AI agents for applicants, and what halved our text volume

The first thing I did when we stood up AI-powered agents in the admissions office wasn't to pick a model. It was to figure out what questions we were actually getting — and which ones weren't worth a human's time.

What the agents handle

Two audiences, two agents:

• For applicants — deadlines, documents, program eligibility, where a given application is in the pipeline, how to fix common submission errors. Everything a well-briefed first-line operator would handle on the phone.
• For staff — pulling the applicant data a staff member needs for a specific case, running the repetitive validation checks (is the passport scan legible, does the transcript match the declared grades, does the document set look complete for this program), and flagging the outliers for a person.

What they don't

Anything with a policy exception. Anything where the right answer depends on who a candidate actually is, not which bucket they fall into. Anything where the cost of a wrong answer is high enough that "pretty confident" isn't good enough. The agents route those to the humans they should have been going to in the first place.

The measurable bit

Call-center text-based interactions dropped by about 50% after the agent went live. That isn't the agent "winning" — it's the agent covering the conversations that shouldn't have been taking a person's attention in the first place. The remaining 50% is more valuable to humans because it's the half that actually needed a human.

What I'd flag for anyone copying this

Two things. First, make sure the data the agent reads from is clean — see the admissions-pipeline post for that part. Second, instrument the hand-offs. When the agent escalates to a person, you want to know why: was it uncertainty, a policy case, a user who just pressed the button? The escalation logs are where the next iteration of the product lives.