HR teams carry a heavy writing and answering load: job descriptions, policy questions, offer letters, onboarding guides, internal communications. Claude handles this kind of work well, and much of it is genuinely low-stakes. But HR is also where AI can do the most durable harm — in decisions about who gets hired, promoted, or let go. This article separates the two zones plainly, because the reference design for HR is less about architecture and more about scope.
The safe wins: writing and policy Q&A
Start where mistakes are cheap and review is natural. Job descriptions: Claude drafts from a role profile and your template, and can flag unnecessarily narrow requirements or gendered, exclusionary phrasing for a human to reconsider — a use of the model that works against bias rather than encoding it. Policy Q&A: an assistant grounded in your actual handbook and benefits documents (the retrieval pattern from the internal knowledge assistant article) answers "how much parental leave do I get?" with citations, and is instructed to route anything personal, contested, or disciplinary to a human. Communications and onboarding: announcement drafts, guide restructuring, interview-question preparation against a defined competency model. In all of these, an HR professional reviews before anything is sent or published, and the model never has access to more employee data than the task requires.
Screening support: proceed only with legal review
Using any AI system on candidates is different in kind, for three reasons. First, bias risk is real: language models learn from human-written text and can reproduce its patterns, including associations about names, schools, employment gaps, and phrasing that correlate with protected characteristics. A model can filter systematically and invisibly in ways a biased human cannot. Second, employment decisions are regulated: multiple jurisdictions now specifically govern automated employment decision tools, with obligations that can include notice to candidates, impact assessments, and audits — and the list of such laws is growing. Third, the decision must remain human as a matter of policy, not just compliance theater: a human who rubber-stamps a model's ranking has delegated the decision in every way that matters.
So the rule is: no use of Claude that scores, ranks, filters, or recommends candidates goes live without review by employment counsel — for every jurisdiction you hire in — plus a documented human-decision process and an audit trail. What can be defensible, with that review, is assistive use: extracting stated qualifications from a CV into a structured summary a recruiter reads in full, or standardizing interview-note formats. Even then, humans read the underlying material, the model's output is an index into it rather than a verdict on it, and you monitor outcomes for disparities.
A reference design for the safe zone
The architecture follows the standard split. Claude drafts and summarizes; deterministic code moves data between your HR systems and enforces access control — payroll, health, and disciplinary records stay out of model calls entirely unless a specific, reviewed task requires them; and humans hold every decision that affects a person's pay, status, or employment. Employee and candidate data is personal data everywhere and sensitive data in many places, so run this inside your governed cloud boundary (Bedrock, Vertex AI, Foundry, or Claude Platform on AWS all keep traffic within the provider relationship you already have), log usage, and confirm data-handling specifics with your provider and privacy team. Sonnet 5 covers nearly all HR drafting work; nothing here needs exotic features.
Rollout advice and pitfalls
Start with job-description drafting and policy Q&A — visible value, low risk, natural review. Bring employment counsel and, where applicable, works councils or unions in before any candidate-facing or screening-adjacent use, and write an explicit internal policy listing allowed and prohibited HR uses so well-meaning recruiters don't improvise. The pitfalls to guard against: quiet scope creep from "summarize this CV" to "which CVs are worth my time"; treating model output as the documented rationale for a decision; policy Q&A answering benefits questions from stale documents; and piloting screening tools on live candidates instead of historical, consented data with outcome analysis.
Where to go next
The decision-gate patterns that keep people in charge are detailed in Human-in-the-Loop Design: Where People Stay in Charge, and Writing an Internal Acceptable-Use Policy for Claude shows how to codify the allowed/prohibited line this article argues for. For the platform basics, see the platform overview.