The materials are perishable, the syllabus is not
Stop building courses for Microsoft 365 Copilot. Build a small, dull core that teaches judgement on a handful of real tasks, and treat everything screenshot-shaped as stock that goes off. I design learning for a living, and this is the part of my trade the product broke.
A course is a promise. It says that what you learn on Monday will still be true on Friday, that the menu will be where the screenshot says it is, and that the vocabulary in the glossary is the vocabulary in the product. Copilot cannot keep that promise. Nobody at Microsoft is being careless. The release cadence is just faster than the production cadence of good learning material, and no amount of authoring discipline closes that gap.
What survives is the part that was never about the interface. How to tell a good answer from a plausible one. What the tool can and cannot see. When to do the work yourself. Those hold for years. Everything else has a shelf life measured in weeks, and the sooner you price that in, the cheaper your learning gets.
Nothing about this product held still
This is not a feeling. Take the last twelve months of things a training deck would have stated as fact.
The billing vocabulary changed. Copilot Studio moved its unit of consumption from messages to Copilot Credits on 1 September 2025, per Microsoft’s billing documentation. Any slide that says “messages” is now describing a currency that no longer exists.
A named governance control is being withdrawn. Restricted SharePoint Search is retiring, with new enablement blocked from 31 July 2026, and Microsoft points customers to Restricted Content Discovery instead. If your governance module names the old control as the answer, the module is wrong, and it is wrong in the section where being wrong matters most.
Surfaces get deprecated. Copilot chat in model-driven apps is deprecated from January 2026 in non-Dynamics environments, replaced by Microsoft 365 Copilot.
The runtime story got more complicated. Copilot Studio now exposes three harnesses, and they determine how work is billed. Describing Copilot Studio as one undifferentiated thing was fine last year and is inaccurate now.
And the licensing shape moved underneath all of it. Microsoft 365 E7 reached general availability on 1 May 2026 as a superset of E5, and Agent 365 went GA for Commercial on the same date.
Five changes, five sections of a typical curriculum, one year. A foundation course rewritten to cover all of that is not a course any more. It is a maintenance liability with a completion certificate attached.
The prompt library was my worst idea
I built one. Organised by department, reviewed monthly, sitting in a list people could filter, with a short note on each entry explaining what it was for. It was well made. It was also the least useful thing I have shipped in this area, and it failed twice over.
It went stale, which I expected and had a plan for. What I did not expect was the second failure, which was worse. It taught people to copy strings. Someone would find the nearest entry, paste it, get an answer that was shaped roughly right, and stop. The library removed the exact moment I was trying to create, which is the moment where you look at what came back and decide whether it is any good. I had built a machine for skipping the learning.
The monthly maintenance also bought me less than I thought. People trusted the library precisely because it looked maintained, so the entries that had quietly drifted did more damage than no entry at all would have done. I retired it and replaced it with about a page of guidance on how to describe a task, plus worked examples that are explicitly dated and explicitly disposable.
I got something else wrong for longer than I should have: the “what’s new in Copilot” session. Monthly, half an hour, genuinely well attended for the first two. Then it emptied. The diagnosis is obvious in hindsight. New is not a reason to attend anything. Nobody in an operational job has a slot in their week for a product’s release notes. When I stopped running it by release and started running it by job task, with the new capability mentioned only where it changed how that task is done, attendance came back and stayed.
Teach the task, and teach the editing
If you want a defensible place to spend the training budget, the evidence points at specific work rather than at general capability.
Microsoft Research ran a randomised experiment with over 6,000 workers across 56 firms over six months, published in April 2025. It found around 30 minutes less time spent reading email per week and documents completed 12% faster. It is Microsoft’s own research, so treat the numbers with the caution that deserves. The useful part is where the effects showed up: email triage and drafting. Not everywhere. Two specific pieces of work.
The Australian government’s whole-of-government trial, run January to June 2024 across more than 5,765 licences and evaluated by Nous Group for the Digital Transformation Agency with the Australian Centre of Evaluation, points the same way and is more honest about its own limits. Sixty-nine per cent of participants said it improved task completion speed and 61% said it improved work quality. The evaluators then say plainly that the figures are “approximations and likely quote the upper bound”, that “editing was almost always needed”, that up to 7% said Copilot added time because of the verification it required, and that quality gains were “more subdued than improvements to work efficiency”. The trial is also 2024 data, which is two product generations old as I write this.
Read that as a curriculum, because it is one. Editing was almost always needed. A minority of people ended up slower because checking cost more than drafting saved. So the skill that carries the value is not producing the output, it is the twenty seconds after the output arrives. Almost nobody teaches that twenty seconds. It gets one bullet on a slide about responsible use, and then everyone goes back to demonstrating features.
I now build the editing in as the exercise itself. Give people a draft the tool produced from a real prompt against real content, with a real flaw in it, and have them find the flaw and fix it. It is unglamorous and it works, because it is the actual job.
The training is not the constraint you think it is
Here is the finding that should reorganise a learning plan, and it comes from a survey rather than a measurement, so label it as such. Microsoft’s 2026 Work Trend Index, a first-party survey of 20,000 knowledge workers conducted by Edelman Data x Intelligence, reports that organisational factors account for more than twice the AI impact of individual factors, and that only 19% of AI users are in what it calls the Frontier, where organisational capability and individual readiness reinforce each other.
Take that seriously and a large training programme aimed at individuals is aimed at the smaller lever. It does not mean training is worthless. It means training cannot compensate for work that has not been redesigned, permissions that were never cleaned up, or a manager who has not said which parts of the job this is for.
The usage evidence agrees. In the randomised study, roughly 40% of workers with access used Copilot regularly. In the Australian trial, only one in three used it daily. Those are not knowledge gaps. You do not fix a two-thirds non-daily-use rate with a better module on slash commands.
There is a compliance thread here too, and it is smaller than vendors suggest. Article 4 of the EU AI Act has required organisations to take measures on AI literacy since 2 February 2025. The Digital Omnibus, in force since 27 July 2026, softened the wording from taking measures to ensure a sufficient level of AI literacy to taking measures to support its development, and added text confirming it does not require anyone to guarantee a specific level. It still applies, and it reaches organisations outside the Union where the output of the system is used in the Union. Softened, not removed. It is a reason to keep records of what you taught and to whom. It is not a reason to buy a course library.
What I do instead
This is the whole shape of it, and it fits on one page.
- A short core that does not mention the interface. Judging output, knowing what the tool can see, knowing when not to use it. Rewritten yearly at most.
- Task clinics, not feature tours. One real task, one team, their own content, forty-five minutes. The unit of learning is a task with an artefact at the end, not a capability.
- Dated, disposable how-tos. Every screenshot-bearing asset carries the date it was made and an explicit note that it will drift. When it drifts, delete it rather than patch it.
- A change note that is written by task. Not “what’s new”, but “this is now done differently, and here is the one line that changed”.
- Editing exercises with real flawed output. The most valuable twenty seconds in the workflow deserve more than a bullet point.
None of that is hard to build. The hard part is telling a sponsor that the deliverable is a page and five clinics rather than a twelve-module curriculum with a completion dashboard, when the dashboard is the thing they were expecting to show their board.
I would rather have that conversation than maintain another library.