Skip to content

Operating Philosophy: Business

How to Measure the Real Value of AI in Education

By Veton KrasniqiAI Product Owner and EdTech product leaderNovember 13, 2026

AI features need to deliver measurable value at a sustainable cost. A practical look at adoption, learning outcomes, model economics, and the difference between a successful demonstration and a viable product.

Business is the second pillar of my People, Business, Technology framework. It asks what outcome we are pursuing, how we will measure it, and whether it is economically sustainable.

Good intentions don't fund themselves. In education, a feature that is technically impressive but unaffordable at scale helps no student, because it never reaches one.

Business defines the outcome. Goals, economics, and measurable results turn ambition into execution. In AI, both halves of that sentence have become harder than they look.

Value is still benefit over cost

Working in enterprise software, I learned to treat value as a simple ratio: benefit over cost. The weak point was always the benefit, because buyers routinely misjudge which features actually matter.

AI changes the other half. Traditional SaaS products already have variable operating costs, but generative AI can make those costs much more sensitive to individual product decisions. The model selected, the number of requests, the context sent with each one, the length of the output, and how much information repeats can all materially change the economics.

A feature that looks affordable in a controlled pilot can become considerably more expensive once it is available across an institution. Cost stops being an accounting detail and becomes a product decision.

Model the economics before launch, not after

I don't sign off on an AI feature as commercially viable until its cost is modeled bottom-up at realistic usage volume and a margin target is defended against it. That happens before launch, not after.

Institutions run on fixed, often public budgets. A price that only works at pilot volume becomes a renewal problem in year two. A renewal problem becomes a student who loses access to a tool halfway through a program.

The engineering discipline that follows is not exotic:

  • Route each request to the smallest model that can do the job well.
  • Cache context that repeats, such as a rubric that every request in a course shares.
  • Cap context windows deliberately, instead of accepting open-ended usage as a given.

None of those choices is visible to a student. All of them decide whether the product survives long enough for students to benefit.

Measure the outcome, not the activity

The second half of the discipline is deciding what success means before anyone builds. Many AI dashboards count activity: logins, prompts, documents generated. Activity is easy to measure and easy to grow. It says very little about whether anything improved.

In education, a better output is not always the same thing as a better outcome. A tool can raise homework scores while students retain less.

Three measures tend to get blurred together. Adoption tells you whether people use a tool. Efficiency tells you whether it saves time or money. Learning outcomes tell you whether students actually learn more. Each needs its own evidence, and strength in one says little about the others.

The same caution applies inside a single metric. When faculty accept AI-drafted feedback, that tells you what they chose to keep. On its own, it does not establish accuracy, feedback quality, or time saved.

So I push teams to name the outcome and the denominator up front: which faculty, which course, which measure, compared with what? When I led AI work for a national education platform, that meant defining KPI systems for model relevance and adoption reporting before calling anything a success.

Goals written as outcomes

The same discipline applies to planning. An objective written as an activity, such as "launch the feature" or "run a pilot," tells a team what to do but not why.

Written as an outcome, it says what has to be true for an institution afterward: faculty in a department spend less time on a task, or a pilot produces evidence a provost can act on. Outcomes force the uncomfortable question early. If this works exactly as planned, who is better off, and how will we know?

What to ask before you buy or build

  • What does it cost at our real usage, not at pilot volume, and what happens to the price as usage grows?
  • Which outcome will we measure, for whom, and by when?
  • Who owns the decision the AI informs? If the answer is "the AI," walk away.
  • What would make us stop? A pilot without an exit criterion is a commitment by another name.

Business is not the enemy of education's mission. It is what keeps a good idea alive long enough to reach the people it was built for.

A measurable, viable objective gives a team direction. Technology then has to deliver it within real constraints of cost, performance, and governance.

Part of my People, Business, Technology series. Next: Technology Should Solve the Problem, Not Define It. Also read: Why AI Adoption in Higher Education Starts With People.