Key facts
| Router model | plugsky-fusion escalates per request across tiers (live) |
| Embeddings | Live embeddings API for policy and handbook retrieval |
| Models | 30+ models; cheap tiers suit policy Q&A and form guidance |
| PII controls | Redaction modes and region pinning for employee data |
| Pricing | Flat monthly self-serve plans with no per-token charges on self-serve |
| Free tier | plugsky-micro and plugsky-lite on the free plan, no card required |
| Governance | Scoped keys, RBAC and audit logs for HR workflows |
| Roadmap | The moderation endpoint is coming soon; apply your own filters today |
TL;DR
- Policy Q&A is high-volume and deterministic: cheap tiers with citations.
- Escalate personal circumstances, disputes and legal-adjacent questions.
- Redact personal data before prompting; keep regions pinned.
- Ask for citations so answers match the current handbook version.
- Track escalations by intent — they show where automation stops being safe.
How it works, step by step
- Index the handbook, benefits documents and HR processes with version metadata.
- Classify employee questions by intent and sensitivity.
- Route policy lookups to cheap tiers with retrieval and citations.
- Escalate grievances, medical situations, compensation negotiations and legal questions to humans.
- Redact personal identifiers before any model call and pin the workspace region.
- Answer only from retrieved current-version policy, and state clearly when to contact HR.
- Measure deflection, escalation and citation accuracy by intent, and re-tune monthly.
Try it yourself
High volume, low complexity
Most HR questions are procedural: how many leave days remain, when enrolment closes, how to submit an expense, what the parental policy says. These questions have stable answers in documents you already maintain, which makes them cheap-tier work with retrieval and citations. The volume is large, the ambiguity is low, and the answer changes only when policy changes.
Routing these to a strong model wastes money without improving outcomes. The value is in accurate retrieval from the current handbook version and in telling the employee exactly which document the answer came from, so they can read further.
Sensitivity, privacy and escalation
Some questions must not be automated at all, and others must be handled with extra care. Build the escalation list explicitly with your HR team:
- Grievances, harassment reports and disciplinary matters go to a person immediately.
- Medical, family and personal circumstances need human empathy and judgement.
- Compensation disputes and contract interpretation require HR or legal review.
- Anything the retrieval cannot ground should say so and offer a contact, not improvise.
Privacy is a design constraint: redact identifiers before prompting, pin the workspace to the required region, and restrict retrieval by role so managers and employees see only what they should.
Accuracy, versioning and measurement
HR answers go stale when the handbook changes. Store document versions with vectors, expire superseded content, and require citations so a wrong answer can be traced to a specific version. Test after every policy update with a question set covering the changed sections.
Measure deflection by intent, escalation rate, citation accuracy and employee satisfaction. A high deflection rate on sensitive intents is a warning, not a win. Build the pipeline on the free plan with plugsky-micro and plugsky-lite and evaluate stronger models for escalation summaries during the 14-day full-access trial; the live pricing page lists plans, and enterprise deployments can add the residency controls HR data requires.
Honest comparison
| HR question type | Routed HR assistant | Strong model for everything | Static FAQ page |
|---|---|---|---|
| Leave and benefits | Cheap tier with citations | Frontier price per lookup | Self-service reading |
| Process guidance | Cheap tier, structured steps | Native strength | Written guides |
| Sensitive cases | Human escalation | Model plus review | Email to HR |
| Policy changes | Re-index and expire versions | Prompt updates | Manual edits |
| Privacy | Redaction and region pinning | Provider defaults | Document access controls |
Frequently asked questions
Is it safe to automate HR answers?
For procedural policy questions, yes, if answers come from the current handbook with citations. Sensitive matters such as grievances, medical situations and compensation disputes should route to people, not models.
Which model tiers suit HR?
Policy lookups and form guidance run well on cheap tiers with retrieval. Reserve stronger models for summarising complex cases or drafting internal responses that a human will review.
How do I protect employee data?
Redact identifiers before prompting, pin workspaces to the required region, restrict retrieval by role, and keep request bodies unstored per the platform's retention controls.
How do I keep answers current?
Store document versions with vectors, expire superseded content, and require citations so every answer maps to a specific handbook version.
How do I decide what to escalate?
Work with HR to list sensitive categories — grievances, medical, disciplinary, compensation and legal — and route those to people immediately, regardless of model confidence.
Can the assistant book or change HR records?
Only through tools with server-side permission checks and audit logs, and only for actions HR approves for automation. Write access should be narrow and idempotent.
How do I measure the assistant?
Deflection and escalation by intent, citation accuracy, and employee feedback. High deflection on sensitive intents is a red flag rather than a success.
Can I pilot without paying?
Yes. plugsky-micro and plugsky-lite are on the free plan with no card, and the 14-day full-access trial covers stronger models for review workflows.