Which Admissions Tasks to Automate with AI: Full Inventory

What this inventory is for

This is a one-off strategic exercise, not a live escalation script. Before you configure a single chatbot flow, you need a written decision on which of the roughly dozen recurring admissions tasks in your cycle are safe to hand to AI, which need AI drafting plus human sign-off, and which must never leave a human's desk.

Most schools skip this step and default to intuition — "fees are fine to automate, interviews obviously aren't" — which works for the obvious cases and fails badly at the edges: borderline admit decisions, disability accommodations, hesitant offer-holders. This article gives you a repeatable framework and a completed grid you can adapt to your own institution. It sits upstream of two related pieces from our AI chatbot and student recruitment guide: automating recruitment without losing the human touch covers the overall philosophy, and AI chatbot vs human agent: when should schools hand off covers real-time escalation triggers inside a live conversation. This one is the audit you run once, before either of those matters.

The four-criteria framework for classifying any admissions task

A task belongs in one of three buckets — AI-safe, AI-plus-human-review, or human-only — depending on how it scores against four independent criteria. No single criterion decides the classification; a task can score high on volume but still land in human-only if the legal risk criterion is severe enough.

1. Volume and repetitiveness. How often does this task recur, and how similar is each instance to the last? High-volume, low-variance tasks (the same fee question asked 500 times a month) are exactly what automation exists to absorb. Low-volume, highly variable tasks gain little from automation and often lose accuracy trying.

2. Rule-based vs judgment-based. Can the task be resolved by applying a fixed set of rules to known inputs, or does it require weighing competing, ambiguous factors? A fee schedule lookup is rule-based. Deciding whether a mediocre personal statement reflects genuine potential is judgment-based, and no amount of training data changes that.

3. Emotional or relational stakes. Does the task involve a moment the applicant will remember — good or bad — or is it purely transactional? Confirming a document was received is transactional. Telling someone they didn't get in, or that their scholarship request was declined, is relational, and it needs a human voice even when the underlying decision was rule-based.

4. Reversibility and legal risk if wrong. If AI gets this task wrong, can you catch and fix it before harm occurs, or is the damage done the moment the message sends? An FAQ answer that's slightly off gets corrected in the next message. A wrong admit/reject decision, a mishandled disability disclosure, or an appeal response that ignores statutory obligations carries reputational and regulatory exposure that a follow-up message cannot undo.

Score each task against all four, and the classification tends to fall out on its own: high volume, rule-based, low stakes, reversible → AI-safe. Anything with high emotional stakes or high legal risk pulls a task toward human-only regardless of how repetitive it looks on paper.

The full task-by-task decision grid

The table below applies the framework to twelve admissions tasks that recur across almost every UK higher education admissions cycle, from first enquiry through to enrolment.

Three tasks in that table need a short note, because they are where directors most often get the classification wrong.

Following up incomplete applications looks purely mechanical — send a reminder, list what's missing — but the tone matters more than it seems. A student who is one transcript away from a complete file responds differently to a templated nudge than to a message that acknowledges where they are in the process. AI can draft and send the routine reminders; a human should review the sequence design and step in once a student has missed two reminders, which is exactly the kind of repeated-failure signal covered in our handoff triggers guide.

Negotiating with hesitant offer-holders is not a single task but a category of moments — a student comparing your offer against a competitor, a parent asking about deferred payment, a student who has gone quiet after an unconditional offer. AI can surface the signal (page revisits, drop in email opens, a support ticket mentioning a rival institution) and even draft a first response, but the actual negotiation — waiving a deadline, discussing a bursary top-up, addressing a specific objection — belongs to a human who can adapt in real time.

Handoff to current-student ambassadors is rule-based in its trigger (a prospect asks "what's it really like") but the value is entirely relational once the handoff happens. AI should identify the right moment and the right ambassador match; it should never attempt to simulate the peer conversation itself.

Why the 72-21-7 split matches this grid almost exactly

The proportions in this grid are not arbitrary — they mirror what actually happens in prospect conversations at scale. Automated classification of 12,000 Skolbot conversations found that 72% of prospective student questions are simple FAQs answerable with no school-specific context, 21% need some institutional context, and only 7% genuinely require a human (Skolbot, automatic classification of 12,000 conversations, 2025). The AI-safe row in the grid — fees, programmes, logistics, slot booking — sits squarely inside that 72%. The AI + human review row corresponds closely to the 21% that needs context but not necessarily judgment. The human-only row is a subset of the 7%, but a consequential one: it's where volume is lowest and stakes are highest.

That distribution is echoed in broader sector commentary. Analysts at McKinsey and Gartner covering AI adoption in education consistently draw the same line: automation gains are largest in high-volume, low-judgment interactions, while decisions with material consequence for an individual student remain a governance and compliance concern, not just an efficiency one. Forrester research on customer-facing AI reaches a similar conclusion outside education — AI performs best where errors are cheap to detect and correct, and worst where a single wrong output is a one-shot, high-stakes event.

Why running this inventory changes what your team actually does

Running this exercise once, in writing, changes how your team spends its week — not by removing tasks from human hands, but by making the remaining human tasks the ones that actually need a person. Institutions that hand the AI-safe row to a chatbot see prospects return to explore further: 34% of prospects return to the institution's website within 7 days of a chatbot interaction, compared with 12% without one — a 2.8x difference (Skolbot, cohort analysis of 8,000 sessions tracked over 90 days, 2025). That reengagement is a byproduct of removing friction from the transactional 72%, not a reason to push AI further into the human-only rows.

The grid is also a governance artefact, not just an efficiency plan. Written classifications matter if a regulator, a QAA review, or an internal audit ever asks how your institution decided which prospect- or applicant-facing decisions involve automated processing. The ICO is explicit that decisions with legal or similarly significant effect on an individual — which a rejection or an accommodation decision plausibly is — require meaningful human involvement, not a rubber-stamped AI output. OfS conditions of registration on fair admissions practice point the same direction. Keeping a dated copy of your own grid, reviewed annually or whenever your admissions process changes materially, is the simplest way to demonstrate that AI-safe classifications were a deliberate decision rather than an accident of convenience.

FAQ

Should every school use the same classification grid?

No — the framework transfers but the specific classifications should be checked against your own institution's process, especially for financial aid and appeals, where internal policy varies significantly between institutions. Use the four criteria to re-score any task where your process differs from the general pattern in the table, particularly around who currently signs off on borderline decisions.

What happens to the "AI + human review" row in practice?

AI drafts, scores, or flags the task, and a named human reviews before anything reaches the applicant. For incomplete-application follow-ups, that might mean AI sends the first two reminders automatically and a human reviews before a third, more personal message goes out. The review step is not optional — it's what keeps the task out of the human-only bucket while still capturing most of the efficiency gain.

Does automating the AI-safe tasks reduce the quality of applicant experience?

The evidence points the other way for the tasks that genuinely belong in that row. Removing a 47-hour email wait for a fee question, or an unanswered call for a slot booking, improves the applicant's experience of speed and reliability, which frees your team to spend the recovered time on the judgment-based tasks that actually need them.

How often should we redo this inventory?

Revisit it at least once a year, and immediately after any change to your admissions policy, funding rules, or regulatory obligations — a new bursary scheme or a change to your appeals procedure can shift a task's risk score enough to move it between buckets. Treat the grid as a living document your admissions leadership team owns jointly, not a one-off memo.

Can AI ever be involved in human-only tasks like admit/reject decisions?

AI can support the process — summarising an application file, flagging missing evidence, surfacing relevant data points — without making or drafting the actual decision. The line is between AI as an input to human judgment and AI as the source of the decision itself; the tasks in this grid's human-only row are where that line needs to be enforced strictly, not blurred.