WorkSym is a neurosymbolic, governed agentic-workflow layer designed for regulated service delivery. Built by the ICS team behind regulated government workflow systems across NSW, Queensland, New Zealand and Singapore.
Decision-making remains the responsibility of organisations and individuals, not the AI system itself.
NT Government, Artificial Intelligence Assurance Framework, 31 May 2024.
Mirrored in the National framework agreed at the Data and Digital Ministers Meeting, 21 June 2024, and in the NIST AI Risk Management Framework.
The accountability principle is one sentence, universally agreed. The operational question underneath it is harder: how does a regulated agency put AI into standing service delivery, across many workflows, while every regulated disposition still lives with a named person or a codified rule?
Most AI product architectures treat this as a policy overlay. Something to be reviewed, signed off, and hoped for. WorkSym treats it as the design constraint the platform is built on. The AI is engineered to be incapable of committing a regulated disposition. Every proposal crosses an explicit neurosymbolic boundary. Every disposition is signed by a named human, with append-only audit.
This is not compliance retrofitted onto a general model. It is the shape of the platform from the first line.
Six words. The whole platform.
A governed AI Skills layer notices, drafts and proposes. A proven workflow engine records, routes and audits. Only a named agency officer commits to any authoritative source. Every proposal and every disposition is logged, traceable and stays entirely the agency's.
Every crossing of the boundary is logged. Every crossing of the boundary is inspectable.
Five Skill families. Each Skill is a versioned, testable, regression-tested artefact bound to a single workflow step. Deployable and rollback-able independently of the workflow engine. Skills are stateless. No autonomous decisions. Ever.
Agency-controlled configuration: domain schema, workflow state machines, access policy (ABAC), forms and views, reports. Configuration compiles into the live tenant without a deployment cycle. Agency configures. ICS supports.
The deterministic core. Owns state, decisions, external effects, retries, idempotency, scheduled jobs, statutory-clock tracking, and append-only audit. Every proposal from Layer 0 arrives as a typed workflow step. Exposes REST and event-driven APIs. SSO via SAML/OIDC. Integrates through message bus and file drops where preferred.
The canonical domain model for the agency. Application, Site, Applicant, Agent, Consultant, Plan-set, Document, Statement, Control, Condition, Referral, Specialist Opinion, RFI, Response, Determination, Notification, Appeal, Interaction, Audit Entry. Every read and write in Layer 2 is typed against this ontology.
Every output is a draft for a named officer. Council ERP, planning portals, land information systems, emergency management systems, grants systems stay as the systems of record. WorkSym proposes into them; it never commits into them.
No AI-issued decision carries statutory effect. Ever. The neural layer is engineered incapable of committing a regulated disposition. That is a design guarantee, not a policy overlay.
Agency data is used solely to run agency workflows. Skills read from the ontology at inference time. They are not trained on agency data, and are never used to improve a foundation model, whether ICS's or a third party's. Contractual, in writing.
No face identification. No behavioural inference. No intent inference. No autonomous decisions affecting individuals. These are engineered exclusions, not policy statements.
In-country hosting available (Australia, New Zealand, Singapore). Point-in-time recovery. Every record exportable at any time. The agency owns the schema, the data, and the evidence trail. Portable to another platform on request.
Every input, every model version, every human reviewer, every timestamp, every disposition preserved in an append-only audit envelope. Any assurance review or ombudsman inquiry can inspect what happened, after the fact, without reconstructing it.
Each anchor's workflow ontology shapes the domain layer above WorkSym. The operating substrate beneath is a single, shared platform. Together they cover parks and nature, precinct and estate, national land, and whole-of-government digital-government operations.
Statutory-board estate management. Wildlife permits, biodiversity, ecosystem stewardship, ranger casework. Builds on a seventeen-year Singapore government operational-system lineage, retired March 2025 with zero Severity 1 incidents across the full period.
Statutory precinct management for a mixed-use national estate. Permit lifecycles across venue, event, tenure and infrastructure surfaces. Workflow ontology tuned to a bounded, high-density, multi-tenant precinct.
National-scale land administration. Tenure, cadastre, land intelligence, statutory land transactions. Direct lineage to a New Zealand national land-system deployment and adjacent land-management workflows.
Central-government instance sitting above line agencies. Cross-agency orchestration for Cabinet-visible workflows: significant-project casework, disaster recovery coordination, and remote-housing program delivery. Northern Territory ontology sits above the other three in scope.
The National AI Assurance Framework leaves the operational model as the delivery question. WorkSym answers it structurally. When an agency runs its self-assurance assessment against a workflow built on WorkSym, the answers are the same answers that fell out of the platform's original architecture. They are not constructed after the fact.
The AI is engineered incapable of committing a regulated disposition. Every proposal is human-signed. Every audit entry names an accountable officer.
Skills are scoped to workflows the agency selects, not to an ICS view of what should be automated. The agency's use cases lead the platform.
Every Skill is regression-tested against a known-good corpus before it can run in production. Rollback is per-Skill and independent of the spine.
Skills are grounded via retrieval against the agency's own domain ontology and case history, not against an open-web index. Bias surfaces are inspectable.
Tenancy isolation per tenant. In-country hosting. Every record exportable at any time. No training on agency data without explicit written agreement.
The provenance envelope on every AI action is exactly the review mechanism the framework calls for. Nothing extra to build for assurance.
If your agency is thinking about how to put AI into standing service delivery while keeping every regulated disposition with a named person, we would like to talk. In Darwin, Sydney, Singapore, or by video. If it does not lead anywhere, no harm done.