Admissions is a product: what we changed at KBTU

The admissions office at a university is a funnel. Applicants enter at the top, documents move between people and systems in the middle, and offers go out at the bottom. If you squint, it's a product pipeline — which means you can treat it as one.

Before

What we had at KBTU before the rework looked like most university admissions operations I've seen: email and in-person paperwork at the front door, spreadsheets in the middle, and a lot of human attention moving between them. Errors in data entry were high, staff hours per application were high, and applicants didn't have a clean path to ask a question and get a useful answer.

What we did

Three things, in order:

1. Data governance first. Before any of the AI work, we defined what "the data" actually was — what we owned, who could write to it, and where it lived. Without that, any automation on top was going to be fragile.
2. A centralised processing pipeline. Moved the document flow into a single pipeline with states we could measure. Processing time came down by more than 30%. Data-entry errors dropped to about 1% once the manual re-keying step was removed.
3. AI agents on top. Only after the pipeline was stable did we put AI-powered agents in front of applicants and staff. The agents handle repeat questions and basic validation; people handle the cases that actually need a person.

What I'd tell you to copy

Don't start with the AI. Start with what the process looks like when you strip out the spreadsheets — who owns what, what states a document passes through, where the hand-offs are. If that's clean, a fairly ordinary agent can cover a lot of ground. If it's not, the agent is going to give confident answers on top of bad data, which is worse than doing nothing.

Outcomes

Document processing time −30%, manual staff input −70%, data-entry errors ≈1%, call-center text-based interactions roughly halved once the agent went live.