email | Recommended AI Connector RuleAllow self-only; strip from multi-worker roster reads |
How Deterministic In-Memory Field Stripping Works
Unlike prompt rules (“Do not repeat Social Security Numbers to the user”)—which still transmit the raw SSN to the AI provider’s servers—deterministic gateway redaction intercepts the response inside volatile RAM before the MCP tool result leaves the connector.
A defense-in-depth Workday AI gateway applies redaction at two stages of the same request:
Stage 1: WQL and SOAP Request Projection Pruning
Before sending an ad-hoc WQL query to POST /wql/v1/data or a SOAP Get_Workers envelope to Workday, the gateway inspects the requested field list and response groups. If an admin rule blocks nationalIdentifiers or paymentElectionsCurrent, the gateway rejects or strips those selectors from the outbound Workday query so Workday doesn’t even serialize them.
Stage 2: Schema-Aware JSON/XML Stream Stripping in RAM
Custom RaaS reports (?format=json) and fixed REST endpoints return preset schemas where the caller cannot pick individual columns on the URL. As the upstream HTTP stream arrives in memory, the gateway walks every object node and drops blocked keys and nested structures deterministically:
Raw Upstream Workday Payload (In Gateway RAM Only — Never Sent to LLM):
{
"Report_Entry": [
{
"Employee_ID": "21097",
"Worker": "Gail Bradley",
"Business_Title": "Warehouse Supervisor",
"Supervisory_Organization": "Warehouse 4",
"National_ID": "999-01-4421",
"Bank_Account_Number": "0048291044",
"Total_Base_Pay_Amount": "94500.00"
}
]
}Filtered Tool Output Returned to Claude, ChatGPT, or Gemini:
{
"Report_Entry": [
{
"Employee_ID": "21097",
"Worker": "Gail Bradley",
"Business_Title": "Warehouse Supervisor",
"Supervisory_Organization": "Warehouse 4"
}
]
}Notice why removing the key cleanly (or replacing it with a compact policy notice "_redacted_fields": ["National_ID", "Bank_Account_Number", "Total_Base_Pay_Amount"] at the envelope level) outperforms regex masking:
- A numeric salary like
"94500.00"or a 10-digit bank account number looks identical to a cost center code or requisition number to a regex; key-path and Workday schema matching never misfires. - Dropping blocked keys across 500 RaaS rows shrinks token consumption in the LLM context window while guaranteeing zero sensitive digits leave RAM.
- If the LLM asks why pay wasn’t returned, the
_redacted_fieldsmetadata tells it clearly “Blocked by tenant admin field policy” so it stops retrying alternative queries.
Testing Your Field Redaction Rules Before Production Rollout
Run these three verification checks in your Workday Sandbox before publishing an AI connector to HR or finance staff:
- Test as your highest-privilege persona, not your lowest: Sign in to Sandbox as an
HR PartnerorPayroll Administratorwho holds unconstrainedViewaccess onPerson Data: ID InformationandWorker Data: Compensation. - Trigger both structured and custom report paths: Ask the assistant to fetch a worker profile via REST, run a WQL query against
allWorkers, and execute a shared custom RaaS roster report containing compensation and ID columns. - Inspect the raw MCP tool-result JSON in the client inspector: Verify that the raw JSON payload delivered to the AI client omits every blocked field—proving redaction happened in gateway memory rather than inside the LLM’s final prose response.
How Kiweely Bridge Automates Field-Level Redaction
Kiweely Bridge (coming next) pairs user-delegated Workday OAuth with deterministic in-memory field rules configured alongside an AI admin agent (see Bridge security details):
- Guided field lockdown in Sandbox: During setup, Bridge’s AI admin agent inspects your tenant’s API client scopes and RaaS schemas, sets field-level redaction rules for National IDs, bank details, and sensitive HR fields, and tests prompts across real personas (
Employee,Manager,HR Partner) before anyone in Production connects. - Stripped in RAM, zero stored payloads: Bridge strips admin-blocked fields in memory and writes zero Workday records, prompts, or answers to disk, databases, or logs.
- Simple company pricing: Flat monthly plans by licensed Workday employee count ($499/mo up to 1,000 employees, $1,499/mo for 1,001–5,000, $3,999/mo for 5,001–20,000, Enterprise by quote above 20,000 or for dedicated/self-hosted VPC containers, and free for one person on a personal Workday developer tenant).
Frequently asked questions
If user-delegated OAuth already enforces Workday security, why do we still need field-level redaction for AI?
Workday security governs what a human role is allowed to see inside Workday—not what should flow into an external LLM context window. An HR Partner or Payroll Partner legitimately needs Person Data: ID Information in Workday for I-9 or tax fixes, yet never needs raw SSNs or bank account numbers sent to Claude, ChatGPT, or Gemini.
Why can't we just remove Person Data: ID Information from HR Partners in Workday?
Editing a core Workday domain security policy affects the Workday web UI, delivered reports, and business processes simultaneously. Removing that domain breaks the HR Partner's day-to-day Workday job. Gateway-level field rules narrow AI exposure without breaking native Workday tasks.
Why is regex PII masking unreliable for Workday compensation and HR data?
Regexes catch formatted US Social Security Numbers, not unformatted foreign national IDs, IBAN/routing numbers, annualized salary integers like 162000, bonus percentages, or disciplinary status codes. Schema-aware field stripping drops exact WQL aliases, SOAP XML nodes, and RaaS JSON keys deterministically.
Does stripping fields in RAM store anything on the gateway server?
No. Streamed JSON/XML payloads are parsed in volatile memory, blocked keys are removed or replaced with a redaction marker, and the filtered structure is returned directly over TLS with zero disk or log writes.