If you want to understand which company will win in an EdTech product category, look past the front-end experience and ask where the assessment signal comes from. The front-end — the interface, the content presentation, the student interaction design — is visible and differentiable, which is why most EdTech product discussions focus on it. But the assessment signal is what makes adaptation possible, and the quality of the assessment infrastructure determines the ceiling on every adaptive experience the product can offer. A beautiful learning interface with a weak assessment layer will produce engagement data; it will not produce instructional intelligence.
Assessment in education serves at least three distinct functions that matter for anyone building AI tools in the sector. The first is measurement: determining what a student knows or can do at a point in time, typically to inform a reporting or placement decision. The second is diagnosis: identifying the specific knowledge gaps or misconceptions that are causing performance difficulties, to inform targeted intervention. The third is feedback generation: producing information that changes what the student does next — whether that's a worked example, a different problem type, additional practice on a prerequisite skill, or a branching discussion activity. Most EdTech products handle the first function reasonably well. The second requires a calibrated item bank and a knowledge model that can infer specific gap patterns from response data. The third requires a closed feedback loop architecture where assessment outputs directly drive instructional decisions.
Building assessment infrastructure that handles all three functions well is genuinely difficult and time-consuming. The calibration problem is probably the hardest: to know whether an item is appropriately difficult for a given curriculum node, and to use student responses to that item diagnostically, you need a large enough response population to estimate item difficulty parameters with statistical confidence. Building that response population requires deployment at scale, which requires having a product that institutions adopt first. The chicken-and-egg problem explains why well-calibrated assessment item banks have historically been assets owned by large publishers and testing organisations, not by EdTech startups. Learnosity's position in the market — an assessment engine API with a large installed base and years of calibration data — is defensible precisely because that data accumulation is not something a new entrant can shortcut.
The AI layer changes what's possible with assessment data in specific ways. Large language models can generate new assessment items at scale and tag them to curriculum standards with reasonable accuracy, substantially reducing the cost of expanding an item bank. They can also generate item variants — alternative surface-level presentations of the same underlying construct — which is important for preventing the item exposure effects that degrade the validity of repeated-use item banks. Retrieval-augmented generation approaches can be used to provide explanatory feedback on student responses that is specific to the item and the student's answer, rather than generic "try again" or "correct" feedback. These are genuinely useful applications, but they all depend on having the underlying assessment infrastructure — the item bank, the curriculum tag taxonomy, the response data schema — that allows AI generation and AI feedback to be grounded in curriculum-relevant constructs rather than free-floating language model outputs.
The investment thesis we hold about assessment infrastructure is that the companies which own the calibrated item bank, the curriculum knowledge graph, and the response data model are building a position that compounds over time in ways that application-layer companies don't. Every additional institutional deployment enriches the calibration data. Every new curriculum standard integration extends the tag taxonomy. Every AI tutoring product that integrates the assessment API adds to the response corpus. The network effects are real, if slow to build. For early-stage investors, the implication is to look for assessment infrastructure companies that are making the right long-term architectural bets — open APIs, standards-aligned curriculum tagging, psychometric rigour — even when early-stage traction looks slower than a consumer-facing application would produce.