Student journey and AI: What to keep 'human'

Student journey as an integrated system

Automation is no longer confined to single services. AI connects marketing, admissions, teaching support and employability into a coherent pipeline, changing how accountability, student agency and institutional purpose are experienced.

Pipeline logic meets educational mission

Many institutions still govern the student journey as a set of departments, each with its own targets and tools. AI does not respect those boundaries. Once a model links enquiry data, application history, learning analytics and careers outcomes, the journey becomes a system with feedback loops. That can be beneficial, but it also concentrates power in the design of the loop.

Examples that make the shift tangible

Consider a prospective applicant who chats with an AI assistant on a course page, receives tailored messages that nudge a particular programme, is fast-tracked in admissions because their profile matches historic success patterns, then receives automated study plans and early alerts about engagement risk. None of these steps is unusual in isolation. Together they form a guided pathway that can feel supportive or constraining, depending on how transparent it is and how often a human chooses to intervene.

Assumptions that no longer hold

• Data collected for service improvement remains local to that service.
• Students experience “support” as discrete interactions rather than a continuous, model-driven relationship.
• Quality assurance can focus on content and assessment without inspecting decision systems.
• Fairness issues appear mainly at admissions rather than across the full lifecycle.

The system question is now unavoidable: what outcomes are being optimised, and who gets to define them?

Why timing matters right now

The issue is urgent because the capability frontier has shifted, regulation is converging, and competitive baselines are moving internationally. The decisions made now are likely to harden into infrastructure.

Capability is moving from assistance to autonomy

Generative AI started as a productivity layer for staff and students. It is increasingly becoming a decision layer, embedded into CRM platforms, virtual learning environments, assessment workflows and customer support. The move from “help me draft” to “decide and route” is where governance and institutional identity are tested.

Regulatory pressure is becoming operational

In the UK context, UK GDPR, Equality Act duties and OfS expectations around quality and student outcomes create a compliance frame that intersects with AI design. Internationally, the EU AI Act sets a direction of travel for risk classification, documentation and oversight, which many multinational providers will standardise around. These are not abstract concerns once automated systems start shaping access, support intensity, or assessment decisions.

Expectations are being set outside the sector

Students and employers experience AI-enabled personalisation in finance, retail and health. That raises expectations for responsiveness in education, but also raises sensitivity to manipulation and opaque profiling. A “digital first” journey can still feel human; an “automated by default” journey can feel transactional. The sector is deciding, implicitly, which experience it wants to normalise.

Early signals from AI-enabled journeys

Before debating principles, it helps to observe what AI is already doing well and where it predictably struggles. Concrete examples reveal the boundary between scale and judgement.

High-value automation already visible

Some of the most credible gains are in high-volume, low-stakes interactions where speed and consistency matter. For example, 24/7 student support triage that routes queries, drafts responses and flags safeguarding indicators can reduce waiting times and free staff to handle complex cases. Similarly, formative feedback tools can offer immediate guidance on structure, clarity and comprehension checks, helping students iterate more often.

Where the risks cluster

Risk tends to concentrate where automation influences life chances or identity: admissions decisions, disability support pathways, suspected misconduct, progression and termination, or employability signalling. A model might be statistically accurate yet still unacceptable if it amplifies historical disadvantage or cannot be explained in a way that a student perceives as legitimate.

A subtle design trap

Personalisation can slide into path dependency. If a system recommends fewer stretch modules to someone predicted to struggle, the institution may improve short-term retention while narrowing long-term capability. In a similar way, “nudges” that increase conversion or attendance can become coercive if they are not transparent and contestable.

Human moments that carry symbolic weight

There are moments where students look for recognition, not information: the first serious academic setback, a shift in career identity, a complaint about fairness, or a decision about whether to pause. Automation can support these moments, but may not be the right “face” of the institution.

Human work that should not disappear

Keeping some elements human is not nostalgia. It is about preserving functions that depend on legitimacy, moral judgement and relationship, especially when AI systems operate at scale.

What kind of “human” is needed?

Not every interaction needs a live conversation. The human role is often to provide accountable judgement, to interpret context, and to hold responsibility when rules collide. In practice, this may mean fewer routine interactions but higher-quality human engagement where it matters.

Candidate areas for explicit human stewardship

• Decisions that change a student’s options: admissions exceptions, progression boards, outcomes of academic integrity cases, and decisions with financial consequences.
• Interpretation of meaning: translating feedback into a personal development plan, helping students reconcile conflicting signals, or navigating identity shifts such as career pivots.
• Relational repair: complaints, perceived injustice, or moments when trust in the institution has been damaged.
• Boundary cases: safeguarding, health crises, and complex disability adjustments where standard policies rarely fit.

Augmentation rather than replacement

In some AI-native learning models, humans use AI to see patterns they would otherwise miss, then choose how to respond. At the London School of Innovation, for instance, the emphasis in platform design conversations has often been less about automating teaching and more about making human time more consequential, using private AI tutors for practice and feedback while reserving live sessions for judgement, coaching and accountability.

A reframing worth testing

The question may be less “what to automate” and more “what must remain contestable”. If a student cannot understand, challenge, or appeal a decision pathway, the institution may be accumulating reputational and regulatory risk, even if outcomes look efficient.

Governance choices hidden in design

Automating the journey turns product decisions into policy decisions. Institutions need governance that inspects objectives, data flows and accountability, not only vendor assurances.

Model objectives as institutional policy

Every automated intervention encodes a goal: reduce dropout risk, increase conversion, improve satisfaction scores, raise average grades. These goals are not neutral. If a system is tuned for retention, what happens to academic standards or intellectual risk-taking? If tuned for satisfaction, what happens to challenge? The governance task is to make trade-offs explicit and reviewable.

Key questions for oversight frameworks

• What decisions are automated, and which are advisory only?
• Where is accountability located when a model recommendation contributes to a contested outcome?
• How are bias and disparate impact tested over time, not only at procurement?
• What is the student’s right to explanation, appeal, and human review in practice?
• How is data minimisation applied when end-to-end journeys make “more data” feel tempting?

Procurement is now capability building

Vendor selection can no longer be treated as buying software. It is selecting an operating model. Institutions may need to build internal capacity for evaluation, model monitoring, and incident response, even when systems are outsourced.

An empirical study that could reduce ambiguity

The sector would benefit from multi-institution evaluations that compare AI-augmented journeys with different “human checkpoint” designs. For example, a longitudinal study could test whether scheduled human reviews at specific inflection points change outcomes such as continuation, attainment gaps, appeals volume and student trust, compared with journeys that rely mainly on automated nudges. The aim would not be to crown a winner, but to identify where human involvement changes trajectories in ways that matter.

Decision tests for the next phase

The next phase is not choosing between human and machine. It is designing a journey where automation increases capability without eroding legitimacy. The hardest work sits in the questions that remain open.

Useful decision tests

Automation tends to expand quietly. A practical approach is to use decision tests that slow expansion at the right moments and invite scrutiny when stakes rise.

• Legitimacy test: Would a student accept this decision pathway as fair if they could see it end to end?
• Contestability test: Is there a real, accessible route to challenge, appeal or request human review?
• Capability test: Does this intervention build student agency over time, or create dependency on prompts and predictions?
• Equity test: Are outcome gaps narrowing across demographics and modes of study, or being masked by aggregate improvements?
• Resilience test: What happens when the model is wrong, the data drifts, or a supplier changes terms?

Difficult questions worth holding

Some questions should remain deliberately unresolved until institutions decide what kind of educational relationship they want to offer in an AI-rich environment.

• Which parts of the student journey constitute a moral relationship rather than a service transaction?
• What would count as unacceptable optimisation, even if it improves retention or revenue?
• Where should human judgement be mandated, and where should it be optional but auditable?
• How will students be informed about automated influence, and what does meaningful consent look like in practice?
• What evidence would justify expanding automation into higher-stakes decisions, and what evidence would require rolling it back?

Automating the student journey is a design choice about the future shape of higher education. The open question is whether institutions can use AI to increase human capability while preserving the trust that makes education socially valuable.