ai-enabled-hr-talent-automation

AI Guardrails Policy

Title: AI Guardrails Policy Version: 1.0 Owner: [TENANT_CONFIGURATION_REQUIRED — AI Governance Lead] Status: Draft Last reviewed: 2026-09-07 Next review: [TENANT_CONFIGURATION_REQUIRED] Reviewers: AI Governance, Security, HR, Legal

Purpose and scope

Defines the non-negotiable boundaries on AI behavior across the platform, enforced structurally (in the Guardrail Pipeline and Workflow Engine — see ADR-002), not merely by prompt instruction.

What AI may do

Recommend, extract, classify, summarize, draft, route, and alert — always producing structured, reviewable output.

What AI must never do (non-negotiable)

Forbidden action	Enforcement mechanism
Autonomously reject or select a candidate	Workflow Engine requires a human `interview_feedback` or `progression-decision` record; no skill has a code path to write this directly
Autonomously release an offer	`offer.send` only proceeds if `offer.approve` was previously recorded by an `hr_approver`
Determine or alter compensation figures	Offer schema requires a `compensationRef` into an external comp system; no free-text AI-generated figure is accepted
Close a discrepancy or grant an exception	`discrepancy.resolve` requires an `ApprovalDecisionRequest` from an `hr_approver`
Create an Employee ID	`employee-conversion` endpoint requires a recorded approval and a fully-passed gate checklist

See human-approval-matrix.md for the complete gate list.

Guardrail pipeline stages

flowchart LR
    A[Untrusted input: CV/JD/feedback/RAG/MCP response] --> B[Input Sanitization]
    B --> C[Prompt-Injection Detection]
    C --> D[Skill Execution - least privilege tools]
    D --> E[Output Schema Validation]
    E --> F[PII/Redaction Check]
    F --> G[Confidence Threshold Check]
    G -->|Below threshold| H[Route to human review]
    G -->|At/above threshold| I[Return structured result to HR Core API]
    I --> J[Approval-Matrix Check before any state mutation]

Confidence thresholds (configurable — see `config/schemas/rag-config.schema.json` and `model-routing.schema.json`)

Skill	Default minimum confidence	Below-threshold action
CV field extraction	0.75	Route to manual review queue
Candidate matching score inclusion	0.60	Exclude candidate from recommendation, do not silently include as a weak match
Document verification	0.80	Flag as `needs_review`, never auto-pass
RAG policy answer	0.65 (retrieval relevance)	Return no-answer fallback with citation prompt to consult HR directly

Required human-review triggers

Any output below the confidence threshold above.
Any detected prompt-injection attempt (see prompt-injection-defense.md).
Any candidate flagged as a possible duplicate.
Any discrepancy above medium severity.
Any output that would (directly or indirectly) touch a forbidden action above.

Governance of prompts and models

Prompt and model changes follow versioning, approval, testing, and rollback rules in prompt-management.md and model-routing-and-cost-controls.md. No prompt/model change ships to production without passing the evaluation suite in ai-evaluation-strategy.md.

Change control

Version	Date	Author	Change
1.0	2026-09-07	Documentation package generation	Initial creation

This site is open source. Improve this page.