Why AI-first learning platforms will get quietly worse

When a company's AI bill climbs faster than its revenue, it has three options: raise prices, cap usage, or switch to a cheaper AI model. The third option is the one to watch if you're being pitched an AI-based learning or assessment platform.

Earlier this year Microsoft, facing internal AI coding costs that had grown faster than planned, canceled most Claude Code licenses in its Experiences and Devices division and moved engineers to its own lower-cost tool. After exhausting its 2026 AI tools budget in four months, Uber moved to spending caps and tighter controls. Even Salesforce, which sells AI agents rather than buying them, has rewritten how it charges for Agentforce more than once as it tries to find a sustainable model. Rationing is what companies do when the meter outruns the budget. There is no reason to expect edtech vendors to be exempt.

AI tutors, AI study tools, and live test engines spend tokens in real time to generate and deliver each question or response as the student works; their running quality is tied to whichever model they can afford to run that month. When that bill climbs, rationing follows — and inside an AI tutoring product, it looks like this. The vendor swaps the frontier model for a cheaper tier. Responses get shorter, or shallower, or subtly less accurate on hard material. Context windows shrink, so the tutor remembers less of the student's history. None of this is announced. Say you're an instructor who evaluated an AI-based tool during a spring demo, when the vendor was running its best model at a loss to win the adoption for fall. The product page still says the same thing that it said in March. But now the platform is requiring your students to pay for something materially different, and you've got no way to see the change.

Instructors and administrators ask vendors about accuracy, alignment, and integrity. I'd add a question to the list: is the pedagogical quality of this product going to remain the same, or does it depend on which model the vendor can afford to run this month? The tools being marketed hardest right now mostly have the second answer, and the question will only become more relevant over time.

As the founder of an edtech company, I can make a structural promise that no metered, AI-dependent product can make: PracticeQuest's quality will not silently degrade because of a budget decision. Students working with PracticeQuest get a fresh question every time, but producing the questions doesn't run a model or burn tokens. There's no meter climbing behind the practice, and no bill that a bad quarter could force me to trim by serving a cheaper model.

Every question a student sees comes from content databases and question templates designed, built, and reviewed by a human subject-matter expert. That doesn't mean the library sits still. The source material and question templates are constantly under review, expanding and improving, and AI helps me do that work upstream, in the workshop. But that improvement is human-directed and happens before students ever log in; it doesn't ride on real-time token usage at all. PracticeQuest changes from one semester to the next via user feedback and deliberate human revision, not from a quiet model downgrade that charges students or institutions the same price for a steadily degrading product. That is the hidden cost of putting AI in front of the student instead of behind the work: the price holds while the product slips, and you find out too late. References

- Microsoft cancels internal Claude Code licenses over cost; Uber's 2026 AI tools budget and spending controls — Fortune, May 22, 2026: https://fortune.com/2026/05/22/microsoft-ai-cost-problem-tokens-agents/

- GitHub Copilot moves to usage-based billing — The GitHub Blog, April 2026: https://github.blog/news-insights/company-news/github-copilot-is-moving-to-usage-based-billing/

- Salesforce revises Agentforce pricing toward consumption-based billing — Salesforce, May 15, 2025: https://www.salesforce.com/news/press-releases/2025/05/15/agentforce-flexible-pricing-news/