Why Generic AI Fails at Psychosocial Risk Management — And What Actually Works

Every modern organisation generates data about its people. Performance reviews, survey responses, incident reports, leadership assessments, absenteeism patterns.

But here is the uncomfortable truth: raw data and general-purpose AI cannot see psychosocial risk.

A generic chatbot does not know that a spike in sick leave in one team, combined with a recently promoted manager who has had no leadership training, against the backdrop of a restructure announcement, creates a confluence of hazards that Australian work health and safety law now requires you to manage.

This isn't a technology problem. It's a knowledge architecture problem.

The ontology revolution in specialised AI

To understand why Gentia exists, it helps to understand what is happening in the broader AI landscape. The most transformative applications are not the general-purpose models everyone talks about. They are the vertical platforms that encode deep domain expertise into intelligent systems.

Consider what has emerged.

Legal AI. Harvey has become a trusted tool for legal professionals — not because it writes better prose than a general model, but because it houses the intricate relationships between precedents, jurisdictions, case law and legal reasoning that lawyers need to trust.

Cybersecurity AI. Vanta, Drata and Delve lead in compliance automation — not because they are flashier, but because they have built the scaffolding that maps every control to every framework to every evidence requirement.

Medical AI scribes. Heidi and DeepScribe outperform generic transcription — not through superior voice recognition, but through ontologies that connect symptoms to diagnoses to treatments to documentation requirements.

The pattern is clear. Vertical AI wins by encoding the complex web of relationships that define a domain.

What is an ontology, and why does it matter?

An ontology is more than a glossary or a database. It is a living model of how concepts in a domain relate to each other.

Think of it this way. A taxonomy tells you that bullying is a category of psychosocial hazard. An ontology tells you that:

• Bullying manifests through specific observable behaviours
• Those behaviours occur in particular contexts — one-on-one meetings, team settings, digital communications
• The same behaviour carries different risk weight depending on power dynamics
• Bullying intersects with other hazards such as workload, creating compounding effects
• Certain leadership gaps create the conditions where bullying flourishes
• Regulatory requirements under Safe Work Australia codes define specific organisational obligations
• Evidence of harm follows particular documentation pathways

This web of relationships is what makes AI useful in a complex compliance domain. Without it, you have a very articulate system that does not actually understand your problem.

Gentia's psychosocial risk ontology

Gentia has built a comprehensive ontology that models the landscape of psychosocial risk in Australian workplaces. It is not a database of hazards. It is a structured representation of four things.

The hazard architecture

Every psychosocial hazard identified in Australian WHS frameworks — job demands, low job control, poor support, lack of role clarity, poor organisational change management, inadequate recognition, poor workplace relationships, poor organisational justice, traumatic events, remote or isolated work, poor physical environment, violence and aggression, bullying, harassment, and conflict — mapped with their interrelationships and the ways each one shows up in real work.

The regulatory framework

The model WHS regulations, Safe Work Australia codes of practice, and state-specific requirements, and how each of those translates into an organisational obligation. Every Australian jurisdiction is treated separately, because the obligations differ.

Workplace systems

Hazards do not live in policies. They live in the systems that create working conditions:

• How work is designed and allocated
• How teams are structured and managed
• How communication flows, or fails to
• How change is implemented
• How leadership capability develops
• How performance is measured and rewarded

Human dynamics

The interactions between employees, peers, managers and leaders that create psychological safety, or erode it. The ontology models manager-employee relationship patterns, leadership behaviour indicators, team cohesion factors, and the communication patterns that signal risk building.

Evidentiary pathways

What constitutes evidence of risk, harm and control effectiveness under Australian law — and how surveys, incidents, claims and qualitative material connect to demonstrate compliance, or expose the gaps in it.

Why generic AI cannot do this

Consider what happens when you ask a general AI assistant to help with psychosocial risk.

You say: our engagement survey shows declining scores in one department. What should we do?

It says: here are some general tips for improving engagement. Ensure clear communication, provide development opportunities, recognise achievements.

This is not useful. It is not even wrong. It is simply disconnected from the regulatory and organisational context you operate in.

What you actually need to know is:

• Does this pattern indicate a psychosocial hazard that triggers your duty to assess and control?
• What specific hazards might be manifesting here?
• What questions would surface root causes without creating legal exposure?
• What does reasonably practicable control look like in this situation?
• How do you document this appropriately?
• What other data points should you triangulate against?

A system grounded in a psychosocial risk ontology can navigate those questions, because it understands the conceptual architecture of the domain.

What sits underneath

The ontology is the foundation, but it is not the whole system.

Purpose-built assessments. Instruments designed specifically to surface psychosocial risk indicators, rather than engagement surveys repurposed in the hope of catching something relevant.

Cleaned and expert-labelled data. AI is only as good as the knowledge it reasons over. Gentia's dataset is domain-specific, tagged and labelled by psychosocial safety specialists, legally informed, and continuously refined.

Semantic retrieval. The system retrieves relevant context conceptually rather than by keyword matching, so a description of a situation surfaces what is actually relevant to it.

Taxonomies. Hierarchical classification working alongside the ontology, so categorisation and reporting stay consistent.

Specialised agents. Rather than one system attempting everything, Gentia uses purpose-built agents that combine into workflows: agents that gather the right information, agents that identify hazards, agents that recommend evidence-based controls, agents that produce records that stand up, and agents that track leading indicators.

Structured reasoning. The architecture requires the system to work through a problem before it answers, rather than responding immediately — which is what makes the reasoning inspectable rather than assertable.

Expert review. Every material finding is checked by a psychosocial safety specialist before it is relied on. AI increases speed, specificity and consistency. It does not replace expert judgment.

The expert intern model

Here is a useful way to think about what this provides.

Imagine you could hire a team of expert interns. Each deeply knowledgeable about one aspect of psychosocial risk management. Each trained by strong practitioners in the field. Each available the moment you need them.

One understands hazard identification. One knows the regulatory landscape cold. One specialises in intervention design. One is strong at data analysis. One writes documentation that would satisfy a regulator.

Individually, they are capable. Stitched together into a workflow, they become a force multiplier for a WHS or HR function.

You still decide what to build. You still design the workflows that make sense for your organisation. But you are building on genuine domain expertise, rather than hoping a general-purpose model will intuit the complexities of psychosocial risk law.

Why this matters now

The regulatory environment has shifted. Australian WHS regulators have made clear that psychological hazards must be identified, assessed, and controlled in the same way as physical hazards. The days of treating mental health as a wellbeing initiative rather than a compliance obligation are over.

At the same time, AI capability has reached the point where genuine assistance is possible — but only where the system actually understands the domain.

Mid-sized organisations face the sharpest version of this. Enough complexity to carry real exposure, without the specialist headcount a large enterprise can employ.

They need to scale WHS and HR capability without scaling headcount proportionally. Generic AI will not get them there.

The trust question

When you are dealing with legal compliance, employee wellbeing and organisational risk, trust is not optional. You need to know that the system guiding your decisions actually understands what the law requires, what good practice looks like, where the hazards hide, which controls work, and how to document defensibly.

That is why specialised platforms have won in legal, in medical documentation and in security compliance. Psychosocial risk management deserves the same foundation.

The ontology is the foundation. The architecture is the scaffold. The agents are the capability. Together they produce a system that can be trusted with the complexity of workplace psychological safety.

Where to start

None of this is useful in the abstract. The question that matters is what it finds in your organisation.

That is what a Gentia Roadmap does. It works from material you already hold — policies, position descriptions, structure, incident and claims data, workforce reporting — and produces a documented assessment of where psychosocial risk actually sits, mapped against your obligations in every jurisdiction you operate in, with a prioritised plan for what to do about it.

The ontology is what makes that assessment specific to you rather than generic. It is why two organisations in the same industry receive materially different findings.

That is the difference between an AI that talks about psychosocial risk and one that can actually help you manage it.