How to Build an AI HR Assistant (Playbook)

HR teams spend a large share of their time answering questions whose answers exist in policy documents, while also handling matters that require human judgment and care. An AI HR assistant takes the first category and protects the second: grounded policy answers, routine requests through HR systems, and immediate routing of sensitive topics to people. This playbook covers the build, following FISTA's AI agents practice. Context is in ai in hr recruiting and the related how to build an ai onboarding assistant.

What does the assistant do?

Capability	Mechanism	Control
Policy and benefits answers	Grounded retrieval with citations	Curated, versioned content; role and location filters
Routine requests	HR system tools	Employee identity; approvals where required
Case creation and routing	HR case system integration	Category rules; sensitive-topic detection
Sensitive topics	Detection and immediate human routing	No assistant handling
Escalation	Human HR with context	Low confidence; employee request

Step 1: Specify scope with HR leadership

Define policy areas, locations, and employee groups in scope; routine requests the assistant may help with and their approval rules; the sensitive-topic list and routing; prohibited behavior (no advice on legal or medical matters, no compensation statements, no access beyond the employee's own records); and quality thresholds. HR owns the specification. See how to write an ai spec.

Step 2: Curate and version policy content

Inventory policies, benefits guides, and procedures; fix contradictions; retire superseded versions; tag by jurisdiction, entity, employee group, and effective date; assign owners; and establish an update process. Readiness method is in the data readiness for generative AI whitepaper.

Step 3: Map identity and permissions

Authenticate employees through the identity provider and map to HR system attributes (location, entity, employee group, manager status) so retrieval filters by applicable policies and record access is scoped to the employee's own data, with manager-only content restricted to managers. Design is in ai access control.

Step 4: Ground answers

Answer from filtered, curated content with citations to the policy and section, state effective dates, and refuse when no applicable policy exists, offering a route to HR. Architecture is in the enterprise RAG reference architecture whitepaper.

Step 5: Detect and route sensitive topics

Implement detection for harassment, discrimination, grievances, disciplinary matters, accommodation, medical and mental health disclosures, compensation disputes, and terminations, using rules and classifiers tuned for recall. On detection, the assistant stops, responds with care, and routes to qualified HR staff with the conversation, telling the employee a person will help. Detection is evaluated on scenarios and tuned with HR. Routing patterns are in how to build an ai ticket routing system.

Step 6: Integrate routine requests with approvals

Connect HR system tools for leave balances, leave requests, document requests, and profile updates under the employee's identity, with approvals where policy requires (manager approval for leave). Tool design is in how to build tool use for llm agents.

Step 7: Enforce privacy

Store minimal conversation data with redaction and defined retention; prevent any cross-employee exposure; log access; provide notice about the assistant, logging, and how to reach a human; and align with employment and data-protection law by jurisdiction. Guidance is in ai data privacy compliance and ai and gdpr.

Step 8: Evaluate and pilot

Build a golden set of employee questions by policy area and location with HR-approved answers, plus sensitive-topic scenarios and permission tests. Measure accuracy, citation validity, routing correctness, refusal behavior, and leakage. Pilot with one location or group, monitor sampled answers and routing, and expand. Method is in the AI evaluation and testing whitepaper.

Worked example: a multinational employer

An employer with staff in several countries deploys the assistant for policy and benefits questions. Policies are curated and tagged by country and entity; identity mapping applies the right set to each employee. Leave balance lookups and leave requests run through the HR system under the employee's identity with manager approval. Sensitive-topic detection routes a message mentioning a manager's conduct to the HR business partner within minutes, with the assistant responding supportively and stopping. Evaluation by country shows high accuracy on leave and benefits questions and reveals a country whose parental leave policy had two conflicting versions on the intranet, which HR resolves. Employees see notice of the assistant and how to reach a person, and HR case volume for routine questions falls while sensitive cases reach people faster.

What does it cost to run?

Cost scales with employee count and interactions and is modest; build cost is driven by content curation and HR system integration. Value is measured in HR time on routine questions and employee time to answer. Drivers are in ai copilot cost.

What are the common mistakes?

• Indexing the intranet without curation and answering from retired policies.
• Ignoring location and entity differences.
• Letting the assistant engage with sensitive topics.
• Record access through a privileged account.
• No notice to employees.
• Evaluating accuracy on average while one location is wrong.

How do you phase the rollout?

Start with policy and benefits questions for one location, where content is best curated, before adding routine requests through HR system tools and before expanding locations. Sensitive-topic detection is live from the first day regardless of scope, because employees will raise sensitive matters wherever an HR channel exists.

How do you handle policy changes?

Policies change, and an assistant that answers from last year's handbook creates liability. Version policy content with effective dates, have HR approve each version before it is indexed, re-run the evaluation set on every policy update, and show the effective date and source in each answer so employees and HR can see what the answer is based on.

How FISTA Solutions builds HR assistants

FISTA Solutions builds HR assistants to this playbook: HR-owned specifications, curated and versioned policy content, identity-mapped permissions, grounded answers with citations and effective dates, high-recall sensitive-topic detection with human routing, HR system tools with approvals, privacy by design, and evaluation by policy area and location. The AI agents practice delivers the assistant, AI enablement the retrieval and integration platform, and forward deployed engineers embed with your HR and legal teams. The record behind the work is 150+ projects with 99.9% uptime.

To scope an HR assistant, message FISTA on WhatsApp, or read ai employee onboarding for the onboarding-specific use case.