A university's agents can only handle as much work as its reviewers can judge well. Admissions, student finance and casework all produce held actions that need a named person, and they peak at the same weeks every year. Without trained, consistent reviewers in each faculty, the held queue becomes the ceiling, and agents never earn more depth.
Where does the agent stop in a university?
Most useful university agents work at depth 2 or 3, which means people are in the loop by design. An admissions agent reads applications and checks that required documents are present at depth 1, then drafts a request for missing evidence at depth 2 for an admissions officer to send. A student finance agent drafts a fee refund or a hardship payment, which is held at depth 4 for a named person to release. An enquiries agent drafts replies about deadlines and procedures, each one sent by a member of staff until the agent has earned more.
Some work never leaves people at all. Decisions on mitigating circumstances, academic misconduct and admission offers outside published criteria involve academic judgement, and the agent's role is to assemble the case, not to decide it.
So scaling a university agent means scaling the people who meet it at those points. More agent capacity with the same reviewer capacity just builds a longer queue.
Why do untrained reviewers cap the scale?
Three ways, each visible in a university setting.
Peaks overwhelm the queue. Applications, enrolment and results all land in a few weeks. Universities meet peaks by borrowing staff from other offices or hiring temporary staff. If those people have not been trained to read a verdict and check where each fact came from, they will either approve drafts on trust or send everything back. The agent looks slow in the first case and dangerous in the second.
Faculties decide differently. Admissions and student services are often devolved. One faculty's officer accepts a scanned transcript as enough; another insists on a verified copy. The agent's record then shows the same kind of case approved in one place and refused in another. That inconsistency is unfair to students and it also blocks the agent's progress, because certification depends on clear, consistent outcomes.
Corrections never reach the agent. Every time a reviewer edits a draft, the business learns something about the agent's brief or sources. Untrained reviewers fix the draft and move on. The same mistake comes back next week, and the reviewer queue never shrinks.
Why does depth depend on the reviewers?
Depth is earned, not configured. A university's plan may allow an agent to execute routine work at depth 3, but evidence from Attest sets the actual depth, and the lower wins. That evidence comes from real cases from the university's own operation, which means cases that reviewers decided.
If reviewers decided those cases inconsistently, the evidence is weak. If they approved poor drafts, the evidence is misleading. Either way, the agent cannot safely be certified for more, and the institution keeps paying staff to send replies an agent could have sent. The same mechanism appears in other sectors: retail banking will not scale AI agents without trained reviewers for the same reason, though the cases differ.
Lapse adds pressure. An agent's certificate lapses after 90 days, and a run that clears no depth drops it to depth 1 the same day. If that happens in clearing week, routine replies return to people who may not have sent one for months.
What should a university train its reviewers to do?
Five skills, each tied to university work:
- Read the verdict. See which bound held the action: the agent's role scope, its certificate, a source, the university's own limit, or a platform guardrail.
- Weigh provenance. A grade from the university's own student record system is sealed; a transcript uploaded by an applicant is external until verified. An external document cannot authorise a refund or an offer.
- Refuse with a reason. Write why a draft was rejected in words another faculty's officer could apply.
- Hold the red lines. For example, never release a refund to bank details that changed in the same week, and never confirm an offer outside published criteria without the admissions tutor.
- Report patterns. Pass recurring edits to whoever owns the agent's brief.
A practical sequence is set out in how universities should train staff before deploying AI agents, and the broader order for any organisation is in the guide to training teams for AI agents.
How do you prepare for the peak?
Train and certify a core of reviewers in each faculty before the peak, not during it. Temporary staff can then be paired with a certified reviewer rather than left alone in the queue. Run calibration across faculties a few weeks ahead: reviewers decide the same set of past cases separately and agree a shared standard.
Can a university start using agents before this is done? At depth 1 and 2, yes, with care, and the trade-offs are covered in whether universities can use AI agents before staff are trained. The rule that travels: no agent reaches depth 3 on a student-facing process until the reviewers handling its held actions are trained and assessed. A university can adopt the practice of granting approval rights only to staff with a current certificate.
What is the trade-off?
Training reviewers takes staff out of service during the months before peak, which is when professional services teams are busiest preparing. Some institutions will prefer to switch on the agent and absorb the risk.
The cost of that choice shows up later, as inconsistent outcomes students can compare, as refunds released to the wrong account, or as an agent stuck at drafting because nobody can show it decides well. A written plan helps decide where the training time comes from; the AI agent training plan template for universities lays out who trains, when and against which cases.
Where the Academy fits
Pryme Intelligence Academy's Operator track is for the people who run a live workspace every day, which in a university means admissions officers, student finance staff and their team leads. Certification combines a written exam with two weeks of real review decisions, signed off by a Pryme assessor. Module tests are set in the learner's own industry and role.
Questions people ask
Should academic staff be trained as reviewers?
Only where they approve held actions, such as admissions tutors confirming offers or panels deciding extensions. Most academics need the shared vocabulary, not reviewer certification. Train the people who actually decide in the queue.
How many reviewers does a faculty need?
Enough that the queue at peak never depends on one person, with cover for leave and illness. Size it from the volume of held actions the agent produced in its shadow period, not from guesses.
Does student data change what reviewers must learn?
Yes. Reviewers handle personal data and sometimes special-category data, such as health information in mitigating circumstances claims. They must know what the agent may read, what it may include in a draft, and what must never leave the case file.