Saying Yes Carefully: Encouraging AI Experimentation Without Buying the Hype

Trying an AI tool now costs almost nothing. A division head can build something useful in a chat window over a weekend. An agent tool can be pointed at real files and running on a schedule by Tuesday. Technology leaders should encourage more of this, not less.

Keeping a tool is a different question. The moment something moves from an experiment to a system the school depends on, you have taken on data exposure, recurring cost, maintenance, and the risk that the vendor behind it does not exist in three years. Much of the current market is priced on the assumption that AI capability is scarce and proprietary, and a good deal of it is not.

This session gives technology leaders a way to hold both positions at once. We will walk through five layers of AI implementation, from a consumer chatbot through agent tools that act on real systems, custom code against a model API, and third-party products built on those same APIs. One ordinary school workflow gets carried through all five so the cost and risk profile of each is concrete.

From there, three questions determine where something should live: what happens if the output is wrong and who sees it, whether it needs to outlive the person who built it, and where the data actually goes. Most work should stop earlier in the stack than instinct suggests. We will spend real time on the economics, including how to tell a substantive AI product from a wrapped API, what you are legitimately paying for when the answer is a wrapper, and how to avoid building on capabilities that will be commodity features within a year.

Throughout, I will share AI proposals my own company adopted, along with the ones we rejected and what the rejections were actually about.

Learning Objectives:

Participants will be able to:

-Distinguish the cost of trying an AI tool from the cost of supporting one, and apply that distinction to proposals from colleagues and vendors.

-Place a given AI use case at the appropriate implementation layer using output risk, persistence, and data sensitivity as criteria.

-Evaluate whether a third-party AI product offers durable value beyond the underlying model API, using a specific set of vendor questions.

-Encourage staff experimentation while maintaining defensible limits on data exposure and long-term commitments.

Facilitator

Joey Nutinsky, Co-Founder and CEO, Ruvna