Platform architecture
The HR OS
for AI agents.
One governed core under every HR app and every agent. Deterministic where outcomes must be exact. Probabilistic where understanding helps. Recorded as it runs.
People
HR · Managers · EmployeesFinance · AuditorsAI agents
MetisExternal agents — plannedHeyHR
Governance orchestratorDeterministic
Compliance & policy engineRules · audit · executionProbabilistic
Language model engineReasoning · explanation · Q&AApps & sub-agents
§01
What AI-native means at HeyHR.
Not a chat layer. Not an automation trick. A platform architecture in which AI works as a trusted part of operations — inside permissions, inside policy, inside reviewable workflows, and alongside deterministic execution wherever outcomes must be exact.
Legacy software
People operate the system.Every task is navigation and manual entry. Knowledge sits with the operator.
AI overlay
A model on top of old data.Access is unclear, answers are generated rather than derived, and nothing is traceable. Acceptable for FAQs; not for payroll.
AI-native
Agents operate inside the rules.Every surface reaches one governed core. AI interprets and explains; rules compute; each run is recorded.
§02
Intelligence that connects.
Governance that holds.
Every request — from a person, Metis or, in future, an external agent — passes the same gate and reaches the same orchestrator. Requests outside a user's scope stop at the gate.
§03
In HR, governance isn't a feature.
It's the foundation the AI stands on.
Data privacy, access control and governance are part of any sound AI-native design. In HR and payroll — salaries, IDs, bank accounts, performance — they decide whether AI can be used at all. So we built them below the model, not around it.
| Control | Mechanism | Why it matters in HR |
|---|---|---|
| 01Identity binding | The acting user is taken from the server-side session — never from the prompt or model output. | A prompt can't talk Metis into being someone else. |
| 02Role- and attribute-based scope | Every request is authenticated, and access is set by user, role, company and legal entity. | A manager and a payroll admin asking the same question get different, correct answers. |
| 03Tenant isolation | Each customer's data is logically separated by company and legal entity. | Group structures with many entities stay cleanly apart. |
| 04Minimum necessary data | Only the fields an answer needs, and only fields the user may see, are passed to a language model. | The model never holds more than the answer needs. |
| 05Deterministic execution | Pay, contributions and tax are computed by a versioned rules engine. Models never produce a payroll number. | The same inputs give the same pay — every time, provably. |
| 06Explicit confirmation | Nothing changes until the user confirms it. | AI can prepare; only a person can commit. |
| 07Lineage & audit | Inputs, rule versions, actions and approvals are recorded as execution happens. | Audit reads recorded truth, not a reconstruction. |
| 08Encryption & residency | Encrypted at rest and in transit; primary data stored in Singapore. | Payroll data stays protected, with its main home in-region. |
§04
Trustis the architecture.
HR data is not one flat surface. Roles, permissions and visibility are webbed together, so the same question must not yield the same answer for every user. An AI that merely mimics user behaviour can't respect that; an architecture can.
Governed AI4 patents pending →Employee sees: own personal data, payroll, bank account, tax
§05
Any surface. One governed path.
Web app, Slack or WhatsApp — the surface changes. The permission model, data access, policy logic and execution controls don't.
David TanReviewBank account missing
Sarah LeeReviewCPF rate pending
44 othersReadyRun staged
Why is my take-home lower this month?
A CPF adjustment and one day of unpaid leave were applied. This answer comes from your actual payroll trace and your permissions.
@Metis who's on leave in SG next week?
4 people — Arjun, Wei, Siti and Daniel. Sales has 2 out on Tuesday. Shown because you manage the SG office.
Illustrative.
§06
From inferred reconstruction
to recorded fact.
Traditional payroll audit is reconstruction after the fact: inputs gathered again, rules re-read, outcomes inferred. HeyHR records provenance and lineage as each calculation runs — where every input came from, which rule version applied, and who approved it — so explanation and audit start from recorded fact, not inference.
Layer 1 · Execution breadcrumbs
PATH /process/payroll REF SG.PAYROLL.V3 REF CPF.RATE.2026 INPUT USER_SCOPE · ENTITY_SG01
Layer 2 · Run-level catalog
| ref | hash / id |
|---|---|
| SG.PAYROLL.V3 | 0x4a7b9d2e |
| CPF.RATE.2026 | 0x1f8c5e3a |
| USER.SCOPE | 0x9d2e4a7b |
Layer 3 · Materialized graph
What a person sees · Net pay, Nov 2026SGD 5,284.00
- Basic salary6,500.00Employee record · effective 15 Feb 2026
- Overtime 1.5×412.50Timesheet #T-1182 · approved by K. Lim
- Unpaid leave (1 day)−298.85Leave #L-447 · rule: daily rate, 5-day week
- CPF · employee 20%−1,329.65CPF.RATE.2026 · age band ≤55 · OW ceiling applied
Run 2026-11-SG · rules v2026.4 · approved 26 Nov 10:42 · figures illustrative
§07
Where the HR OS is heading.
What comes next is not more fragmentation. It is a governed ecosystem built around the same orchestrator.
Current
One orchestrated platform, inside a single trust boundary.Payroll and core HR run as one governed, multi-agent system: AI, permissions, policy-aware workflows, deterministic execution and audit lineage in the same architecture.
- Multi-agent system
- Governance orchestrator
- Deterministic execution
- Role-aware access
- Audit lineage
Next · on the roadmap
External agents and apps — connected through the orchestrator.Your own agents, partner HR apps, recruiting systems and other AI services connect through HeyHR's orchestration layer, inherit the same governance model, and collaborate through agent-to-agent (A2A) and agent-to-orchestrator (A2O) patterns — without breaking the trust boundary. Timing will follow how agent standards mature over the next one to two years.
- A2A / A2O
- External agents
- Third-party HR apps
- Partner services
§08
Under the hood.
- Hosting
- Primary application data, files and backups in Singapore
- Encryption
- Encrypted at rest and in transit
- Tenancy
- Multi-tenant; separated by company and legal entity
- Workflows
- Durable orchestrated workflows — runs complete or stop safely
- Models
- Provider-agnostic agent layer — models can change; the controls around them don't
- AI data
- Employee NRIC, passport and bank numbers never sent to a model; customer data is not used to train AI models
- Countries
- Singapore and Hong Kong on HeyHR; others connected or uploaded into one record
- Assurance
- Internal security reviews, tracked to remediation
Platform questions
What is an HR OS?
An HR operating system is the governed core that HR apps and AI agents run on: one employee record, one set of rules and one permission model. HeyHR provides that core, so payroll, HR apps, Metis and — in future — external agents all work from the same data under the same controls.
What does AI-native mean for payroll?
It means AI is part of the architecture rather than a chat layer on top: agents act inside permissions, policy and reviewable workflows, while a deterministic engine computes every payroll number and lineage is recorded as execution happens.
Does AI calculate payroll in HeyHR?
No. Pay, contributions and tax are calculated by HeyHR's versioned rules engine. Metis, the AI HR agent, explains results, checks runs and prepares actions that a person confirms.
How does HeyHR stop an AI agent from seeing data it shouldn't?
The acting user is bound from the server-side session, every request is authenticated, and access is set by user, role, company and legal entity. Employee NRIC, passport and bank numbers are never sent to a model.
What is payroll lineage?
Payroll lineage is the recorded trail behind every figure — inputs, rule versions, actions and approvals — captured as payroll runs, so audit and explanation don't depend on reconstruction.
Why do provenance and lineage matter for accounting and audit?
Payroll is a major expense and a liability on the books, so finance and auditors need to trace each figure in the general ledger back to its source: the employee record, timesheet or claim it came from (provenance), and the rule version, calculation steps and approval that produced it (lineage). With both recorded at run time, journals reconcile to payslips, accruals and statutory filings can be explained line by line, and audit samples can be checked without rebuilding spreadsheets. Because changes record who made them, when and under which approval, it also supports internal controls over financial reporting.
Will external AI agents be able to connect to HeyHR?
Yes — it is on the roadmap. External agents and partner apps will connect through HeyHR's orchestrator and inherit the same governance model. Timing will follow how agent standards mature over the next one to two years.