AI agents that work with employee records
Let HR agents draft and look things up, and put a person in front of offers, payroll changes and anything that ends someone's access. With NobleCloak, policy is set per tool and scoped by role or by person, so compensation tools stay out of reach outside the comp team and an offer goes out only after an approver says yes. Every decision is recorded.
100 scenarios in 5 areas
These are scenarios for how teams govern AI agents in People & HR. Some depend on capabilities still in development. Products named in a scenario are systems your team runs, named to make the example concrete. The policy applies to the AI agent's calls into them.
Recruiting & Hiring20 examples
Offers go out only on a yes
The recruiting agent drafts offer letters in Greenhouse, but `send_offer` pauses for the recruiter to confirm in-session, so no candidate gets an offer the team hasn't read.
Sponsored job posts on a budget
The sourcing agent can boost postings on LinkedIn and Indeed within a monthly budget, and any single boost above $500 needs approval.
Bulk rejections pause
Rejecting a batch of candidates in the ATS asks first, so a misread filter doesn't send 300 rejection emails to people still in process.
Offers above band need sign-off
An offer whose base salary exceeds the approved band for the level routes to the head of talent before it can be sent.
Background checks under a cap
Background checks through Checkr are limited to that vendor, with a per-call cap matched to the standard package and a monthly budget per hiring team.
Hiring managers see scorecards, not comp
The hiring manager role can read and submit interview scorecards; compensation and offer-detail tools are not entitled to that role, so their assistants never surface them.
Sourcing stays on approved sites
The sourcing agent's web access goes through the governed proxy to approved job boards and professional networks only, so it can't scrape arbitrary personal sites.
EEO data walled off
Tools that read voluntary EEO self-identification data aren't entitled to recruiters, hiring managers, or the screening agent; only the HR compliance role can reach them for reporting.
Bias-audit evidence for AI screening
Every resume-screening call the agent makes is in the ledger with inputs and output, giving the company a record to support its NYC Local Law 144 bias audit.
Agencies submit, can't browse
An external staffing agency's agent may call in over A2A to submit candidates to open reqs, but pipeline-read tools aren't on its peer allowlist.
Scheduling runs on its own
Creating and rescheduling interview events in Google Calendar is allowed without confirmation, since it's the highest-volume task and easy to undo.
Candidate PII scrubbed in committee summaries
When the agent summarizes candidates for the hiring committee, home addresses, dates of birth, and photos are removed from the output.
Recruiting cost per requisition
Token and tool usage from recruiting agents is attributed per requisition, showing talent leadership which roles consume the most AI effort.
Section 19 screening at a bank
The FDIC Section 19 criminal-history screening tool is visible only to the HR compliance role, so recruiters' assistants can't run or read those checks.
Coordinator gets the reschedule override
The lead recruiting coordinator has a per-user allow to cancel interviews in bulk during a hiring freeze, while other recruiters stay on ask.
Sourcing sub-agent can't write to the ATS
The recruiting agent spawns a sub-agent to research candidates; it gets search tools only and does not inherit the parent's ability to move candidates between stages.
Withdrawing an offer is human-only
Rescinding an offer in Greenhouse is denied for every AI app, because the legal exposure needs a person making that call directly.
Nursing license checks through the proxy
A hospital's recruiting agent verifies RN licenses against state board lookup sites reachable only through the governed egress proxy, and each lookup is logged.
Outbound candidate email limits
Each recruiter's agent is capped at 200 outreach emails a day, protecting the company's sending domain from spam flags.
Check a new hiring manager's view
Before a new director starts interviewing, HR uses "Inspect as user" to confirm she can see scorecards for her reqs and no offer or comp tools.
Onboarding & Offboarding20 examples
Payroll setup is the gated step
The onboarding agent creates the Okta, Slack, and Google accounts on its own, but creating the employee record in ADP pauses for HR ops to confirm pay details.
Deprovisioning asks first
For a departure, the offboarding agent disables Okta, revokes sessions, and removes SaaS licenses only after an in-session confirmation, so a mix-up of two people with the same name gets caught.
Final paycheck needs two people
In states with same-day final pay, the payroll agent prepares the final check, and both the payroll lead and the HRBP must approve before it runs.
Mailbox handoff asks
Granting a manager access to a departed employee's mailbox pauses for confirmation and is recorded, since the mailbox may hold personal or privileged messages.
Workday hires: sandbox free, prod asks
The onboarding agent creates hire records freely in the Workday sandbox tenant connection, and on ask in the production tenant.
Laptop shipping to new hires
The agent orders and ships laptops through the approved hardware vendors, with a per-call cap and a monthly budget tied to the hiring plan.
I-9 and E-Verify stay with HR ops
The I-9 and E-Verify tools are visible only to the HR operations role, and submission is denied for agents even there; a person files it.
Welcome kits from approved vendors
The onboarding agent can order swag and welcome kits only from the approved merchandise vendors, under a per-hire cap.
Contractor onboarding through Deel
The contractor-onboarding tools in Deel are entitled to the people ops role only, keeping contractor setup separate from employee onboarding in Workday.
Physical badges stay human
Issuing or reactivating building badges in the access-control system is denied for all AI; facilities staff handle physical access in person.
Onboarding buddy without HR reach
The onboarding buddy agent answers new-hire questions from the handbook; when it spawns a sub-agent to look up benefits FAQs, the sub-agent has read-only docs access and no HRIS tools.
Access-removal record for SOC 2
The ledger shows each of the 40 SaaS deprovisioning calls for a leaver with timestamps, which becomes the SOC 2 evidence that access was removed within 24 hours.
Epic access only after credentialing
A new physician's Epic access tool is callable only inside the credentialing-complete workflow, not on request from chat.
Transfer data, don't delete it
On departure, the agent can transfer Google Drive ownership but deleting the user's Drive data is denied, keeping records for legal hold and retention schedules.
PEO's agent coordinates benefits
The PEO's benefits agent is peer-allowlisted to call in over A2A for enrollment status on new hires, but cannot read compensation or performance data.
Rehires get a fresh check
When a former employee is rehired, HR runs "Inspect as user" to confirm none of their old per-user overrides came back with the account.
Knowledge-transfer access expires
A departing engineer keeps read access to GitHub for their two-week notice period to hand over work, and it ends automatically on the last day.
Onboarding cost per hire
Agent token and tool usage for each onboarding run is attributed per hire, giving people ops a real per-hire figure for AI-assisted onboarding.
Exit interview notes de-identified
Exit interview summaries sent to leadership have names and identifying details removed before the agent returns them.
Compensation & Benefits20 examples
Payroll submission for payroll only
`adp.submit_payroll` is entitled to the payroll role and on ask even there, so a run is never submitted without a person confirming totals.
Large off-cycle payments go to the controller
An off-cycle bonus or correction over $5,000 routes to the controller for approval before the payroll agent submits it.
Employees view benefits, changes ask
Each employee's own assistant can read their benefits elections; mid-year changes pause for the employee to confirm the qualifying event.
Teller incentive changes on the record
Every change to branch incentive-comp plans at a bank goes through an agent with confirmation, and the ledger shows who changed what, the kind of record examiners may ask for on incentive programs.
Equity grants for finance and legal
Issuing option grants in Carta is visible only to finance and legal roles and asks first, since grants need board approval behind them.
Comp letters sent on confirmation
Sending 400 merit-increase letters at once asks first, so the comp team gets one last look before the letters go out.
Payroll test connection for dry runs
The payroll agent can run full payroll previews against the Gusto demo company connection, while the live company connection stays on ask.
Salaries invisible outside comp
Workday compensation tools are entitled only to the comp team and HRBPs; for everyone else, salary data isn't reachable by any AI app.
Read the full exampleComp benchmarking data purchases
Buying salary survey cuts from approved benchmarking vendors goes through spending policy with a vendor allowlist and annual budget.
Merit modeling in a sandbox
During the merit cycle, the comp agent reads current salaries and models scenarios in the sandbox, and write tools on the production Workday connection are denied.
401(k) contribution changes denied
Changing an employee's 401(k) deferral is denied for all AI; employees make that change themselves in the recordkeeper's portal.
Direct deposit changes blocked
Updating direct deposit bank details is denied for every agent, closing the payroll-diversion scam where a spoofed email asks the help desk to "update my account."
Broker's agent gets plan data, not people
The benefits broker's agent may call in over A2A for plan-level enrollment counts at renewal, and individual employee records aren't on its delegation.
Salaries scrubbed for non-comp users
When a manager's assistant pulls a team roster, salary fields are filtered out of the response unless the user holds the comp role.
Stipends under per-employee limits
The reimbursement agent pays home-office and wellness stipends through Ramp with a per-call cap and monthly budget per employee.
Comp review cost tracked
Token usage during the annual comp review is attributed to the comp team, so next year's budget reflects what AI-assisted modeling really costs.
Wage garnishments for payroll only
Garnishment setup and changes are visible only to the payroll role and logged in full, since errors carry court and legal exposure.
Pay equity analysis stays in-house
The pay-equity agent can't upload salary data to outside AI tools; its egress goes only through the governed proxy to the approved analytics vendor.
Preview an HRBP before comp cycle
Before merit season, HR runs "Inspect as user" on each HRBP to confirm they see compensation only for their own business unit's employees.
Employee Relations & Performance20 examples
Draft freely, submit on a yes
A manager's assistant can draft performance reviews in Lattice, but submitting the final rating pauses for the manager to confirm.
PIPs logged and confirmed
Issuing a performance improvement plan asks first and lands in the ledger with the drafts, giving HR a clear record if the PIP is later challenged.
Slack eDiscovery for legal only
Exporting Slack DMs for an investigation is denied for all AI except a per-user override for in-house employment counsel, who confirms each export.
Disciplinary letters to outside addresses pause
Sending a written warning via Gmail goes through, but the agent asks for approval when the recipient is outside the company domain.
Promotions route to the HR business partner
A promotion entered by a manager's assistant routes to the HR business partner for approval.
ER case files for ER only
Tools reading the employee relations case system are entitled only to the ER role, so no other HR assistant can surface an open investigation.
Investigation sub-agents get only notes
The ER agent spawns a sub-agent to build a timeline from interview notes; the sub-agent has no access to the HRIS or email tools the parent uses.
Hotline reports untouchable
Tools reading the ethics hotline system aren't entitled to anyone outside compliance, and every compliance read is logged, protecting whistleblower identity.
Managers see their own team only
Each manager's AI can read reviews for direct reports only; reviews for peer teams sit behind a role the manager doesn't hold.
Calibration data can't leave
Exporting calibration spreadsheets from the performance system is denied, so stacked rankings don't end up in someone's personal Drive.
Small-group survey results masked
Engagement survey breakdowns for teams under five people are suppressed in the agent's response so individuals can't be identified.
Outside counsel's agent, narrowly
Outside employment counsel's agent may call in over A2A to retrieve documents for one active matter, and other ER cases aren't on its delegation.
What the AI touched during litigation
When a former employee sues, the ledger shows every record any agent read or wrote about them, which counsel uses to answer discovery.
Goal-setting runs on its own
Employees' assistants can create and update their own OKRs in the performance tool without confirmation, since it's low-risk and frequent.
ADA accommodations stay confidential
Accommodation request records are entitled only to the accommodations specialist role, separate from the manager's view of approved schedules.
Review season cost by department
Usage attribution shows how many tokens each department's managers spent drafting reviews, which HR uses to plan licensing for next cycle.
Private 1:1 notes stay private
A manager's 1:1 notes workspace is reachable only by that manager's own sessions, through a per-user grant, and not by HR assistants.
No AI status changes to terminated
Changing an employee's Workday status to terminated is denied for every agent; the termination is entered by a person after the conversation happens.
Grievances in a union shop
At a unionized hospital, grievance-tracking tools are entitled only to labor relations, and every step is logged for the arbitration record.
Confirm a new manager can't see medical leave
HR runs "Inspect as user" on a newly promoted manager to verify she sees reviews and goals for her team but no leave or accommodation data.
Learning, Policies & Workforce Planning20 examples
Handbook changes publish on a yes
The policy agent drafts handbook updates in Confluence, but publishing a new version asks first and is limited to the HR policy role.
WARN notices need dual sign-off
Sending WARN Act notices requires approval from both legal and the CHRO before the agent sends them.
Coaching seats over threshold
Coaching platform seat purchases go through automatically up to $2,000, and larger orders need approval from the head of L&D.
Publishing shift schedules asks
At a hospital, the scheduling agent builds shift schedules in UKG, but publishing them pauses for the unit manager, since posted schedules trigger overtime and union notice rules.
Training records can't be edited
The LMS agent can assign courses and send reminders, but editing completion records is denied so compliance training can't be marked done after the fact.
BSA/AML training evidence
At a community bank, the ledger shows every annual BSA/AML training assignment and reminder the agent sent, answering the examiner's "how do you track completion" question.
Course purchases from approved providers
The L&D agent buys courses and seats only from approved training vendors, within each department's annual learning budget.
Headcount plan: finance reads, HR writes
In Adaptive Planning, finance assistants can read the headcount plan and only the HR planning role can change it.
Org chart changes stay manual
Changing reporting lines in Workday is denied for every agent, since one wrong supervisor change re-routes approvals across the company.
Acknowledgement reminders throttled
The agent's policy-acknowledgement nudges are capped at one per employee per week, so a new code of conduct doesn't generate daily spam.
Training content from approved sources
The course-building agent fetches external material only through the governed egress proxy to licensed content libraries, not from arbitrary websites.
Workforce analytics in aggregate only
The HR analytics role can query aggregated attrition and headcount views in Snowflake, while individual-level employee tables aren't entitled to that role.
RIF modeling for three named people
Reduction-in-force planning tools are available only through per-user grants to the CHRO, the CFO, and the head of legal; nobody else's AI knows they exist.
Nurse certification expirations
The credentialing agent flags nurses with BLS or ACLS certifications expiring in 30 days and opens tasks, but it can't change their scheduling eligibility in the staffing system.
AI tutor usage by department
Usage attribution for the internal AI tutor shows which departments use it and how much, guiding where to expand training budgets.
Open enrollment window enforced
Benefits change tools open to employees' assistants only during the November enrollment window and close automatically afterward.
Staffing agency's agent fills shifts
The travel-nurse staffing agency's agent is peer-allowlisted to call in over A2A to accept open shifts, but cannot see staff schedules or pay rates.
Policy drafting sub-agent stays in docs
The policy agent spawns a sub-agent to compare state leave laws; it gets research and docs tools only, not the publish tool its parent can request.
Translation through approved vendors
Translating the handbook into Spanish and Vietnamese uses only the approved translation API vendor, with a per-call cap and a monthly budget.
Succession plans invisible
Succession and high-potential lists are entitled only to the executive team role, so a manager asking their AI "who's my backup" gets nothing from the talent review data.
| Action | Who | Decision |
|---|---|---|
| draft_offer_letter | recruiting agent | allowed |
| send_offer_letter | recruiting agent | needs approval |
| view_salary_data | agents outside comp | denied |
| submit_payroll_change | payroll team's assistant | needs approval |
| deprovision_account | offboarding agent | needs approval |
| edit_training_record | any agent | denied |
Approval requests go to the HR business partner or payroll lead in Slack or the web inbox (Teams and email coming soon).
Illustrative: names no package, import or API call
AI agents in People & HR: common questions
Can an AI agent send an offer letter?
It can draft one. Sending is a separate tool, and marking it as needs approval means the offer leaves only after the recruiter or hiring manager says yes.
Who can reach compensation data through an agent?
Only the roles you scope it to. Salary and comp tools can be limited to the comp team, and Inspect as user shows exactly which tools a given person's agent would get.
Can an employee's own assistant change their benefits?
Policy is set per tool: an employee's assistant can call the benefits lookup, and the tool that makes a change can be set to needs approval, so each change waits for a person to approve it.
Can sensitive fields be hidden from an agent's results?
Redaction is coming soon, and the scenarios that need it carry the Coming soon label. Until then the tool that returns those fields stays out of scope for roles that shouldn't see them.
Set this policy up for your own agents.
Start with your own agents, your tools and the approvers you already have.
Other functions
Finance & Accounting
Payments, the close, treasury, budgets and payroll.
Software Development & Engineering
Code review, releases, production systems and on-call.
IT & Security Operations
Identity, devices, threat response, SaaS admin and cloud.
Sales & Revenue
Outreach, the CRM, quotes, contracts and renewals.
Marketing & Communications
Content, campaigns, paid media, social and the website.
Customer Support & Success
Ticket replies, refunds, privacy requests and escalations.
Legal, Risk & Compliance
Contracts, regulatory work, audit evidence, vendor risk and records.
Operations, Procurement & Facilities
Purchasing, vendors, inventory, facilities and projects.
Data, Analytics & AI Operations
Warehouse access, pipelines, agent-to-agent work, models and AI cost.

