Skip to content
GEERD
← All insights

Leadership

Seven questions to ask before letting an AI talk to your applicants

An AI demonstration always goes well: you are shown the questions it can answer. These seven are about the rest, and they should be put to everyone, ourselves included.

6 min readBy the GEERD team

Institutions being approached right now all get roughly the same demonstration: an agent answers three well-chosen questions in seconds, in impeccable French, and the room is convinced. That is to be expected: it is the easiest exercise in the whole chain.

What decides the rest is not visible in a demonstration. It is in what the supplier wrote down before plugging anything in. Here are the seven questions that make that document speak, each with what a serious answer sounds like.

The seven questions

  1. 01Where do the facts in the reply come from?

    A serious answer names a source you hold: a knowledge document per programme, written and maintained by your team, outside which the agent has nothing to say. If the answer is “the model knows your website” or “it learns from your conversations”, one day it will talk about a course you closed two years ago.

  2. 02What does it do when it does not know?

    There are only two possible answers and only one is acceptable: it stops and hands over. Ask to see what the team receives at that moment. If the message does not say what was missing, the hand-over is worth little: nobody will know what to fix.

  3. 03Who sends, in the first weeks?

    The right answer is: your team. The agent drafts, a person reads and sends, and that lasts as long as it needs to. A supplier who offers full automation from day one is making you carry the cost of their tuning.

  4. 04What is checked before sending, and by what?

    Ask for the list. A real list is boring and precise: do the cited facts exist, is the language right, does the reply promise an admission, does it judge the applicant, does it repeat the previous turn. Then ask what happens when a check fails: if the reply goes out anyway “with a warning”, the check is decorative.

  5. 05What does it do besides reply?

    This is the question that separates an agent from a chatbot, and it has two halves. First: what actions can it take on your records: record an interest, open a task, write a note? Second, and more important: is the applicant told? The answer must be no, and the reply sent must be identical with or without the action. You do not announce to a family that they have just been ranked.

  6. 06What will I be able to read back in six months?

    A complete, timestamped conversation, with what the agent relied on and what was blocked. That is what will let you answer a family who disputes something, and settle an internal argument about “what the robot said”. Without it you will have to take your supplier’s word, and your teams will have to take yours.

  7. 07What costs, and what stops?

    An agent consumes on every message, useless ones included. Ask how the spend is bounded and what happens when the limit is hit: the agent should stop and say so, not carry on and bill. An AI invoice that surprises you at month end is a design flaw, not an accident.

What the answers actually tell you

Taken together, these seven questions do not measure the quality of a technology. They measure whether someone, at the supplier, took the time to write down what the machine is allowed to do, before plugging it into families.

If the answer to any one of them is “the model handles that”, you do not have a supplier, you have a subscription. The difference will show on the day of an incident, that is, at the worst moment of your admissions campaign.

Our answers, to the same seven questions

We deploy an agent in admissions, and these seven questions are exactly the ones we want to answer in writing before going live. The page that follows gives our answers in detail: where the facts come from, what is checked, what is forbidden, and what can be read back.

Agentic AI in an institution

Put the seven to us

A free diagnostic, half a day. We look at what your institution has written about itself, what your teams answer today, and whether a machine has any business in it. Sometimes the answer is no, and we say so.