AI for Employee Onboarding in Higher Education: A Termly Surge, Not a Steady Hire
Most writing about onboarding assumes a steady trickle of permanent hires arriving one at a time with a known start date. An HR operations team in a university or college is dealing with something else. The work arrives as a cohort, compressed into the days before teaching starts, made up of people whose relationships with the institution have almost nothing in common except that each needs clearing before anybody stands in front of a class.
The fortnight before teaching starts
The permanent academic appointment made in March is not the problem. The problem is the casual teaching, demonstrating and invigilation contracts that land in the two weeks before teaching begins, because departments only know who is covering what once enrolment numbers settle and the teaching allocation is finalised.
The volume arrives at the same point every year and the team absorbing it does not grow, and everything else continues alongside it. The gap is not caused by any one case being hard. It is caused by hundreds of cases each needing the same small acts of chasing, inside the same fortnight, against a teaching date that does not move. The crunch is the arithmetic of doing them all at once.
A contract type for every relationship with the institution
A single intake will contain a fractional academic post, a researcher funded by a grant with an end date tied to the award, an honorary or visiting appointment carrying no pay at all but needing full systems access, a casual worker engaged by the hour against a departmental budget, and a postgraduate student teaching on the module below the one they are studying.
Each has a different check set, payroll treatment, access profile and end date behaviour. The honorary appointment never touches payroll but needs a library account and a building pass. The grant funded post must not run past the award. The casual teaching contract may generate no record at all until the first claim arrives. Getting the type wrong at the start means rebuilding the case later.
The departmental request that arrives as prose
The trigger for most of this work is not a form. It is an email from a departmental administrator reading something like: can you get her set up for the autumn, she is taking the seminars on "module 2, from week 3", roughly a day a week, starting as soon as possible.
An agent can read that and produce a complete case: the likely engagement type, the dates implied, the cost centre, the access the teaching will require, and a named list of what is missing, as specific lines rather than a request for documents. No tax reference. No rate band confirmed. No prior check on file. No named budget holder approval. The chase then goes out the day the request lands rather than a week later when someone finally opens the email.
Checks that must clear before any work is done
Right to work is the hard gate. It has to happen before the person does any work, the evidence has to be recorded in a particular way, and liability for getting it wrong sits with the institution, not the department that was in a hurry. It is also the step most likely to be quietly devolved to a departmental secretary with a scanner. A photocopy taken by someone untrained to look at it is not a check and will not survive being read back later.
An agent is useful either side of that decision, never in it. It can collect the documents, tell the starter which combination is acceptable for their circumstances, book the appointment with the team authorised to carry it out, assemble the pack in the order the reviewer reads it, and flag expiry dates falling mid engagement. Where immigration status or sponsorship duties are in play, the case routes straight to the team that owns it.
Qualifications the department assumes you already hold
Departments frequently assume a person's credentials were verified years ago, during a previous engagement, or by someone else. Often nothing is on file. An agent can gather certificates and transcripts, extract the award, awarding institution and date into structured fields, check professional registration where teaching is practice facing, and identify what is unreadable, expired or inconsistent with what the department said.
What it must not do is decide whether a qualification is sufficient to teach the material. That is academic judgement, it belongs to the head of department or the appointing panel, and it becomes indefensible if nobody can explain how it was reached.
The payroll cutoff is a wall, and the claim that follows it
Payroll runs monthly and the cutoff does not negotiate. Miss it and a casual teacher who has already taught a fortnight waits weeks to be paid, then submits a retrospective claim someone has to reconcile against a contract that may not have existed when the work happened.
This is where delay has a direct human cost, and where agents do straightforwardly useful work. Work backwards from the cutoff, know which steps must clear to make it, know how long each takes here, and escalate the cases that will miss by name. A list naming the starters who will not be paid this month, and the document each is waiting on, beats any dashboard.
A card, a reading list and a timetable entry
Nobody owns provisioning. Virtual learning environment enrolment sits with a faculty learning technologist, borrowing rights with library systems, laboratory access with a technical manager, door access with estates, and the timetable entry with a scheduling team that has already published the week.
An agent can hold the dependency chain in one place: raise each request in the right queue, know which steps need a staff identifier to exist first, notice a request sitting unactioned for days, and chase the person who owns it rather than a shared inbox. It also catches the familiar failure where someone is contracted, paid and timetabled but cannot get into the building.
Induction that gates the laboratory before the lecture theatre
Some training is a genuine prerequisite rather than a formality. A demonstrator cannot enter a teaching laboratory without local safety induction. Anyone invigilating needs briefing on the assessment regulations they will be enforcing, and fieldwork, workshops and placements each carry their own requirement.
An agent can derive the required set from the engagement type and the activity described, issue it, track completion, and hold the dependent access until the record exists. The point is not completion statistics. It is that a technical manager can see, for a given session, which of the people rostered are cleared to be in the room.
The demonstrator who is also on your student roll
A postgraduate teaching assistant exists twice in your systems, once as a student and once as a worker, and the two records rarely know about each other.
Pay interacts with a stipend. Engagement hours may be capped by the terms of their studies. Access granted as a worker can conflict with access held as a student, particularly around assessment material for a module they are themselves assessed on. An agent that recognises the person already exists, links the records rather than creating a third, and flags those conflicts for a human to rule on removes a category of recurring error.
The worker who left in June and is back in September
The same people return term after term, and treating each return as a brand new starter is the largest avoidable cost here.
An agent can match an incoming request against prior engagements before any record is created, then report what can be reused and what cannot: this check remains valid, this one expired over the summer, this qualification evidence is on file, this training lapsed, these access rights were revoked at the leaver step and need raising again. The useful output is a short delta, not a fresh checklist. A leaver step that preserves what will be needed again makes the next intake cheaper.
Eleven versions of the process, and one list across them
Devolved practice is the reality in most institutions. Departments run the same process differently because they have different budgets, software habits, readings of the same policy and tolerance for lateness, and centralising that is a multi year change programme.
An agent does not need the process to be uniform. It needs to read each department's version and normalise the state into one view: every starter in the intake, their engagement type, the checks and access outstanding, the payroll run they are tracking against, and who was last chased about what. It can draft both sides of the communication too, the chase to the department and the plain explanation to the starter waiting to hear where they stand.
Where AI fails here
The statutory right to work decision stays human. A named trained person looks at the document, judges whether it is genuine and whether it relates to the person in front of them, and records that they did. The liability is institutional and personal. An agent returning a pass or fail there would create exposure rather than remove work, and we will not build it.
Anything touching immigration status stays human and stays with the team that owns it. Visa conditions, sponsorship duties, engagement hour limits tied to a person's permission to be in the country: these affect someone's right to remain. They are not a workflow branch.
Academic judgement stays with academics, as does the appointment decision. Whether a particular doctorate qualifies someone to teach a particular module is not a document check.
Then the practical limits. Document quality is poor: phone photographs, scans of scans, certificates in formats nobody here has seen before. Accuracy on clean files tells you nothing, so measure on contested cases from your own records. Staff and student systems in this sector are old, and writing back to them safely is usually the longest part of any build. The load is annual too, so a system that looks fine in a quiet month meets its real test once, at the worst moment.
Last, the honest limit. Someone's first week is a relationship, not a sequence of tasks. Knowing who to ask about a photocopier, being introduced to the person who runs the lab, understanding what the department expects: none of that is provisioned. Clearing the administration faster gives colleagues time to do that part properly. It does not replace it.
Take one term's intake of new starters
Take the last intake in full, from the departmental request to the first payslip. Count how many cases missed the payroll cutoff, how many people taught before access worked, how many returning workers were rebuilt from scratch, and how many chases went to the wrong person. Build against the largest of those, measure the next intake the same way, and expand only where the numbers hold. If it would help to work through which part of your cycle to attack first, hello@adoptaipartners.com.