ChatGPT for HR: What It Is Great At, and Where It Breaks

ChatGPT for HR is two different stories depending on the task. For drafting (job descriptions, policy first drafts, interview questions, emails), it is a genuine accelerator that has changed how People work gets done. For answering (what is our parental leave policy, who approves this), it is a confident liability, because it does not know your company and will not admit it. This guide maps the line precisely, so you get the acceleration without the incidents.

Where ChatGPT genuinely helps HR

The pattern across every good use: blank-page tasks where a human reviews the output before it matters.

• Job descriptions and postings: a strong first draft from a role summary in seconds.
• Policy first drafts: a generic remote work or conduct policy skeleton to edit toward your reality.
• Interview and survey questions: solid starting sets to refine.
• Emails and announcements: tone-shifting, shortening, untangling difficult messages.
• Brainstorming: onboarding ideas, recognition programs, meeting agendas.

Used this way, it is the most useful intern HR has ever had: fast, tireless, and always needing review.

The four places it breaks

1. It does not know your company. Ask it your PTO policy and it generates a plausible PTO policy, not yours. The danger is precisely that the answer sounds authoritative. For anything where the true answer lives in your documents, an ungrounded model is the wrong tool, the same gap covered in Guru vs Slack AI vs Sakha.

2. Privacy. Pasting employee or candidate personal data into a consumer AI tool can violate privacy obligations and your own policies. This is the number one rule in any AI usage policy: real people's data stays out.

3. No memory, no workflow. Every chat starts from zero. It cannot run your onboarding, track who completed what, or remember last week's decision. It is a conversation, not a system.

4. No sources. Its answers cite nothing, so errors look identical to facts. For HR work, where wrong answers have consequences, unsourced confidence is the failure mode.

The rule of thumb

Use ChatGPT where the human is the system: you prompt, it drafts, you review, you own the result. Use a grounded system where the tool must be trusted on its own: answering employees, running processes, tracking state. The mistake is letting the drafting tool drift into the answering job, which happens naturally the day a busy manager asks it a policy question and ships the plausible answer.

This is why "just use ChatGPT" is not an onboarding strategy, however good the model is. Onboarding needs memory (who is on day 4), schedule (what goes out today), grounding (your policies, not generic ones), and visibility (what got done). Those are system properties, not model properties, and they sit much closer to the agent pattern covered in AI agents in HR.

Where Sakha sits in this split

Sakha is the grounded half. It answers employee and new hire questions from your actual company knowledge base, with the source cited on every answer, and says honestly when it does not know rather than improvising. It runs the onboarding flow on schedule, remembers state, and tracks completion, the system properties a chat window cannot have. And on the drafting side, its policy generator does what ChatGPT drafting does for policies, but structured, reviewed for gaps, and published straight into the knowledge base where employees can query it.

The practical setup for a modern People team: ChatGPT (governed by a clear policy) for drafting, Sakha for everything that answers, runs, or remembers. See AI onboarding vs traditional onboarding for the broader division of labor.