Skip to main content

Human resources assistant

People are not workflow states.

A human resources assistant helps employees and HR teams find applicable policies, prepare requests and track review or correction. It can draft explanations tied to the correct employer, employment relationship and effective policy version. Werkon would validate this service workflow with protected employee information and accessible challenge routes; qualified owners retain entitlement and material employment decisions.

Ground the service in the applicable record

An employee statement, manager observation, verified record, policy term and qualified HR finding are different evidence. Keep policy, handbook, benefit-plan, collective-agreement and source-system versions tied to their audience and effective period. Calculate dates and balances from approved values while exposing missing facts and jurisdiction uncertainty.

Collect only the information needed for the request. Compartmentalize medical, disability, family, demographic, complaint, investigation, pay and performance details. A service classification cannot establish eligibility, and absence cannot become misconduct or a generated performance fact.

Preserve review and employee challenge

Qualified HR, legal, payroll, benefits, occupational-health and other authorized owners handle entitlement, accommodation, compensation, performance, discipline and employment decisions. Each reviewer receives only necessary information. Validate the destination, field and approval before any personnel or payroll write.

Keep submission, review, decision, notice, acknowledgment, record update, payroll effect and appeal separate. Employees need an accessible route to inspect, supplement, correct and challenge records. Preserve original evidence through reconsideration and correction, and evaluate service differences and harm without treating one statistical threshold as proof of fairness.

People-service boundary

Support the process without deciding the person.

HR work combines private facts, policy, administration and consequential judgment. Four boundaries keep service useful without hiding authority or exposing more than the task requires.

01

Identity, purpose, access, and service intake

Resolve the employer entity, person or representative, work relationship, location, role, purpose, audience, language and accessible path; classify the service without deciding it; collect only required information; and route emergencies, complaints and sensitive matters safely.

Required evidence: Employer entity, person and representative identifiers, relationship and status, location and jurisdiction candidate, role, purpose, service class, language, accessibility path, notice version, authority, source and field access, requested outcome, urgency, sensitivity, conflict, safe contact and denial or escalation reason.

02

Employee facts, policy, and eligibility evidence

Preserve employee assertions and original documents; retrieve current source records and applicable policies; separate supplied, observed, verified and derived facts; apply exact date, balance and condition rules; and expose missing, conflicting or stale evidence.

Required evidence: Original request and attachment, source and author, employee assertion, manager observation, system record, policy and plan version, effective period, collective term, job and work schedule, date and balance inputs, deterministic calculation, missing fact, contradiction, supersession and access boundary.

03

Administrative candidate and qualified review

Draft clarification, checklist, calculation, route, notice or field change from permitted evidence; show the exact rule and uncertainty; minimize reviewer data; and require the qualified owner for entitlement, accommodation, performance, pay, discipline, complaint, investigation and other material decisions.

Required evidence: Candidate type, cited rule and source facts, calculation trace, missing information, alternatives, sensitive-field mask, reviewer role and conflict check, decision authority, rationale, professional input where applicable, approved wording, rejected candidate and unresolved state.

04

Notice, record effect, reconsideration, and correction

Deliver an accessible, audience-appropriate notice; separate decision from acknowledgment; validate and approve each system update; reconcile payroll, benefit, time and case effects; and preserve correction, reconsideration, appeal, complaint, withdrawal, retention and deletion paths.

Required evidence: Decision and effective date, notice and language, accessible alternative, recipient and safe channel, delivery and acknowledgment, approved field diff, destination, idempotency key, write receipt, downstream effect, reconciliation, employee response, correction, reconsideration, appeal, complaint, supersession and retention state.

Request-to-record path

Keep private facts, policy, decision, and system effect distinct.

A completed form is not an entitlement and a policy match is not a decision. Each stage preserves the evidence and authority needed to challenge the next one.

  1. 01

    Authorize and minimize intake

    Verify the person or representative, employer entity, relationship, location, purpose and requested audience; provide language and accessibility paths; request only necessary fields; separate sensitive channels; and escalate safety, complaint or retaliation concerns without exposing them broadly.

    Owner
    HR service, identity, privacy and accessibility owners
    Evidence
    Identity, representation, entity, relationship, location, purpose, notice, language, accessibility path, required fields, collection reason, safe channel, sensitivity, emergency or complaint trigger, access decision and escalation.
  2. 02

    Assemble the applicable record

    Preserve the request, retrieve current employee and system facts, select policy and plan versions by entity, location and effective date, resolve source authority, retain contradictions and mark facts that require employee, manager, professional or records-owner confirmation.

    Owner
    HR records, payroll, benefits, policy and domain owners
    Evidence
    Original submission, employee and system records, source timestamps, policy, plan and collective versions, job and schedule, dates and balances, source authority, contradiction, missing information, superseded state and confirmation owner.
  3. 03

    Calculate and draft

    Run deterministic date, balance, threshold and field validations; draft questions, explanation, checklist, route, notice or record change with exact sources; expose uncertainty and alternatives; and prevent employee content from controlling tools or expanding access.

    Owner
    HR operations, payroll, benefits, data and security owners
    Evidence
    Rule inputs and version, calculation trace, field validation, candidate wording, source links, model and prompt versions, uncertainty, alternative, missing fact, sensitive mask, prompt-injection result and proposed destination.
  4. 04

    Make the qualified decision

    Route the minimum evidence to the authorized HR, legal, benefits, payroll, occupational-health, accessibility, labor-relations or management owner; check conflicts; record rationale; preserve human disagreement; and never let a score or generated recommendation become the decision.

    Owner
    Qualified HR, professional and employment-decision owners
    Evidence
    Review trigger, reviewer role and scope, conflict check, evidence viewed, rule interpretation, professional input, alternatives, employee context, decision, rationale, effective date, conditions, dissent, second review and unresolved issue.
  5. 05

    Notify, update, and maintain

    Issue an accessible notice through the safe channel, validate approved field changes, commit once, reconcile downstream effects, preserve acknowledgment separately and maintain employee correction, reconsideration, appeal, complaint, withdrawal, retention and deletion paths.

    Owner
    HR service, records, payroll, benefits and review owners
    Evidence
    Notice, reason and sources, recipient, language and accessible alternative, delivery, acknowledgment, approved field diff, write receipt, payroll or benefit effect, reconciliation, employee challenge, correction, appeal, complaint, supersession, retention and deletion event.

Authority map

Separate exact administration, model assistance, and employment authority.

A model can organize a request. It cannot determine a person's rights, credibility, health, performance, value or future at work.

01

Deterministic HR software

Software owns identity and entity binding, field permissions, policy and plan versions, source records, exact date and balance arithmetic, workflow states, destination validation, approved writes, notices, access logs, reconciliation, retention and correction history.

  • Employer, person, relationship, policy, plan, request, case and decision identifiers
  • Field-level access, purpose, audience, effective-period and retention checks
  • Date, service, schedule, balance, threshold and approved field calculations
  • Review, notice, acknowledgment, write, reconciliation, appeal and correction receipts
02

Bounded AI assistance

AI can propose service classification, clarifying questions, document extraction, policy retrieval, checklists, summaries and notice wording from permitted evidence, but every output remains a candidate with visible sources, uncertainty and sensitive-data limits.

  • Service, urgency, language and safe-routing candidates
  • Source-linked extraction, policy passage and missing-information candidates
  • Plain-language explanation, checklist, notice and record-summary candidates
  • Conflict, sensitive-field, inconsistent-treatment and review-trigger candidates
03

Human HR and professional authority

Qualified people own legal and policy applicability, entitlement, accommodation, health interpretation, performance, pay, promotion, discipline, termination, complaint, investigation, exceptions, appeal, correction and every material employment decision.

  • Jurisdiction, worker status, collective terms and policy interpretation
  • Leave, benefit, accommodation and occupational-health decisions
  • Performance, compensation, mobility, discipline and employment decisions
  • Complaint, investigation, appeal, correction, retaliation response and stop authority

HR-assistant components

Build a private case path, not a profile about a person.

People, policies, plans and employment records change independently. Four components preserve the context and permissions behind every HR action.

01

Identity, relationship, purpose, and access registry

Version employer entity, person and representative identity, work relationship, role, location, jurisdiction candidate, team, manager, safe contact, language, accessibility needs, purpose, audience, delegated authority, field permissions and revocation.

Operating contract: Authentication is not HR authority, manager is not entitled to every field, employment status is not consent, one entity's policy does not govern another, prior access is not current need and a representative's authority must be explicit and revocable.

02

Employee, policy, plan, and evidence ledger

Preserve requests, attachments, employee assertions, manager observations, system records, policy and plan versions, collective terms, job and schedule, dates, balances, professional evidence, provenance, classification, access, effective periods and corrections.

Operating contract: Employee assertion is not verified fact, manager note is not finding, generated summary is not evidence, policy text is not applicability, medical information is not general HR data and a corrected record must not erase the original or its challenge history.

03

Service, review, and decision ledger

Bind service class, deterministic calculations, source-linked candidates, missing facts, sensitive masks, review triggers, qualified reviewers, conflicts, professional input, alternatives, decision, rationale, effective date, conditions and dissent.

Operating contract: Classification is not eligibility, calculation is not entitlement, score is not fairness, policy match is not judgment, reviewer click is not expertise and the assistant cannot hire, promote, rate, pay, discipline, accommodate, dismiss or resolve a complaint.

04

Notice, system-effect, and review ledger

Record accessible notices, safe recipients and channels, delivery, acknowledgment, approved field diffs, destinations, write receipts, payroll, benefit, time and case effects, reconciliation, employee correction, reconsideration, appeal, complaint, withdrawal, supersession and retention.

Operating contract: Sent is not received, acknowledgment is not agreement, personnel update is not payroll effect, approval is not waiver of review, correction must reach affected systems and complaint or appeal access must remain independent of the original decision path.

Delivery path

Prove one employee service before widening the casework.

A narrow service with real privacy and decision boundaries exposes more than a broad HR chatbot demo. Five steps test the full operating path.

  1. 01

    Choose the service and population

    Select one recurring administrative service, named employer entities and eligible user groups, qualified owners, safe manual path and prohibited decisions; record jurisdictions, collective contexts, languages and accessibility needs.

  2. 02

    Define the minimum record

    Map required facts, sources, classifications, field permissions, policy and plan versions, effective periods, calculations, sensitive channels, notices, retention and employee correction rights before connecting systems.

  3. 03

    Build the reviewable workflow

    Implement source-linked intake, exact calculations, candidate drafting, qualified review, accessible notice, approved field diffs, idempotent writes, reconciliation and separate correction, appeal and complaint paths.

  4. 04

    Test people-shaped failures

    Test wrong entity, representative, policy and effective date; sensitive leakage; inaccessible intake; ambiguous facts; conflicting records; proxy traits; inconsistent treatment; harmful wording; unauthorized writes; retaliation risk and correction failure.

  5. 05

    Operate and review outcomes

    Assign service, source, privacy and decision owners; review access, completion, review burden, consistency and harm across relevant groups; rehearse policy change, reconsideration, correction, appeal, revocation, incident and retirement.

Release controls

Six controls before an HR answer can change a record.

Administrative speed is not a safe outcome when the record affects a person. These controls keep purpose, privacy, judgment and review visible.

The person, entity, and purpose are verified
Resolve employer, identity, representation, relationship, location, purpose, audience and effective date; apply least privilege; expire access; and block requests whose identity, authority or safe channel is unresolved.
Collection and display are minimized
Request only fields needed for the named service, compartmentalize medical and other sensitive data, mask reviewer views, prevent derived-trait exposure and keep credentials and private records outside model context where possible.
Policy and calculations are inspectable
Use the correct policy, plan, collective term, entity, jurisdiction candidate and effective period; preserve exact inputs and arithmetic; show conflicts and never turn a generic policy match into an individual determination.
Qualified people own consequential decisions
Require the authorized HR, legal, benefits, payroll, accessibility, occupational-health, labor-relations or management owner, with conflict checks and rationale, for every material effect on a person.
Notice and system effects stay separate
Validate recipient, language, accessibility, safe channel, effective date, destination and exact field diff; distinguish delivery and acknowledgment; reconcile payroll, benefit, time and case effects before closure.
Challenge and correction remain available
Give the person a usable path to inspect, correct, supplement, reconsider, appeal or complain; preserve non-retaliation controls and history; propagate approved changes and stop the workflow when harm or unauthorized access appears.

Outcome evidence

Measure safe service and fair process, not tickets closed.

A faster queue can conceal inaccessible service, sensitive leakage or inconsistent decisions. Evidence must include the people affected and the work shifted to reviewers.

Baseline

  • Service types, populations, entities, locations, policies, qualified owners, manual paths and consequence tiers
  • Current intake completion, time to clarification, review, decision, notice, update and reconciliation with human effort
  • Current missing, duplicate, stale and conflicting records, policy exceptions, inaccessible paths and correction volume
  • Current access incidents, sensitive-data exposure, inconsistent treatment, disputes, appeals, complaints and unresolved harm

Outcome evidence

  • People reaching an accessible service path and receiving complete, source-linked information for the correct context
  • Requests resolved with minimum data, complete evidence, qualified decision and reconciled record effects
  • Time to safe clarification, decision, notice, system effect, correction and appeal, including reviewer effort
  • Consistency and service outcomes across relevant groups with documented reasons, uncertainty and human review

Guardrails

  • Cross-entity access, sensitive-field leakage, excessive collection, inferred traits, unsafe channels and credential exposure
  • Wrong policy or date, fabricated facts, hidden contradictions, inaccessible notices and unqualified interpretation
  • Automated significant decisions, proxy discrimination, inconsistent treatment, retaliation, reviewer overload and harmful reliance
  • Unauthorized writes, payroll or benefit mismatches, blocked challenge, correction misses, retention failures and unresolved incidents

Fit test

Use this pattern when service and decision authority are already named.

Good reason to begin

  • One recurring employee service has defined entities, populations, policies, source systems, qualified reviewers and a safe manual path.
  • Minimum data, sensitive fields, access by purpose, policy versions, effective dates and system effects can be made explicit.
  • People need accessible, source-linked guidance and transparent case status without surrendering correction, appeal or human contact.
  • The organization can test privacy, inconsistent treatment, review burden, accessibility, correction reach and harm before release.

Resolve before beginning

  • No one owns policy applicability, employee records, sensitive-data access, qualified decisions, notices, appeals or corrections.
  • The request assumes that all HR data may be pooled or that manager access covers medical, complaint, demographic and case details.
  • Success is defined as ticket reduction or manager convenience without service access, data minimization, consistency or employee outcome evidence.
  • The assistant is expected to rank people or decide hiring, pay, performance, promotion, accommodation, discipline or termination without accountable human authority.

Source basis

Sources behind the control model.

  • 01

    European Union

    Regulation (EU) 2016/679, General Data Protection Regulation

    The official regulation includes personal-data principles in Article 5, special-category conditions in Article 9, collection-time information in Article 13, access rights in Article 15 and Article 22 provisions for decisions based solely on automated processing that produce legal or similarly significant effects, subject to scope, exceptions and safeguards. Applicability, controller role, lawful basis, employment-law interaction, national rules, exceptions and remedies require qualified analysis. It is not a global HR design, consent mandate, universal ban on automation or compliance certificate.

  • 02

    Electronic Code of Federal Regulations

    29 CFR Part 1607, Uniform Guidelines on Employee Selection Procedures

    The continuously updated eCFR text, current through 24 August 2026 when reviewed, sets US federal principles for covered tests and selection procedures used in specified employment decisions, including adverse-impact analysis, job-related validity, alternatives, operational use and documentation. The eCFR is authoritative but unofficial. The part is limited by covered laws, users, groups and selection procedures and does not govern every HR service, protected characteristic, jurisdiction or employment decision or validate this pattern.

  • 03

    United States Department of Labor

    Fact Sheet 28D, FMLA employer notification requirements

    The Wage and Hour Division guidance distinguishes general, eligibility, rights-and-responsibilities and designation notices for covered employers and qualifying US FMLA contexts, including stated timing, content, language and accessibility considerations. It is jurisdiction-specific guidance, not a universal leave rule, individual eligibility or designation decision, legal advice, proof of delivery or authority for an assistant to decide leave.

  • 04

    National Institute of Standards and Technology

    NIST AI 600-1, Generative AI Profile

    NIST published the voluntary cross-sectoral profile on 26 July 2024 and updated its landing page in April 2026. It addresses risks including confabulation, privacy, information integrity, harmful bias, human-AI configuration and provenance with evaluation and monitoring actions. It does not define employment law, prove fairness, authorize an employment decision, validate a model or certify compliance or outcomes.

[ WORKFLOW / SYSTEMS AUDIT ]
THE FIRST ENGAGEMENT

Start with one real workflow

A Systems Audit is the usual starting point. If the opportunity is already clear, we can move directly into a focused build.

Show Us the WorkflowStart with the free automation readiness checklist

OBSERVEQUANTIFYDECIDEBUILD

Continue with the work

Choose the next useful path.

Stay with the operating question. The technology can wait until the work is clear.

01

Explore governed AI agents

Compare other bounded agent patterns and the evidence, authority and recovery contracts they require.

Open path02

Design HR and workforce automation

Map employee service, workforce records, qualified decisions, system effects and review before automating steps.

Open path03

Build a governed AI agent

Enclose model-assisted work with explicit identity, permissions, tools, approvals, evidence and stop paths.

Open path04

Build the service application

Create an owned workflow around users, business rules, records, integrations and long-term operation.

Open path