What should remain human in a university?

Universities already have AI policies, committees and formal academic governance. In Australia, TEQSA has required providers to develop institutionally governed responses to generative AI. The problem is therefore not that nobody has considered AI. It is that principles such as “human oversight” or “academic judgment must be retained” do not specify what role AI may perform inside an actual decision.

An academic might determine a mark and the reasons for it, then use AI to turn those reasons into clear feedback. Alternatively, AI might read the submission, apply the rubric, propose the mark and generate the rationale, after which the academic approves the result. Both processes can be called AI-assisted marking, and both can leave the formal delegation with the academic. In practice, however, they allocate the evaluative work differently. A final human approval does not establish that the human formed or meaningfully tested the judgment.

The operational boundary is being established through local adoption, product configuration, procurement, platform upgrades and individual working practices. Academic governance determines standards, technology teams govern systems, procurement governs vendors, faculties design processes and individual staff decide how tools are used. Each function controls part of the arrangement, but the existing structure does not necessarily connect the academic principle to the system configuration and working practice that determine where judgment actually occurs.

AI can therefore change the effective location of authority without any formal delegation being amended. That makes its adoption an organisation-design problem as well as a policy problem.

The academic/professional distinction remains important to academic authority, career structures and employment conditions, but it was not designed as a comprehensive classification of university work. It accumulated as universities grew, as functions once bundled into academic roles were separated or professionalised, and as new specialist work emerged. AI is now reopening that process at the level of individual tasks: what should be automated, what requires human-machine judgment and what must remain the responsibility of a person.

Employment classification does not reliably indicate the kind of intelligence a task requires. Both academic and professional roles contain routine execution, analytical judgment and work in which human responsibility is integral to the outcome.

An academic scoring fixed-answer questions may be overseeing automatable work. A professional employee applying an explicit eligibility rule may be doing the same. Curriculum design, research strategy, learning analytics and resource planning may involve genuine combinations of human and machine analysis. Complex student support, contested misconduct findings, supervision and ethical decisions may require human responsibility regardless of whether the person performing them is classified as academic or professional.

AI acts on tasks rather than employment categories. Earlier transfers generally moved whole activities from academics to other people or functions. AI can separate research, recommendation, drafting, decision and communication within a single role, allowing substantive work to move while the job title and formal delegation remain unchanged.

The academic/professional distinction remains important, but its current form reflects earlier transfers and institutional growth; it does not determine how work should now be divided between people and systems.

PressureTest’s Intelligence Tier model makes this next round of sorting explicit. It starts with the nature of the task rather than the employment category of the person currently performing it and divides organisational activity into three operating tiers.

The first tier contains bounded processes in which an explicit rule can be executed without human participation in every transaction. University examples could include records, timetabling, fee calculations, routine eligibility, workflow routing, standard reporting and objective scoring.

Automation does not remove human ownership. People still determine the rules, validation requirements, thresholds, exceptions and appeal arrangements. The relevant question is whether the rule can be made sufficiently explicit, whether the system can be tested against it and whether unusual or contested cases have somewhere legitimate to go.

The fact that a decision has consequences does not automatically prevent automation. A person repeatedly applying a predetermined rule may add delay without adding judgment. The governance requirement is to distinguish the execution of the rule from the authority to create, change and challenge it.

The second tier contains work in which AI provides evidence, comparisons, patterns, scenarios or drafts and a qualified person interprets that material and owns the conclusion. Much of curriculum and assessment design, feedback support, research discovery, learning analytics, planning and risk analysis is likely to sit here.

The phrase “human in the loop” is too imprecise to govern this work. Human ownership is only meaningful when the person can inspect the relevant evidence, understand the recommendation, reject it, provide different reasons and accept responsibility for the result.

If the workflow does not give them the information, competence, authority or time to do those things, the nominal human approval does not preserve human judgment. It only preserves the appearance of it.

The third tier contains work in which relationship, legitimate authorship, accountable judgment or human participation forms part of what the university is providing. Likely examples include supervision, oral defence, contested assessment, complex academic-integrity findings, pastoral response, ethical deliberation and original scholarly interpretation.

AI may support these activities by retrieving evidence, exposing inconsistencies, generating questions or challenging assumptions. It should not replace the person or collegial body that must give reasons and answer for the outcome.

The boundary cannot be determined by technical capability alone. It also depends on legitimacy, contestability, due process, institutional purpose and whether the relationship itself contributes to the result.

This is not an argument that human judgment is inherently more accurate. Human decisions can be inconsistent, biased, hurried and poorly documented. A well-designed AI-assisted process may be more reliable and auditable than the process it replaces.

The purpose of Human Primacy is not to preserve human labour for its own sake. It is to preserve work that changes its nature when the responsible human participant disappears.

Content and explanation are becoming abundant. Models can generate examples, exercises, seminar questions and feedback in seconds. What a university cannot fully outsource without becoming something different is formation: the development of people who can reason, revise their views, exercise disciplinary judgment and become responsible for conclusions of their own.

Some teaching and assessment processes can be automated without damaging formation. Others can be improved through AI assistance. Complete delegation of supervision, scholarly dialogue and consequential academic judgment would change the nature of what the institution provides because the human relationship and the location of responsibility are part of the educational process.

Applied to a university, the three tiers suggest an automated institutional operations spine, a set of augmented academic and professional practices, and a protected domain of formation and scholarly judgment.

PressureTest adds four functions needed to govern that arrangement. Intelligence Architecture maintains the classification of work and manages proposed movement between tiers. Frontier Sensing tests what new models can actually do in relevant settings. Human Development prepares staff and students for the capabilities the institution continues to need from people. Value Realisation determines whether the new combination of human and machine intelligence is improving institutional outcomes.

The translation is not one-to-one. Almost every current university domain contains work belonging in more than one tier.

The proposition is narrower: the existing academic/professional boundary describes where work and authority currently sit, but not how individual tasks should be divided between people and systems. The intelligence tiers cut across the current domains, while the new functions govern their interaction and the movement of work between them.

The proposition is narrower: the university can no longer assume that academic work belongs in one block, professional work in another and technology somewhere underneath both. The intelligence tiers cut across the current domains, while the new functions govern their interaction and the movement of work between them.

Control over the boundary should remain federated rather than being handed to a single new executive.

Academic Board should determine protected academic acts and the principles governing delegation. Faculties and disciplines should interpret those principles in their own contexts. Relevant professional authorities should retain responsibility for regulated professional judgments.

Intelligence Architecture would maintain the institutional view and translate those decisions into use-case permissions, procurement requirements, system configurations and working practices. Frontier Sensing would provide evidence when a capability changes, but it would not possess unilateral authority to move work merely because a new model appears capable of performing it. Procurement and technology functions would enforce approved uses, while academic quality, risk and audit would provide independent assurance.

Every university needs these responsibilities to be connected, although they need not be placed in an identically named office. The new function owns the coherence of the boundary rather than the academic and professional judgments on either side of it.

The urgency comes from path dependence rather than a single imminent catastrophic decision. Once AI-generated recommendations become the normal starting point for a process, the human role changes. Once a platform feature becomes embedded in workload assumptions, reversing it becomes expensive. Once a capability disappears from a role, the institution may lose the capacity to recover the previous process.

The current academic/professional boundary shows how repeated transfers of work eventually harden into institutional structure. The alternative to deliberate redesign is therefore not preservation of the current university. It is another accumulated redesign, this time produced through local adoption, configuration choices, embedded product features and model upgrades that were never considered together.

Universities need to classify work according to the intelligence, authority and accountability it requires; automate bounded execution; use genuine human-machine combination where it improves judgment; reserve human responsibility where it is constitutive of the institutional outcome; and establish an explicit process for moving work between those categories.

This is the Intelligence Tier model applied to a university. It is a worked design hypothesis rather than a complete target operating model or a claim that every university should adopt seven new departments.

PressureTest takes a short description of an organisation and diagnoses where its existing boundaries conflict with the kinds of intelligence its work now requires. It does not audit policies or produce a finished organisation design. Its role is to identify the structural tension clearly enough for the organisation to decide whether it warrants further work.

Pressure-test your organisation →