Education AI has entered a new stage.
For many years, the largest-impact systems were “silent AI”: recommender systems that selected the next practice problem, mastery models that estimated what a learner did not know, and analytics that could identify missing assignments to teachers. Now we also have “loud AI”: large language models that can tutor in natural language, write lesson plans, create questions, and converse with students around the clock.
This presents a real opportunity and a real risk.
The potential is that good education AI can make learning more personalized, more adaptive, and more equitable – even in large classes, or poorly resourced environments. The trap is that many applications of AI boost short-term performance rather than learning, and can produce a “false mastery” effect, where students generate better outputs but don’t develop enduring skills. That distinction — performance versus learning — features strongly in the new OECD policy and research brief, which explicitly cautions that delegating cognitive work to general-purpose chatbots can lead to better task performance but not necessarily to better learning, particularly if access is taken away during exams.
An effective EdTech team, therefore, has to think like this: every AI feature had better answer three questions.
Firstly, what learning problem does it solve, in a way that can be measured not just by how engaging it is? Secondly, what data and pedagogical assumptions underpin it, and are those assumptions validated in the real world? Thirdly, what type of risks does it introduce—privacy, bias, integrity, safety—and how can those risks be managed over time?
This post is a deep dive, implementation-focused overview of the ten actual solution types for AI that are relevant to current educational platforms, organized by adaptive learning, assessment, and student support. It is “build-ready” – who each solution is for, what it takes to implement, what could go wrong, what evidence exists, and what governance is increasingly expected.
Adaptive learning is the foundation of many successful EdTech products, even if it does not claim to be “AI.” It is the system that determines which concept a learner should be exposed to next, which problem should come next, and whether review of previously covered knowledge is warranted so that it can be retained in long-term memory.
Technically, an adaptive learning engine is the combination of a content graph and a learner model. The content graph represents skills and dependencies (prerequisites, degree of difficulty, learning goals). The learner model predicts mastery, forgetting, and uncertainty. It then applies those estimates to customize sequencing, pacing, and practice. Today’s reviews of adaptive learning solutions make it clear that these are not basic “recommendation widgets” – they are sophisticated pedagogical systems that embed assumptions about the nature of the learning process and the way mastery can be inferred from learner engagement.
In reality, “magic” is never the model type. The differentiating factor is content organization. Adaptive systems break down when things are inconsistently tagged, skill definitions are inconsistent, or the graph is too coarse (“Algebra”) or too fine (“Solving 2-step linear equations with negative coefficients under time pressure”). The scale is the right granularity for your product. When you’re building a K–12 practice tool, you generally want to be able to model skills more finely for more detailed feedback. When you’re developing a higher education courseware product, you may be more concerned with broader competency outcomes.
But the most important detail to implement is that adaptive learning needs to be optimized for learning outcomes, not for time-on-platform only. A “smart” engine can increase engagement by always serving up easier questions, but that can slow learning. Alternatively, it can accelerate learning through productive struggle, but that can decrease engagement if the UX is not supportive. You need to select your goal consciously and measure it truthfully.
A good adaptive engine also allows for teacher and curriculum control. The system would be adapting within guardrails – the curriculum sequences, the required units, the assessment windows, and teacher-selected priorities. However, if the engine is completely autonomous, it may come into conflict with reality in the classroom and be discarded.
Lastly, a cold start is also critical for adaptive engines. New learners, new courses, and new skills suffer from weak historical data. A practical pattern is to start with evidence-based heuristics and add personalization as signals accumulate.
An intelligent tutor explains how to solve a problem if the adaptive learning system selects the next problem.
Intelligent Tutoring Systems (ITSs) represent one of the most investigated categories of AI-in-education. They usually offer immediate feedback, hints, and scaffolding at the step level, often relying on a model of the domain (a “knowledge model”) and a model of the learner’s current state.
The evidence base is significant, but subtle. Meta-analyses and reviews have shown that ITS can enhance learning outcomes relative to a wide range of traditional or non-adaptive methods, with effect sizes that depend on the nature of the design, subject area, comparison group, and outcome measure. The practical interpretation isn’t “It always works.” It is “tutoring-like scaffolding can work at scale when it is informed by how students actually err, and when the tutor’s feedback is timely and specific.”
ITS works most effectively when the skill granularity is matched. Step-based tutors are great for procedural domains, such as math, coding, grammar, and some science problem-solving. They usually have trouble when the targeted skill is open-ended writing, creativity, or complex argumentation, unless accompanied by very strong rubrics and human evaluation.
Large language models alter this space by making tutors conversational. They are allowed to ask questions, rephrase explanations, and change tone. Khan Academy has openly talked about using OpenAI’s GPT-4 to run “Khanmigo,” a tutor that leads learners like a tutor rather than just giving answers. Duolingo has outlined applying machine learning models (including its “Birdbrain” model) to make difficulty personalised, and on “Duolingo Max” has brought GPT-4 powered experiences such as roleplay and explanations.
These product examples are instructive as they illuminate the core design decision for LLM tutors: do you allow the model to answer freely, or do you constrain it to be a scaffold? The OECD’s 2026 Digital Education Outlook puts a strong spin on this message: general-purpose GenAI can enhance task performance, but in the absence of pedagogical guidance, it might not lead to learning gains; education-specific solutions, employed purposefully, are more likely to result in enduring improvement.
Implementation-wise, LLM tutoring is not “just plug in an API.” You need concrete content. The tutor should be referencing the course materials, not making it up. You need a pedagogy policy that says what the tutor is allowed to do: when they can give a hint, when they must ask a question, when they can reveal an answer, and when they must escalate. You also need a strong evaluation harness: test the tutor on known student error patterns, assess accuracy, assess “learning-friendly behavior” (is it scaffolding or solving), and track hallucination rates over time.
And one more point that almost all teams miss: tutoring systems generate new types of data. Tutor dialogues are a rich signal for misconceptions, affect, and metabolic behavior. But that also creates privacy risk, because student chats can include sensitive personal information. And that takes us to the next solution buckets: assessment and support need to be designed with governance.
The majority of learning occurs in the process of formative assessment. It’s not the unit test. It’s the continual cycle of try, feedback, revision, and thinking.
AI is powerful here because it can shorten the lag time between a student’s action and useful feedback. However, the key crux is “meaningful feedback.” If a system just tells a learner “wrong” — it’s not very useful, is it? A system that tells them why they are wrong and what to try next is tutoring.
Formative feedback systems can be divided into two approaches in practice.
The first is structured: you create items with predetermined solution steps or rubrics, and the system compares student submissions to those templates. That’s typical with math, coding, and structured problem-solving. It is a robust, explainable, and easier to validate methodology.
The second level is generative: the system scans student work and produces feedback, hints, or subsequent steps in natural language. This is where big language models get both easier and more dangerous. It’s easier to comment on open-ended work en masse. It’s more dangerous because feedback can be incorrect or contradictory, and because it can inadvertently teach misconceptions.
One practical application of generative techniques that is increasingly attracting attention is automatic question generation. Recent studies investigate the extent to which large language models can generate questions of various cognitive levels, which include higher-order levels according to Bloom’s taxonomy, and highlight the importance of expert review and quality assurance. This is relevant, since the quality of assessment is not just about being correct, but whether you are assessing what you want to assess, at the cognitive depth you intended.
Implementation Insight: When it comes to formative feedback, “accuracy” is not sufficient. Tone and pedagogy must remain consistent. A tool for student support that shames or baffles students may dampen motivation and raise dropout rates. That is why many education systems stress human-centered design and teacher engagement. The UNESCO recommendations on generative AI in education and research explicitly situate GenAI usage within a humancentred approach that upholds agency, inclusion, equity, and accountability, and calls for regulation and policy frameworks to facilitate the ethical and meaningful application of GenAI.
You’ll probably want to build in “feedback quality monitoring.” If the system drifts, teachers and students will eventually. You need a mechanism for teachers and students to report incorrect or useless feedback, and those reports must feed into retraining, prompt changes, or rule changes.
Automated scoring is probably the most attractive AI application, since it could save time for assessors and allow large-scale assessment. It is also extremely sensitive, since grading affects life outcomes and can incorporate bias.
Automated Essay Scoring (AES) has been around for decades. The evidence base is large and includes findings supportive of, and strongly critical of. The research of ETS has identified various validity threats and potential failure modes for the automated scoring system if used as the only scoring method in a high-stakes testing context. More recent scholarship continues to explore AES methods and limitations, including how systems might overemphasize surface features and have difficulty dealing with meaning, coherence, and relevance unless thoughtfully designed to do so. Fairness in AES research has also been broadened, which includes work on individual fairness assessment as opposed to group-level comparisons.
In terms of EdTech products, there are two secure ways of employing auto-grading.
One is the low-stakes formative feedback: the system outputs feedback and a provisional score, but the grade that “matters” is human-reviewed or relies on multiple signals. This can help reduce teacher workload and enhance revision cycles without the model being the ultimate arbiter.
The other is tightly-regulated high-stakes scoring with extensive validation, human review, and well-defined policy controls. If your product resides in a jurisdiction that considers educational assessment systems as high risk, then you must anticipate stronger governance. Such as, the EU AI Act’s high-risk annex explicitly includes AI systems for evaluating learning outcomes or for guiding learning processes, as well as systems for monitoring and detecting prohibited behaviour in tests.
The details of the implementation are what make automated grading defensible, including how the rubrics are defined, how student work is modeled, how the model’s uncertainty is exposed, and how the appeals process works. In education, a model that is not contestable breeds distrust. Your product has to enable contestability — provide students and teachers with a transparent reason code for the score, show the evidence that was considered, and include a human review path.
An Important technical point: if you use LLMs to grade, do not engage in “single-shot grading.” LLM outputs are not deterministic. The common safety pattern is multi-pass scoring with consistency checks, rubric anchoring, and constraints that force the model to cite evidence from the student’s answer. You need adversarial testing too; students will try prompt injection and “grading hacks,” and at least some will succeed unless you engineer against it.
Predicting student outcomes is among the most useful and potentially damaging fields of education AI.
That’s valuable because dropout and failure rarely happen abruptly. They tend to be apparent in trends: work not turned in, low participation, work turned in late, misconceptions repeated, poor quiz performance, lack of forum participation, and pacing changes. It is dangerous because the labels of “at-risk” can become self-fulfilling. When it tags a student and educators subtly drop their expectations, the system does more harm than good.
Contemporary work in learning analytics stresses that predictions are only part of the story and that interventions are equally important. The Journal of Learning Analytics published a systematic review on learning analytics-based interventions in the context of LMS, which illuminated design considerations for interventions (if not models) in that specific context. A systematic review of predictive models in education similarly depicts these models as means to forecast performance and detect students at risk, while implicitly reminding readers that the use of models within educational systems ultimately influences outcomes.
The main execution pattern is a three-stage pipeline.
The first stage is instrumentation. Your LMS and tools need to have the same definitions for events. If you can’t trust your “assignment submitted” event, you can’t build anything on top of it. That’s the case with interoperability and analytics standards for EdTech platforms. The 1EdTech’s LTI 1.3 specification level is designed to allow third-party tools to be integrated with LMS platforms in a standard, secure manner and provide consistent launches and services. Caliper Analytics is a learning standards technology that seeks to standardize learning activity event vocabulary to enable the collection and exchange of usage data between digital learning tools that can be used to drive learner success and drive decision-making.
The second stage is prediction. The risk is predicted employing the model, although with uncertainty. A well-developed system outputs not only a “risk score” but also a “why the score changed,” linked to interpretable features such as missed work, low mastery, low pacing, and low participation. But if the system is a black box, teachers will not trust it or will use it improperly.
The third stage is the orchestration of interventions. This is where the majority of projects die. A warning that is ignored is useless. A warning that results in a negative reaction is detrimental. Intervention still needs to be evidence-driven – tutoring support, targeted practice, teacher outreach, parent outreach, advising, study planning, or mental health resources, as the case may be. The system needs to know if the intervention took place and if it helped, because that feedback is what makes the system better over time.
In the early warning systems, they must not become punitive on ethical grounds. They should never be used to deny access to or tag students indefinitely. They need to be established as support distribution systems with explicit guardrails.
Regulating these systems tends to dwell in “high-risk” land when they assess learning outcomes and affect chances. If you do business in Europe, consider them as high-risk AI under the EU AI Act, recognized in the Education and Vocational Training sectors.
Student support isn’t just tutoring. Everyday frictions that lead students to drop out include uncertainty about deadlines, policies, where to find resources, how to register, what to study next, and who to call for help.
AI can eliminate such friction by offering on-the-spot answers and guidance. The danger is that support staff may unwittingly become answer machines that short-circuit learning or offer incorrect policy advice.
A prominent historical example study is “Jill Watson,” a virtual teaching assistant implemented in an online course at the Georgia Institute of Technology, detailed in a chapter about using a conversational TA in course forums starting in 2016. In recent times, scholars have been building still more of these conversational virtual teaching assistants, intelligence-powered by today’s LLM capabilities, including publications on extending Jill Watson concepts with newer models.
The product design takeaway is steady: the best student support assistants can be defined by two things.
They answer “where is X,” “what is due,” “what is the policy,” “how do I submit,” “how do I appeal,” and “where can I get accommodations” with grounded, authoritative institutional content, easing navigation friction. They promote learning by leading students to solutions instead of giving them the answers, particularly in homework-adjacent contexts, mirroring the OECD’s focus on GenAI as a learning ally and not a shortcut.
In terms of execution, retrieval grounding, source citation, and access by role are needed. A student helper should not be able to see notes reserved for teachers. A teacher aide will have different tools and permissions. In higher ed, assistants often have to honor departmental boundaries. It’s less “AI” and more “systems design” — but it determines whether an assistant is safe.
You want escalation paths, too. The assistant needs to know when a question is out of scope, when the student is upset, and when the query refers to personal information. It must route to humans and log context appropriately.
Workload for teachers is a significant lever in education systems. Teachers’ lesson plan, differentiate, provide feedback, communicate, grade support, document, and complete administrative tasks. AI can assist, but only if it respects teacher agency and preserves quality.
The most recent U.S. Department of Education’s report on artificial intelligence and the future of teaching and learning was authored with the explicit intent of supporting sound policy-making, and it underscores ethical and equitable considerations as AI becomes integrated into EdTech and other common tools in education. The OECD’s 2026 Outlook also stresses co-design with teachers and the alignment of tools with learning science and educational aims, rather than simply unleashing generic chatbots with no instruction.
The real value-add cases for a teacher-copilot tend to be predictable.
One function is lesson planning assistance: producing draft lesson plans, differentiated samples, and formative assessments aligned to standards. Another is feedback support: producing comment starters, rubric-based feedback templates, and summarizing shared misconceptions identified in student work. Another is administrative summarization: Processing discussion threads, assignment submissions, or notes on meetings into effective recaps.
But that’s the difference between a helpful copilot and a harmful one.
Bad copilots generate plausible-sounding content that’s wrong, which teachers then unwittingly propagate. Or it generates generic content for a classroom that doesn’t fit the context. Or it adds to the workload by compelling teachers to check facts and rewrite everything.
Therefore, it needs parameters for a production teacher copilot: it needs to be rooted in the curriculum, mapped to local standards, and calibrated to the school’s policies. It also must allow for “teacher voice.”Teachers have a style. It is families’ perception to trust or not to trust if every message is corporate AI text.
The well-known systems treat teacher copilots as “drafting engines,” with strong context control, not full autonomous planning. They also record what was generated and what was edited, because that feeds back into quality improvement and accountability.
Education platforms are increasingly catering to a mix of learners, including multilingual students, students with disabilities, neurodiverse learners, and students with varying levels of background knowledge.
AI has the potential to enhance accessibility in a meaningful way, including speech-to-text, text-to-speech, translation, reading support, simplified explanations, and adaptive scaffolds. This is among the most societally impactful and ethically compelling applications of EdTech, but it’s also sensitive given that these systems interact with vulnerable populations and often handle sensitive data.
Policy advice is increasingly portraying the uptake of AI as a human-centred enterprise. The UNESCO guidance highlights inclusiveness, equity, cultural and linguistic diversity, and underlines the necessity of regulation and policy instruments at the national level to foster a safe and meaningful use of generative AI in education and research. The OECD has also released work on the uptake and integration of AI in education systems that focuses on school-based risks and risk mitigation.
The teams that adopt accessibility as a buzzword, “adding captions,” really miss out on the details of implementation here.
First, translation and simplification must be true to the original in instructional intent. If your tool simplifies a scientific explanation but the meaning is altered, you are creating a potential misconception. Secondly, the accessibility features must be developed in accordance with the rules of the assessment. If a student takes a test with text-to-speech, it needs to align with the accommodations policy. Thirdly, it must be impervious to bias; inclusion features are inclusion features. Speech recognition systems may vary in accuracy depending on how you accentuate your words and the dialect you speak, and translation may alter meaning.
A mature perspective is to view accessibility features as core components of pedagogy: test them with real learners, assess learning outcomes, and add educator review as appropriate.
AI has transformed academic integrity in two ways.
Students utilize AI to create answers, essays, and code. Institutions can employ AI to monitor exams, alert them to suspicious activity, and flag text generated by AI. Both directions are prone to failure modes.
The OECD’s 2026 Outlook highlights teachers’ concerns that AI undermines academic integrity by facilitating students to plagiarize, and stresses that assessment should be redesigned rather than based solely on detection-based measures. The high-risk annex of the EU AI Act explicitly contains AI systems for surveillance and detection of prohibited conduct during examinations in education, indicating that such systems will be subject to more stringent requirements.
Remote proctoring, and AI proctoring in particular, exemplifies the tradeoff rather starkly. Reported harms include invasions of privacy, false positives, and bias issues – these have been followed in incident databases and legal analyses. Even if the system is “correct,” it can still be harmful if it identifies behavior of neurodiverse people as suspicious, or if it causes stress that decreases their performance.
An effective approach to academic integrity is progressively shaping up as: mitigate the pressure of high-stakes surveillance, enhance authentic assessment design, and leverage AI as an assistant rather than an arbiter. In practice, this will involve more open‐book, process‐based assessments; oral defenses; multiple drafts; and assessments that focus on thinking rather than end products. The OECD underscores that effective task completion in GenAI does not automatically result in learning, which means that assessment design needs to differentiate between “I can produce an answer” and “I can reason.”
And if you’re going to use integrity tooling, design it with human review and transparent appeals as a risk triage mechanism rather than an automated accusation engine. Hold onto as little data as possible, limit retention, and be clear in your communications to students.
Manufacturing was taught something by Industry 4.0, which education is now relearning: Automation is worthwhile if it allows experts to do expert work.
In education, that expert work falls squarely on teaching and attending to students. So operational AI is not “back office fluff.” It can be one of the highest ROI areas, particularly in systems under staffing strain.
Day-to-day operational AI is the scheduling and timetabling, enrollment forecasting, capacity planning, automated communications, processing of transcripts and credentials, and helpdesk automation. In the OECD’s 2026 Outlook, it is stated explicitly that GenAI can be used to optimize backend processes, provide guidance, and perform administrative duties, but that policy frameworks must protect learners and uphold learning.
This category also applies to the knowledge management related to institutions: policies, program requirements, course catalogs, the accommodations process, and services to students. A well-engineered AI helper could ease the administrative burden and boost student navigation, which is one of the best predictors of retention in several programs.
Execution details: Operational AI systems have access to personally identifiable information. You have to design for privacy laws and for consent expectations. FERPA in the U.S. gives rights with respect to education records and limits disclosure of personally identifiable information, subject to certain exceptions and requirements. In Europe, the requirements of the GDPR for children and in the educational context bring further safeguards and duties regarding lawful basis, transparency, and consent in applicable cases. These aren’t “legal footnotes.” They are the product needs. They govern what data you can have, how long you can keep it, and what you need to tell.
Thus, a production-ready AI system has role-based access control, audit log, data minimization, and strong vendor management—since most EdTech platforms bundle several third-party tools.
An AI feature or tool, pretty much any AI feature, can increase clicks and time on the platform. That is not to say it improved learning. The OECD’s 2026 Outlook repeatedly highlights the difference between performance improvements and learning gains, and advocates for tools developed with pedagogy rather than generic chatbots. That’s why strong evaluation matters.
In EdTech, the best unit of truth is not just “user satisfaction.” It’s durable learning: delayed post-tests, transfer tasks, and performance without AI support. If you fail to measure these, you risk creating an AI system that merely appears to be helpful but actually impairs skill development.
However, if you are building an EdTech platform rather than an individual app, your AI success hinges on clean integration and uniform event data. LTI 1.3 allows tools to be securely integrated into LMS platforms. Caliper defines a structured vocabulary for learning event data. Otherwise, your analytics and personalization will be partial and skewed towards tooling that logs well.
Expectations have started to be codified by policy bodies: privacy, age appropriateness, bias testing, transparency, and teacher agency. The guidance from UNESCO clearly includes calls for regulation and a human-centered approach. U.S. Department of Education report leans toward common policies and ethical implementation. The EU AI Act contains an explicit list, including several education-related uses of AI that are classified as high-risk.
A practical founder takeaway is straightforward: if your AI touches grading, admission, high-stakes assessment, or student risk classification, you should treat it like a high-risk system even outside of Europe. That adds up to documentation, monitoring, human oversight, and a transparent appeals process.
