Most of what is written about AI in higher education is either a promise or a warning. If you run a registrar’s office or a financial aid office, neither helps you decide what to do on Monday.
I have worked in both offices, and I now work every day with AI agents on the systems behind them, building the tooling, the rules and the checks they run under. This is the practical view from where I sit.
Start with what the work is
A great deal of office work in a college is careful reading. Does this list match that list? Does every student who should have a record have one? Which of these four thousand rows is the one that is wrong?
People are good at judgment and poor at reading four thousand rows the same way every time. Software is the opposite. An AI agent is useful for the part in between: working out which comparison to run, running it, and explaining what it found. That tells you most of what you need to know about where agents fit.
Three places they fit
Reconciling one system with another. Enrollment in the student system against the learning platform. Awards against disbursements. What one report says against what another says. An agent can set up the comparison, run it over every record, and hand a person the short list of differences.
Checking a bulk change before and after. When records are moved or corrected in bulk, an agent can verify every record against the source, not a sample. Better still is a second agent, separate from the one that did the work, whose only job is to check it. I describe that arrangement in the two-agent verification case study.
Preparing work for a person to review. Sorting incoming support cases, drafting the exception report, running the pre-flight checks before a change goes live. The agent does the gathering. A person makes the call.
In all three, the agent reads and proposes. It does not decide, and it does not change the system of record.
Three places they do not
Decisions about a student. Admission, eligibility, an appeal, how much aid. These are judgments your institution is accountable for, and they should be made by a person who can explain them.
Unattended changes to live records. An agent that can write to production can be wrong at scale, quickly. In my work production is read-only to agents, and a person applies every change.
Answering regulatory questions. An agent will give a confident answer about a financial aid rule whether or not it is right. That is not a risk worth taking with Title IV. Ask your compliance staff.
The controls that make the difference
The gap between a useful agent and a dangerous one is not the model. It is what surrounds it.
- Limits that are enforced, not requested. What it may read and what it may never write is set by how access is configured.
- An independent check. Something other than the agent verifies the agent’s work.
- A record. Everything it read, proposed and checked is kept, so the work can be reviewed.
- A person in charge. Someone accountable reads the evidence and applies the change.
- An agreement about data. Which tools may see which data is settled in writing before anything starts.
If a vendor cannot tell you how each of those five is handled, the answer is that they are not.
How to start small
Pick one task that is tedious, frequent and easy to check. A weekly reconciliation is ideal: you already know what the right answer looks like, so you will know quickly whether the agent is earning its place.
Run it beside the manual process for a few weeks. Compare the results. Keep the person in the loop even after it has proved itself, because the point was never to remove the person. It was to give them back the hours they spend reading rows.
If you want help choosing that first task, or a plain opinion on whether it needs AI at all, get in touch. Sometimes the honest answer is that ordinary code would do it better, and I will say so.