architecture

HR AI Agents

AI agents for HR: resume screening, interview question generation, and policy Q&A. Includes off-the-shelf options, custom build approaches, and training examples with important bias safeguards.

Overview

HR tasks span recruiting, onboarding, policy, and employee support. Many involve document processing, communication drafting, and information synthesis.

Efficiency Potential: 5-7 hrs/week at Adoptive tier

Agent Summary

Agent Efficiency Target Complexity Priority
5 min β†’ 30 sec per resume Medium High
30 min β†’ 5 min Low Medium
70% self-served Medium High

Resume Screening Assistant

Problem Statement

Recruiters review hundreds of resumes per role. Initial screening is pattern-matching that AI can accelerate while maintaining (and improving) consistency.

Efficiency Target: Resume screen from 5 min β†’ 30 sec per resume; better consistency

Important Safeguards

⚠️  BIAS MITIGATION REQUIREMENTS

1. Never use demographic indicators in screening
2. Validate AI decisions with human review
3. Track outcomes for bias detection
4. Maintain human decision authority
5. Regular audits of screening patterns
6. Document all criteria and decisions

Training Example: Resume Screener

Screen this resume against job requirements.

JOB REQUIREMENTS:
{{job_description}}

MUST-HAVE CRITERIA:
{{must_haves}}

NICE-TO-HAVE CRITERIA:
{{nice_to_haves}}

DISQUALIFIERS:
{{disqualifiers}}

RESUME:
{{resume_text}}

EVALUATE:

## Candidate Screening: {{candidate_name}}

### Quick Verdict
[Strong Match / Potential Match / Not a Match / Needs Discussion]

### Must-Have Assessment
| Requirement | Met? | Evidence |
|-------------|------|----------|
[Each must-have with specific resume evidence]

### Nice-to-Have Assessment
| Criteria | Present? | Evidence |
|----------|----------|----------|
[Each nice-to-have with evidence]

### Disqualifier Check
[Any disqualifiers present?]

### Strengths
- [Top 2-3 relevant strengths with evidence]

### Concerns/Questions
- [Gaps or items to probe in interview]

### Suggested Interview Focus
- [What to validate based on resume]

### Bias Check
[Evaluation based only on job-relevant criteria; no demographic inference]

CONFIDENCE: [High/Medium/Low] - [why]

Architecture Pattern

Consider ENT-FINE-TUNING for consistent scoring:

# Training data for resume scoring
{
    "messages": [
        {"role": "system", "content": "Score resume against criteria."},
        {"role": "user", "content": "Resume: [text] Criteria: [requirements]"},
        {"role": "assistant", "content": json.dumps({
            "verdict": "Strong Match",
            "must_haves_met": 5,
            "must_haves_total": 5,
            "confidence": 0.92
        })}
    ]
}

Interview Question Generator

Problem Statement

Creating good interview questions takes time. Questions should be role-specific, behavior-based, and consistent across candidates. Most interviewers wing it.

Efficiency Target: Interview prep from 30 min β†’ 5 min; better question quality

Training Example: Interview Guide

Generate an interview guide for this role.

ROLE: {{role_title}}
LEVEL: {{level}}
KEY COMPETENCIES:
{{competencies}}

TEAM CONTEXT:
{{team_context}}

CANDIDATE BACKGROUND:
{{candidate_summary}}

INTERVIEW STAGE: {{stage}} (phone screen, technical, behavioral, final)
INTERVIEW DURATION: {{duration}}

GENERATE:

## Interview Guide: {{candidate_name}} for {{role_title}}

### Interview Objectives
[What we're assessing in this round]

### Competency: {{competency_1}}

**Question 1:**
"{{behavioral_question}}"

*What to listen for:*
- Strong answer: [indicators]
- Weak answer: [red flags]

*Follow-ups:*
- "What would you do differently?"
- "How did you measure success?"

**Question 2:**
[Additional question for same competency]

### Competency: {{competency_2}}
[Repeat structure]

### Role-Specific/Technical Questions
[Based on job requirements]

### Candidate-Specific Questions
[Based on their resume - gaps to probe, claims to validate]

### Sell Questions (if candidate is strong)
[Questions to let them evaluate us]

### Wrap-Up
- Time for candidate questions: [X min]
- Next steps to communicate

### Evaluation Rubric
| Competency | 1-Poor | 2-Below | 3-Meets | 4-Exceeds |
|------------|--------|---------|---------|-----------|
[Quick reference for scoring]

Policy Q&A Bot

Problem Statement

HR answers the same policy questions repeatedly. Employees can't find information in lengthy handbooks. HR time goes to answering routine questions vs. strategic work.

Efficiency Target: 70% of policy questions self-served; HR focuses on complex cases

Architecture Pattern

Uses ENT-RAG for policy retrieval:

β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”     β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”     β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”
β”‚ Employee     │────►│ Embedding    │────►│ Policy       β”‚
β”‚ Question     β”‚     β”‚ Search       β”‚     β”‚ Chunks       β”‚
β””β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”˜     β””β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”˜     β””β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”˜
                                                 β”‚
                                                 β–Ό
                                          β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”
                                          β”‚ Answer       β”‚
                                          β”‚ + Citation   β”‚
                                          β””β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”˜

Data Sources:

  • Employee handbook
  • Benefits documentation
  • Leave policies
  • Expense policies
  • IT/Security policies
  • Manager guidelines

Training Example: Policy Q&A

Answer this HR policy question using only the provided policy documents.

QUESTION: {{employee_question}}

RELEVANT POLICY SECTIONS:
{{retrieved_policy_chunks}}

RULES:
1. Only answer based on provided policy text
2. If the answer isn't in the policies, say so
3. Always cite the specific policy section
4. If the situation is complex, recommend talking to HR
5. Never give legal advice
6. Be helpful but accurate

GENERATE:

## Answer

[Direct answer to the question]

### Policy Reference
> [Relevant quote from policy]
> β€” {{policy_name}}, Section {{section}}

### Additional Context
[If helpful - related policies, common follow-ups]

### Need More Help?
[If complex or edge case]
Contact HR at {{hr_contact}} for situations involving:
- [exceptions]
- [approvals needed]
- [sensitive matters]

Implementation

from langchain.retrievers import EnsembleRetriever

# Hybrid search for better policy retrieval
bm25_retriever = BM25Retriever.from_documents(policy_docs)
vector_retriever = vectorstore.as_retriever(search_kwargs={"k": 10})

ensemble = EnsembleRetriever(
    retrievers=[bm25_retriever, vector_retriever],
    weights=[0.4, 0.6]
)

# Chain with citation
qa_chain = RetrievalQA.from_chain_type(
    llm=llm,
    retriever=ensemble,
    return_source_documents=True,
    chain_type_kwargs={
        "prompt": policy_qa_prompt
    }
)