Hands On: Copilot Study and Learn

What Microsoft built

Study and Learn is a first-party agent inside Microsoft 365 Copilot, generally available since mid 2026 and pre-installed for all education licence tiers (A1, A3 and A5) at no extra cost. It sits in the Copilot app's left navigation and in the chat dropdown, on web and Windows desktop. Everything in this paragraph and the next is Microsoft's own description of its product, and it is worth hearing in full before testing it, because the design is thoughtful.

The agent is framed around a principle Microsoft states plainly: the student does the thinking, not the AI. It is built on a four-pillar learning-science foundation: scaffolded guidance that activates prior knowledge and adjusts until a concept clicks; productive struggle, meaning hints before answers and mistakes treated as learning opportunities; practice and active learning through flashcards, quizzes, matching and fill-in-the-blank activities generated from the student's own uploaded materials, with remediation that re-teaches what was missed; and application and transfer, teach-back and new-problem work. For writing tasks, Microsoft's documentation states the agent coaches on knowledge gaps and process and "never writes for the student". It leads with questions rather than dumping information, and calibrates difficulty to the student in front of it.

Read as a piece of educational design, this is the state of the art in commercial tutoring agents, and much of it maps onto the same learning science this workshop covers in the cognitive outsourcing section: retrieval practice, desirable difficulty, feedback specific to the error. The interesting question is not whether the design is good. It is what the design is made of.

For a closer look at the learning science Microsoft cites for those four pillars, and how far it has actually been shown to improve learning, see the companion walkthrough deck The learning science behind Study and Learn.

How CDU accounts are actually configured

Before assuming anything about what students can and cannot do, it helps to know what CDU's own systems team says. In July 2026, Digital and Technology Services answered exactly this question for this workshop, and their response is the primary source for this section.

Copilot Basic, which includes both the general web chat and the Study and Learn agent, is bundled with the Microsoft 365 A5 student licence. Because CDU is registered as a Higher Education tenant, general Copilot chat is enabled by default on student accounts. CDU does not enforce age-group restrictions that would lock chat down, and feature-level enforcement for Copilot in the M365 apps, chat web search and the Study and Learn agent had, at the time of the response, not been fully settled; DTS flagged that as still under investigation. CDU had also recently lifted restrictions on staff accounts to allow Copilot Basic chat, without including student accounts in that change.

Microsoft's own deployment material completes the picture. Study and Learn can be set as the pre-selected agent for Copilot chat, as an IT-controlled option; the same material notes that students can dismiss it. Even at its most enforced, the learning posture is a default, not a wall.

A prompt, not a property

Here is the central argument of this section, and the reason the hands-on exercise below matters. Study and Learn is not a different model from ordinary Copilot. It is an agent: a set of behavioural instructions, the coaching persona, the hints-before-answers rule, the refusal to write the essay, layered on top of the same general-purpose model that answers questions plainly in the next dropdown item. The pedagogy is real, but it lives in the instructions, not in the architecture. Nothing about the underlying system is incapable of just giving the answer; it has been asked nicely not to.

That distinction has practical consequences for teaching staff. A restraint that is behavioural can be stepped around (switch agents), negotiated with (reframe the request until the persona yields), or simply never encountered (a student who opens plain Copilot chat first meets no pedagogy at all). None of this makes the tool bad; used as designed, it is a genuinely good study companion. It means the tool's existence cannot carry institutional weight it was never built to bear. "Students have access to a learning-safe AI" is a statement about a default setting, not about what students experience, and assessment design that assumes the restrained version is the version students used is building on the wrong fact.

The exercise: use it properly, then break it

This is a laptop exercise in two halves, and the first half is not a warm-up; it is half the point. Work in pairs if numbers allow.

First, use it as designed

Sign into the Microsoft 365 Copilot app with your CDU account and open Study and Learn from the left navigation or the chat dropdown. Bring a real task from a unit you teach: paste in a section of lecture notes or a reading and ask the agent to help you understand it, then let it generate flashcards or a quiz from that material and work through a few items, getting something wrong on purpose to see the remediation. If your unit involves writing, ask for help with an essay task and watch how it handles the request. Notice the craft: it asks what you already know, it questions rather than tells, and when you miss a quiz item it re-teaches before letting you retry. Notice also your own reaction; most staff find the experience better than they expected, and that matters for the advice you give students.

Then, try to break the posture

Now spend ten minutes trying to make the agent abandon its pedagogy and just answer. This half is deliberately open-ended, because the finding is whatever happens in the room. Approaches worth trying: ask it directly to stop coaching and give the answer; insist you are a staff member preparing materials, not a student; reframe the study question as a different kind of task entirely; escalate urgency; or hand it the assessment question verbatim and ask for a model answer. Keep notes on what it refuses, what it reframes, and what, eventually, it does.

Then do the one that is not a trick at all: click the agent dropdown, switch from Study and Learn to ordinary Copilot chat, and ask your question again. No persuasion is required to step around a persona that was only ever one option in a menu, and on CDU student accounts that menu is open by default.

One boundary for the exercise: use your own materials or content you are free to share, not student work or anything the Data, Privacy and Security rules exclude. Breaking the tool's pedagogy is the exercise; breaking policy is not.

Group activity: Be a Champion Prompt Critic

Before or after the hands-on work, the room plays Prompt Critics: a presenter-driven game built for this workshop. Eight everyday prompts from around the university, an assignment announcement, a Study and Learn revision quiz, an exam email, a marking shortcut, an extension request, are about to be sent to Copilot, and every one of them has a problem. The room's job is to call it out before the send button gets pressed; the presenter takes answers, fixes the prompt live, and the improvement is visible in the reply that comes back.

The game runs full-screen from a browser, needs no login and no real AI, and ends with a takeaway checklist. It teaches the prompt-writing half of what this section's exercise teaches by demonstration: what you put in is what you get back, and what you should never put in at all.

The second tool: Gemini Notebook

Google's Gemini Notebook, renamed from NotebookLM in July 2026, makes a useful contrast because its restraint is built differently. Where Study and Learn coaches the model's manner, Gemini Notebook restricts the model's world: it works from the specific sources the user uploads, lecture slides, readings, transcripts, and grounds its answers in those documents with citations back to them. That grounding is closer to an architectural constraint than a behavioural one; the tool's answers are tethered to a corpus the student chose, which changes the failure modes. It is less likely to confidently import facts from nowhere, and correspondingly useless when the answer is not in the sources.

For a motivated student, the tool rewards exactly the study behaviours the learning science favours, provided it is driven that way. The productive pattern is to use it against the forgetting curve rather than as a summary machine: close the notes and have it quiz you cold from the uploaded sources, generate flashcards for spaced review, ask it to connect a new concept to something already understood, and have it challenge you to apply the material to a new scenario. The passive pattern, "summarise week three", is also available, produces something that reads like understanding, and builds none; the same fluency-illusion trap the cognitive outsourcing section examines. Its audio overviews, which turn uploaded material into a two-host discussion the student can interrupt with questions, are popular for commute-time review and worth demonstrating if time allows.

The teaching point from putting the two tools side by side: "AI as a learning tool" is not one thing. One product constrains how the model behaves, the other constrains what it knows, and neither constraint is the model itself. In both cases the study value comes from how the student drives, which is a skill staff can teach, not a property staff can assume.

Sources

Microsoft material describes Microsoft's own product and is identified as the vendor's account; the CDU configuration facts come from institutional correspondence.

• Study and Learn Agent overview, Microsoft Support, published 21 July 2026, updated 18 August 2026; confirmed live 20 August 2026. General availability, licence tiers, features and platforms.
• Study and Learn: AI built for your student, Microsoft Education Blog, 13 May 2026. The learning-science framing and the "learner does the thinking" principle.
• Microsoft, Study and Learn Agent general-availability briefing pack (PDF, in the workshop's source folder). The four pillars, activity types, "never writes for the student", and the pre-selected-agent option with student dismissal.
• CDU Digital and Technology Services, response on Copilot configuration for student accounts, July 2026 (institutional correspondence). Licensing, Higher Education tenant default, enforcement status and the staff-account change. DTS flagged feature-level enforcement as still under investigation at the time of the response.
• Google renames NotebookLM to Gemini Notebook, 9to5Google, 16 July 2026.