The thirty-billion mistake: when edtech mistook engagement for learning
A new AEI report puts US edtech spend at roughly $30 billion a year with limited evidence of effectiveness. The deeper diagnosis — picked up by World Bank voices and a growing chorus of researchers — is that the category optimised the wrong variable. It measured engagement and called it learning. The next category will not make the same mistake.
An American Enterprise Institute report this month put US school edtech spend at roughly $30 billion a year — projected to nearly double over the coming decade — alongside the uncomfortable observation that the evidence base for effectiveness has not kept pace. Cristóbal Cobo and others writing from inside the development-finance and education-research community have picked up the same thread: a category that grew very fast, spent very large, and rewarded the wrong variable.
The instinctive read is that edtech failed. That is too generous to the diagnosis and too harsh to the people who built it. The category did not fail at what it optimised for. It optimised for the wrong thing — and then the wrong thing became the metric the market paid against, the metric the procurement officer asked for, and the metric the next product had to beat to get bought.
How engagement became the proxy
In the first wave of consumer software, engagement was a reasonable proxy for value. If users came back, the product was probably useful. The pattern transplanted directly into edtech, because the same designers, the same investors and the same product playbooks moved across. Minutes on platform, daily active learners, streaks, badges, completion rates, lessons started.
Every one of those metrics is easy to instrument, easy to put on a sales deck, and easy to grow. None of them is a measurement of learning. A student can complete a unit without mastering it. A class can have ninety-five percent daily active use and a flat attainment curve. A platform can post strong engagement numbers in a system whose learning outcomes are quietly drifting backwards.
The market did not stop the drift, because the market was buying the proxy. Engagement metrics fit neatly into renewal conversations, board reports and case studies. Outcome measurement is slower, harder, jurisdiction-specific, and frequently inconvenient for the product it is measuring. The system selected for what it could measure cheaply.
Why this is structural, not cultural
It is tempting to read this as a failure of intent — vendors who knew better and chose engagement anyway. A few did. Most did not. The deeper cause is structural: the architecture of the products made outcome measurement very difficult, because the products had no internal model of what the student was supposed to be learning.
A platform that does not know which learning outcome a particular task is evidencing, how that outcome sits inside the national curriculum, what prerequisite outcomes it depends on, and what would constitute mastery, cannot report on learning even if it wants to. The most it can report on is activity. The procurement officer, faced with a vendor who can show engagement and a vendor who can show nothing, buys engagement. The category settles around the variable the architecture can produce.
What the AEI and World Bank framing is really saying
Read carefully, the $30 billion critique is not an argument against technology in education. It is an argument against rewarding the wrong variable for long enough that the variable becomes the category. The remedy the report points to — closer alignment between research, procurement and what actually gets deployed at scale — is the same remedy ministries are now articulating in their own language: buy the outcome, not the platform; buy against the curriculum, not against the demo.
Cobo's framing adds a sharper edge. For low- and middle-income systems, the cost of the proxy is not just wasted spend. It is a generation of learners whose time on a screen was confused with progress, and a procurement habit imported from richer systems that the receiving country can least afford to repeat.
What a category measured on learning looks like
A category that takes the critique seriously looks structurally different from the one being critiqued. Three shifts, in particular, are non-negotiable.
First, the unit of measurement changes. Not minutes on platform, but mastery against specified outcomes. Not lessons completed, but evidence collected against the curriculum the deploying authority has actually committed to teaching. Engagement remains useful as an operational signal — it is not useful as a sales metric.
Second, the unit of integration changes. A system that reports on learning has to be built on top of the curriculum as structured data — outcomes, prerequisites, progression, assessment evidence — not on top of a content library tagged for marketing convenience. This is the layer most of the previous category never built, and it is the layer the next category cannot avoid.
Third, the unit of accountability changes. The buyer stops being the procurement officer choosing between dashboards and becomes the system owner accountable for attainment. That buyer asks different questions. What did learners actually know at the end of the year that they did not know at the start? Against which outcomes did the system contribute, and against which did it not? Where did teacher cognitive load fall, and where did it rise?
Why this is not a critique we deflect
It would be easy, and dishonest, to treat the $30 billion finding as a problem for other companies. It is a problem for the category. Any organisation building educational AI now inherits the credibility deficit the previous decade created. That deficit is not closed by a better demo. It is closed by being measurable on learning — publicly, repeatedly, against the curriculum of the deploying authority — and by accepting the asymmetry that comes with it: harder to sell, slower to scale, more expensive to build, and considerably more durable once it works.
The companies that take that trade seriously are the ones the next decade will be built on. The ones that keep selling engagement will keep finding buyers for a while longer, and will keep being absent when the ministry comes back the following year and asks what changed.
The investor read-through
The $30 billion observation is not a reason to write the sector down. It is a reason to underwrite it differently. The defensible position is no longer the company with the best engagement metrics; it is the company whose architecture makes outcome reporting structurally possible — curriculum as structured data, evidence collected against specified outcomes, governance the ministry can actually enforce, and a commercial model that survives being measured on attainment rather than activity.
Capital that backs the previous category's metrics will keep funding renewals until the renewals stop. Capital that backs the next category's architecture is funding the substrate every credible educational AI system will eventually have to sit on.
"Thirty billion dollars a year is not the cost of edtech. It is the cost of measuring the wrong variable for a decade. The next category will not be measured on engagement; it will be measured on what learners actually know — and the architectures that cannot answer that question will quietly stop being bought."
aime is built to be measured on the right variable.