HR handles some of the most sensitive data in any organization — salaries, performance reviews, disciplinary records, medical accommodations, personal information. Sending that to a cloud AI service creates real compliance and privacy risks.

Local AI eliminates that risk entirely. The model runs on your hardware, processes data locally, and nothing is transmitted externally.

Why It Matters for HR

• Employee privacy — performance data, salary info, and personal details stay on-premises
• GDPR/compliance — easier to demonstrate data doesn’t leave your infrastructure
• No vendor risk — no third-party data processing agreements needed
• Unlimited use — generate 50 job descriptions in a row without rate limits
• No per-seat costs — free for the entire HR team

Setup (15 Minutes)

# macOS or Linux
curl -fsSL https://ollama.com/install.sh | sh

# Windows: download from ollama.com

Which Model to Use

Model	RAM Needed	Speed	Best For
llama3:8b	8GB	Fast	Emails, quick JDs, simple communications
qwen2.5:14b	16GB	Medium	Policy documents, detailed reviews, analysis
qwen2.5:32b	32GB	Slower	Complex policies, long handbooks, nuanced writing

ollama pull qwen2.5:14b
ollama run qwen2.5:14b

Recommendation: qwen2.5:14b for most HR work. Policy documents and performance reviews need nuance that the 8b model often misses — it tends to produce generic, corporate-sounding output. The 14b model writes more naturally.

HR Workflows With Real Examples

Performance Review Drafts

Prompt:

Write a performance review for a [job title] who has been in the 
role for [duration].

Strengths:
- [specific strength with example]
- [specific strength with example]

Areas for improvement:
- [specific area with context]

Overall rating: [Exceeds/Meets/Below expectations]

Tone: constructive, specific, and encouraging. Avoid vague phrases 
like "good team player" — use concrete examples. Include 2-3 
development goals with measurable outcomes.

What to expect: Well-structured reviews with specific language. The key is giving the model specific examples — “consistently meets deadlines on the Q3 project” produces much better output than “good at time management.” The model is excellent at turning your bullet points into professional paragraphs.

Batch processing: During review season, you can draft 20+ reviews in a day. Paste each employee’s bullet points, get a draft back, review and personalize. What used to take a week takes a day.

Job Descriptions

Write an inclusive job description for a [role title] at a 
[company type/size] company.

Department: [department]
Reports to: [title]
Location: [remote/hybrid/office in city]
Salary range: [range]

Key responsibilities: [list 3-5 main duties]
Must-have qualifications: [list]
Nice-to-have qualifications: [list]

Requirements:
- Use gender-neutral language throughout
- Avoid unnecessary requirements that could exclude qualified candidates
- Include benefits and perks
- Include an equal opportunity statement
- Keep it under 600 words

What to expect: Clean, inclusive JDs that avoid common bias patterns. The model is good at catching gendered language (“rockstar,” “ninja”) and unnecessary requirements (“must have 10 years experience” for a mid-level role). Always review the qualifications section — the model sometimes adds requirements you didn’t specify.

Policy Documents

Draft a [remote work / PTO / social media / AI usage / expense] 
policy for a company of [size] employees in [industry].

Include:
- Purpose and scope
- Eligibility
- Detailed guidelines
- Exceptions and approval process
- Enforcement and consequences
- Effective date and review schedule

Tone: professional but approachable — this will be read by all 
employees, not just HR. Avoid legalese where possible.

What to expect: Comprehensive first drafts that cover the standard sections. The model is particularly good at AI usage policies (ironic, but true) and remote work policies. For anything with legal implications (termination, discrimination, leave), always have legal review the final version.

Interview Questions

Generate 10 behavioral interview questions for a [role] position.
Focus on these competencies: [list 3-4 competencies].

For each question:
1. The question itself
2. What you're evaluating
3. What a strong answer looks like
4. One follow-up probe

Important: Avoid questions that could create legal liability 
(age, family status, religion, disability, national origin).

What to expect: Well-structured questions with useful evaluation criteria. The legal compliance note is important — the model generally avoids obviously illegal questions, but review carefully. The follow-up probes are often the most valuable part.

Employee Communications

Draft an all-hands email announcing [organizational change: 
restructuring / new benefit / policy change / leadership change].

Context: [what's happening and why]
Impact on employees: [what changes for them]
Timeline: [when this takes effect]

Tone: transparent and empathetic. Acknowledge concerns directly. 
Don't use corporate euphemisms ("right-sizing," "synergies"). 
Include who to contact with questions and next steps.

What to expect: Clear, human-sounding communications. The model is good at striking the right tone for sensitive announcements — it avoids the corporate-speak that makes employees distrust the message. Review the “acknowledging concerns” section to make sure it addresses the actual concerns your employees will have.

Onboarding Checklists

Create a comprehensive onboarding checklist for a new [role type] 
employee. Company size: [size]. 

Include phases:
- Pre-start (before day 1)
- Day 1
- Week 1
- First 30 days
- First 90 days

For each item include: task, responsible person (HR/Manager/IT/New hire), 
and deadline. Include both administrative tasks (paperwork, systems access) 
and cultural tasks (team introductions, mentor assignment).

Separation/Offboarding Documentation

Draft a [voluntary resignation acknowledgment / termination letter / 
exit interview questionnaire] for an employee in [state/country].

Include relevant information about:
- Final pay and PTO payout
- Benefits continuation (COBRA if applicable)
- Return of company property
- Non-compete/NDA reminders (if applicable)
- References policy

Tone: professional and respectful regardless of circumstances.

Is Local AI Good Enough? Honest Comparison

Task	Local AI (14b)	ChatGPT Plus	Claude Pro
Performance reviews	90% as good	Very good	Best (most natural tone)
Job descriptions	90% as good	Very good	Very good
Policy documents	85% as good	Very good	Best (most thorough)
Employee emails	95% as good	Excellent	Excellent
Interview questions	90% as good	Very good	Very good
Legal compliance	Review always needed	Review always needed	Review always needed

The honest truth: For day-to-day HR writing, local AI is close to cloud AI quality. The biggest gap is in nuanced policy writing where Claude’s longer context window and instruction-following give it an edge. But for 80% of HR tasks — reviews, JDs, emails, checklists — local AI is more than good enough.

Security Best Practices

For HR data specifically, go beyond the basics:

1. Dedicated machine — run the AI on a workstation that only HR has access to. Not a shared office computer.
2. Enable disk encryption — FileVault (Mac) or BitLocker (Windows). Non-negotiable for HR data.
3. Don’t include SSNs or sensitive identifiers — use placeholders in prompts and fill in manually
4. Clear conversation history — if using a UI like Open WebUI, configure auto-delete after each session
5. Network isolation — keep the AI machine off shared network drives that contain employee records. Copy only what you need.
6. Access logging — if multiple HR staff use the system, enable logging to track who used it and when

For full sandboxing with Docker and VMs, see How to Sandbox Local AI Models.

Troubleshooting

“Performance reviews sound too generic”

• The #1 fix: give more specific input. “Good at communication” → “Led the Q3 client presentation that resulted in a contract renewal.” The model can only be specific if you are.
• Use the 14b model minimum. The 8b model defaults to corporate boilerplate.
• Add “Avoid generic phrases. Every statement should reference a specific behavior or outcome.”

“Policy documents miss important sections”

• Be explicit about what sections to include. Don’t rely on the model to know your company’s needs.
• For legally sensitive policies, always start with a template from your legal team and use AI to draft the content within that structure.
• Run the output through a second prompt: “Review this policy for gaps, inconsistencies, or missing sections.”

“Model is too slow during review season”

• Review season means batch processing. Use the 8b model for quick drafts and the 14b for final versions.
• Pre-write your system prompt once and reuse it: “You are an HR professional writing performance reviews. Tone is constructive and specific. Always include measurable development goals.”
• Process reviews in batches of 5-10, not one at a time.

“Out of memory”

• 8b: ~6GB RAM needed. 14b: ~12GB. 32b: ~24GB.
• During review season, close everything except Ollama and your text editor.
• If your machine has 8GB total, the 8b model is your only option — but it’s still better than writing from scratch.

Team Setup

For an HR department with multiple people, run Ollama on a shared server:

docker run -d \
  --name open-webui \
  -p 3000:8080 \
  -e OLLAMA_BASE_URL=http://host.docker.internal:11434 \
  ghcr.io/open-webui/open-webui:main

Access at http://[server-ip]:3000. Set up individual accounts so you can track usage and maintain audit trails.

The Bottom Line

HR writing is repetitive and time-sensitive — reviews, JDs, policies, communications. Local AI gives you instant first drafts without sending employee data to the cloud. The 15-minute setup pays for itself the first time you draft a batch of performance reviews. And during review season? It’s a lifesaver.

Related reading: BambooHR AI Review · Best AI Tools for HR · AI for Engagement Surveys · AI for Employer Branding

🛠️ Try our free tools: Job Description Generator · Performance Review Generator · Interview Question Generator · Onboarding Checklist Generator · Policy Document Generator