Medly AI is one of the more interesting companies in British education technology, and it has nothing to do with medical education, which is precisely why it is worth medical educators' attention. Built by medically trained founders but aimed at GCSE, A-level and IB students, Medly has reportedly raised close to £6 million and supported more than 400,000 UK students, per Times coverage (last checked August 2026), on a proposition that sounds almost boring until you sit with it: an AI tutor that knows the learner's actual exam. The interesting question is not whether Medly could teach medicine. It is what its model tells us about where AI-mediated learning is going, because medical education is converging on the same ideas from the other direction.
Why ChatGPT-style Q&A isn't really tutoring
The first generation of AI study help was answer generation: ask a question, receive a fluent explanation. Useful, and educationally thin, because supplying an answer can bypass the very process it is meant to support; the learner's misconception is never surfaced, their attempt is never examined, and the fluency of the response papers over the gap it was meant to fill. Tutoring, as any human tutor knows, is mostly not answering. It is questioning, marking, diagnosing and deciding what happens next, and the platforms now pulling ahead, in schools and in medicine alike, are the ones that understood this early; our earlier piece on that distinction is at /blog/socratic-ai-tutors-vs-answer-first-chatbots.
What Medly gets right
Medly's design choices read like a checklist of what answer-first chatbots lack. Curriculum awareness: its tutoring is constrained to the relevant exam board and specification, so help is framed in terms of what this learner's assessment actually rewards, not what the internet knows. Marking the learner's attempt: it takes typed or handwritten answers and marks them, which forces the learner to produce before receiving. Loss-point identification: feedback identifies where marks were lost rather than presenting a model answer, which is the difference between a mirror and a lamp. And active involvement throughout: the learner works; the system responds. None of these ideas is novel individually; a good human tutor does all four. What Medly demonstrates is that they now scale, and that families will pay for the scaled version, at hundreds of thousands of learners.
Why medicine makes the same problem harder
Transplanting that model into medicine multiplies every difficulty. The curriculum is enormous, spanning thousands of conditions and a working lifetime rather than a two-year specification. The assessments are plural: UKMLA, PLAB, MRCP, MRCGP and dozens of specialty exams across jurisdictions, each with its own blueprint, so curriculum awareness means exam-by-exam construction rather than one specification per subject. The material is ambiguous: clinical questions have distractors that are nearly right, presentations that overlap, and guidance that changes underneath the learner. And, most fundamentally, knowledge alone is insufficient: the medical learner's errors are frequently reasoning errors, the missed discriminating clue, the differential never considered, which means the marking problem is not where marks were lost but where thinking went wrong.
What the medical equivalent looks like
Solving that harder version is the design brief iatroX set itself, and the components map directly onto Medly's checklist, upgraded for the domain. Exam-specific question banks across 40+ examinations supply the curriculum awareness. Adaptive selection and spaced repetition supply the what-next decision continuously, informed by performance rather than preference. The Socratic Tutor supplies the marking-of-thinking: when an answer is wrong, it does not present the model explanation but interrogates the reasoning, which clue was weighted wrongly, which alternative was never raised, until the actual misconception is named and corrected. The AI Study Planner turns performance into a schedule. And the whole system runs on web and native mobile apps, moving between question practice and tutoring in one flow, because the transition between attempting and being taught is where the learning actually happens.
From answer engine to learning system
The lesson Medly teaches medicine, in one sentence: the winning educational AI will not be the model with the longest answer; it will be the system that best decides what this learner should do next. Answer quality is being commoditised by the frontier models themselves. What is not commoditised, in schools or in medicine, is the loop around the answer, assessment, diagnosis, prescription, retest, and the platforms building that loop, Medly for the exam hall, others for the ward, are describing the same future from different starting points. The wider survey of that shift across medical platforms is at /blog/how-ai-is-changing-the-medical-question-bank.
Frequently asked questions
Should medical students actually use Medly?
For medicine, no, it is built for school curricula, and that constraint is its virtue. Its relevance to medical learners is as a preview: the feedback-first tutoring model it applies to GCSEs is what medical platforms are now racing to build for clinical knowledge.
Is curriculum-constrained AI just a smaller chatbot?
The opposite: the constraint is the product. Knowing what this exam rewards, and marking against it, is precisely what generic models lack, and it is the hard, unglamorous work that separates tutoring systems from chat interfaces.