E-commerce trends
How to Reduce Cart Abandonment Without More Discounts
Most cart abandonment advice starts with discount emails. The recoverable losses are usually sitting inside your own checkout, where you can measure and fix them.
Ask an edtech founder what question they hear most from their board this year, and it is some version of "what's our AI tutor strategy?" The pressure is understandable. Students are already using AI on their own, a handful of rigorous studies have reported learning gains from AI tutoring that would have sounded like marketing copy three years ago, and the price of always-available tutoring has fallen to less than a streaming subscription. What is missing from most of the coverage is a sober read of the evidence and an honest account of what it takes to turn a demo into a product. That is what this post is for.
What AI tutoring actually is (and what it isn't)
An AI tutor is software that teaches the way a good human tutor does: it asks guiding questions, checks the student's work, adapts to their pace, and refuses to simply hand over the answer. The research literature calls these intelligent tutoring systems, and the idea predates ChatGPT by decades; large language models are what finally made them conversational enough to hold a student's attention. That "refuses to hand over the answer" part matters. A raw chatbot will happily write the essay for the student, which is homework laundering, not tutoring. The systems producing real learning results are built on top of large language models but constrained by carefully designed teaching behavior, curriculum-aligned content, and guardrails that keep the conversation on task.
That distinction, between a general-purpose chatbot and a purpose-built tutor, turns out to be where most of the engineering effort and most of the educational value lives.
The evidence is stronger than you might expect
Two studies are worth knowing in detail, because they bracket the market nicely: one at an elite US university, one in public secondary schools in Nigeria.
At Harvard, physicists built a custom GPT-based tutor for an introductory physics course and ran a crossover experiment with 194 undergraduates in fall 2023. Students who worked through a lesson with the AI tutor at home showed learning gains roughly double those of students in an expertly taught active-learning class, and they reported feeling more engaged. Worth stressing: the comparison class was not a boring lecture. It was a research-refined course the university considers one of its best-taught.
The World Bank result is arguably more striking because of its setting. In a six-week after-school pilot in Nigerian public schools during June and July 2024, first-year senior secondary students used Microsoft Copilot as an English tutor with teacher support. The evaluation found learning gains of about 0.3 standard deviations, equivalent to nearly two years of typical schooling, and the program outperformed 80% of education interventions previously tested in developing countries. Girls, who started behind, gained the most. Every additional session attended added measurable benefit.
Both studies come with caveats a serious buyer should keep in mind. They were short. They measured first exposure to material, not long-term retention. And both tutors were heavily shaped by educators, not raw chatbots dropped into a classroom. But as evidence goes, education technology has rarely had it this good, and the research base predates the current generation of models.
Demand is not hypothetical either. In a Pew Research Center survey of 1,391 US teens conducted in autumn 2024, 26% said they used ChatGPT for schoolwork, double the share from 2023. Students are not waiting for the market to mature.
The economics have flipped
Private human tutoring has always been the most effective intervention money can buy, and the least affordable. AI has broken that price floor. Khan Academy's Khanmigo, one of the most visible products in the category, costs $4 a month for learners and is free for teachers.
For anyone running an edtech or corporate training product, that number is the strategic fact of the decade. When credible one-on-one help costs a few dollars a month, "content library plus quizzes" stops being a defensible product. Personal guidance becomes the baseline students, parents, and L&D buyers expect. The question shifts from whether AI tutoring belongs in your roadmap to how it gets there without wrecking your margins or your reputation.
Why the product is harder than the demo
Here is the part most teams miss: nothing in the studies above was achieved by pointing users at a chat window. The Harvard tutor worked because instructors encoded lesson structure, pre-vetted feedback behavior, and pedagogy into the system before any student touched it. The Nigerian pilot wrapped the model in teacher supervision and a structured after-school program.
We see the same pattern at Encomage in commercial AI work: the model is the smallest part of the project. In our AI integration projects, the effort goes into connecting the model to real product data, constraining what it may and may not do, evaluating output quality continuously, and keeping per-conversation costs predictable at scale. An education product adds its own layer on top: curriculum alignment, age-appropriate guardrails, moderation and parental visibility, and privacy rules for minors that vary by market. None of that is glamorous. All of it decides whether your tutor produces the outcomes the research promises or becomes an expensive answer machine.
Token costs deserve their own line item in any plan. A tutor that holds long, Socratic conversations burns far more compute per user than a support bot that closes tickets quickly. Pricing, session design, and model selection have to be worked out together, which is a product decision as much as an engineering one.
Build, buy, or wrap
Edtech teams generally face three options. Buying an off-the-shelf tutor gets you to market fastest and teaches you what your users actually do with it, at the price of differentiation and control over data. Wrapping a frontier model in your own prompts is the tempting middle path, and it is fine for a pilot, but thin wrappers are easy to copy and hard to keep on-curriculum. Building a properly scaffolded tutor into your own platform is slower and costs real money, but it is the only route where the pedagogy, the data, and the learning outcomes belong to you.
Our advice to most founders is unfashionably boring: run the smallest honest experiment first. A discovery phase and a focused MVP that puts a constrained tutor in front of a few hundred real learners will tell you more than any vendor deck, and it establishes the evaluation habits (measuring learning, not just engagement) that the eventual production system will need. We have watched the same lesson play out in e-commerce AI, where agentic systems only earn trust once governance and measurement are in place.
Where to start
The research window is open. The evidence says AI tutoring works when it is built with pedagogical discipline, the cost of delivering it has collapsed, and your users are already practicing on free chatbots. The teams that win the next few years of edtech will be the ones that treat the tutor as a product with measurable learning outcomes, not a feature checkbox.
If you are weighing how to add AI tutoring to a learning platform, or how to keep quality and costs under control once it ships, this is the kind of problem we work on at Encomage. We typically start with a short technical discovery: what your data supports today, what the guardrails need to be, and what the first measurable pilot should look like.
