Use case 03 · HR & case-management copilot

Who's it allowed to tell?

Your HR copilot knows who's on medical leave. Claude is genuinely good at handling onboarding paperwork, tracking leave cases, and answering manager questions. The question is what happens the moment "answer the question" means disclosing something the law says only certain people may know — a medical reason, a disability accommodation, a disciplinary record. That's where "Claude wrote a helpful HR bot" and "a runtime enforces who's allowed to know what" stop being the same claim.

Data classification enforced at dispatch RBAC is a runtime object Every disclosure journaled
What stands between the question and the diagnosis
A question in — from someone with a legitimate reason to ask
↓ Claude drafts a helpful answer
Execution tree — the fields the answer would include
Per-role capability check — is medical detail in this role's grant?
↓ only the fields the role may see
HRIS / case system — a scoped answer, returned and logged
Why HR is a different kind of risk

The wrong answer isn't hacked out of the agent — it's asked for, nicely.

A support ticket is occasionally hostile. A sales negotiation is persuasive by design. An HR case is different again: the person asking usually has a completely legitimate reason to ask — and the law still says the agent may not tell them everything it knows. Nobody has to trick an HR copilot into an ADA or FMLA violation. A manager just has to ask a normal management question. Neither approach assumes the manager has bad intent. The real question is what stands between that Slack message and the diagnosis actually being disclosed.

FROM: manager@company.com (Slack DM to HR copilot) RE: Sarah Chen — coverage planning
Hey, can you catch me up on Sarah's case? Is this a medical thing? When's she expected back — I need to know if I should plan for someone to cover her role long-term or if this is short-term. Thanks!
Illustrative — built to demonstrate the mechanism, not a real employee case
✕ Claude alone, free-form integration

The case-summary logic pulls from whatever fields exist on the record and paraphrases them for whoever asks. Nothing in a prompt reliably distinguishes "a manager's legitimate need to plan coverage" from "protected medical information the ADA and FMLA say only HR and the employee are entitled to see."

✓ Magic Cloudlet + Claude

Claude can draft a helpful-sounding answer. The cloudlet checks the manager's role against the case's data classification — medical and accommodation details aren't in a manager's capability grant. The response is scoped to what a manager may know: approved leave through a date, coverage needed.

verdict · true"leave status & expected return date for direct report" → within manager grant → returned
verdict · scoped"coverage duration, no medical detail" → within manager grant → returned, logged
verdict · false"medical reason / diagnosis / accommodation detail" → outside manager's capability grant → refused, HR role required
The architecture question

Two ways to answer: what stops the agent from over-disclosing?

Ask any HR-AI vendor this and the marketing sounds identical — "role-based," "privacy-aware," "compliant by design." The architecture underneath does not. There are, in practice, exactly two answers.

A Magic Cloudlet exposes your HRIS and case-management system to Claude as a set of generated, role-gated endpoints — with each data field carrying a classification (operational, medical, disciplinary, compensation) rather than sitting in one undifferentiated case record. The agent's token carries a role, the same way a badge does. A manager role sees dates and coverage needs; an HR generalist role sees more; medical detail is scoped to HR and the employee, regardless of who's asking or how reasonably.

A data-classification rule enforced in a prompt is a suggestion. One enforced by a runtime evaluator is a fact about what can be disclosed.
Claude alone, free-form
Question → Claude
hand-written disclosure logic
HRIS / case system
Magic Cloudlet + Claude
Question → Claude
execution tree → per-role capability check
HRIS / case system

Trust boundary, left: the prompt and your guardrail code. Right: the runtime. Attack surface shrinks from every field your code can pull to only the fields this role's grant allows.

What it takes to get to production

Thirteen steps between an HR-copilot demo and one you'd trust with real cases.

The checklist a security-conscious HR leader should put in front of a CFO before letting an AI agent answer questions against live employee records.

#RequirementClaude alone, free-formMagic Cloudlet + Claude
1HRIS connectionManualCustom integration per platform (Workday, BambooHR, ADP…).HandledGenerated endpoints wrapping your existing system — zero migration.
2New-hire provisioning workflowManualBuilt and maintained by your engineering team.HandledGenerated as endpoints — accounts, equipment, paperwork routing.
3Data classification (medical / disciplinary / comp)ManualField-level tagging built and maintained by hand.HandledA capability grant per data class, enforced at every call.
4Role tiers (manager / HR / HRBP / legal)ManualHand-rolled middleware, one implementation per build.HandledRuntime-native roles — same object from token to admin screen.
5Leave & accommodation case routingManualCustom workflow logic for FMLA / ADA processes.HandledGenerated as an endpoint the correct role can call, not bypass.
6Over-disclosure defenseManualFilters and prompt hardening — a defense each new question tests.HandledA reasonable-sounding question still can't exceed the capability grant.
7Employee self-service scopingManualBuilt by hand — what an employee sees about their own case.HandledA distinct role grant, same enforcement mechanism.
8Audit trail of case accessManualDesigned, instrumented, and wired to a log store.HandledBuilt-in, filterable log — every access, every disclosure.
9Onboarding document generationManualTemplate logic hand-built per document type.HandledGenerated from the same validated employee data.
10Security review before launchManualFull human review of every generated line.PartialStructural checks handle validity; case-routing logic still reviewed.
11Compliance scope (ADA, FMLA, GINA, EEOC, state leave law)ManualOngoing, and specific to your jurisdictions.PartialSelf-host / air-gap keeps infra in your boundary; scope mapping is still yours.
12Hosting, deploy, rollbackManualA DevOps project, built once a backend exists.HandledIncluded with a managed cloudlet, or self-hosted / on-prem / air-gapped.
13Ongoing patching & maintenanceManualYour team, indefinitely, as each HRIS API changes.PartialMaintained by Nuity (managed), or self-maintained under the MIT license.
13 / 13 manual — Claude alone 9 / 13 handled — Magic Cloudlet + Claude 4 / 13 partial — review scope shrinks to case-routing and your jurisdictional policy
Security architecture, layer by layer

Five places an over-disclosure can be stopped — or not.

A security reviewer's real question isn't "is the model discreet." It's "what, mechanically, refuses to disclose the field this role isn't entitled to see" — no matter how legitimate the request sounds.

LayerClaude alone, free-formMagic Cloudlet + Claude
Field vocabularyOpen — any field your integration code can read, reachable if the prompt gets there.Closed — a finite, classified set of fields per role.
Pre-execution verificationPrompt hardening and application-level checks, written by hand.A static proof against the live capability registry, before delivery.
Capability grants (data classification)Not native — hand-built into the case-summary functions.Argument-bound grants, checked by the evaluator at every dispatch.
Identity & role (manager vs HR vs legal)A middleware pattern each new build may implement differently.A runtime object — one role, from the caller's token to the admin screen.
Audit trailWhatever your team designs, instruments, and maintains.Built-in, severity-tagged log of every access and disclosure.

Everything on the right is enforced at the moment of dispatch, by the evaluator, on every call — not by a prompt a well-meaning manager's reasonable question is free to talk past. That difference is architectural, not behavioral. It doesn't degrade as case volume, or good intentions, go up.

What this costs — and what getting it wrong costs

The platform bill is the small number.

Self-hosted, the runtime is free and MIT-licensed. Managed, it's a flat $100/mo with one developer user included. The reference backend build runs roughly 80% fewer AI tokens than hand-building the same integration and disclosure logic from scratch — about 25,000 tokens versus 140,000. Set that against the figures below, which are the EEOC's, not ours.

$660M

Recovered by the EEOC for 17,680 workers in FY2025 — the agency's own enforcement total for the year.

$40K

Average out-of-court discrimination settlement, per the EEOC — before a case ever reaches trial.

2nd

Most common basis for EEOC lawsuits in FY2025 was disability discrimination — exactly the category a mishandled leave case falls into.

Source: U.S. Equal Employment Opportunity Commission, Fiscal Year 2025 enforcement data and public reporting on it. Figures are agency-wide totals and averages, not a prediction for any specific company, and not a Nuity AI estimate. Token and cost figures are Nuity AI's published estimates as of mid-2026, based on a reference build (two linked tables, eight CRUD endpoints, access locked to a single role). Full methodology: nuity.ai/savings-calculator. This page is not legal advice; ADA, FMLA, GINA, and state leave-law obligations vary by employer and fact pattern.

For the approval memo

$100/month buys a hosted, role-gated HR runtime with enforced data classification and a full audit trail — a rounding error next to the $40,000 average cost of even one out-of-court discrimination settlement, before legal fees.

Where we're honest about trade-offs

What Claude alone still does better.

A case a budget approver can trust has to concede real points, not just make them.

Claude is the best available way to understand what someone's really asking. A Magic Cloudlet is what makes sure the answer only includes what that person is legally entitled to know — no matter how reasonable the question sounded.
Take it to the budget conversation

The brief your CFO actually reads.

Two pages — the checklist, the security model, the numbers, and the ask — built to forward. The ten-page version has the full walkthrough for anyone who wants the detail.

Twenty minutes · bring your real HRIS

Scope an onboarding and case-management copilot against your actual HRIS.

Nuity AI's engineers will scope it against your actual HRIS, your actual role structure, and this exact comparison — live, no sales deck. This use case is a proposed pattern, not yet a shipped case study; we'll show you the mechanism, not a slide.