Calibrating Copilot Ep. 7: Creating Interactive Training with Copilot and Cowork
Formal training often means hours of content, but sometimes you just want to reinforce a policy quickly and memorably. For Episode 7 of Calibrating Copilot, Dino Caputo and I asked Copilot and Cowork to turn a policy document into a short, interactive training experience.
The prompt and the policy
Our prompt: using Contoso's attached AI policy, create a five-minute interactive training experience with branching scenarios, feedback, and a final score.
Dino had Copilot find a publicly available AI governance policy, which he anonymized for Contoso (Microsoft's favorite made-up company) and supplemented. Importantly, the policy encourages appropriate AI use rather than simply banning it, and it names specific tools rather than speaking only about Microsoft AI in general.
Microsoft 365 Copilot (Auto), take one: an agent nobody asked for
On my first attempt with M365 Copilot set to Auto, it decided to use Agent Builder and created a 'Contoso AI Policy Coach' agent in about two minutes. That sounded promising until I started the training. It took 1 minute and 8 seconds to present the first question, the experience was plain multiple choice in a chat, and there were only three questions. It also claimed it could adapt to the learner's role, then admitted it had no evidence it actually did. A fail.
Cowork Medium: fast and cheap
Dino's Cowork Medium run took just 1 minute and 25 seconds and $0.72 in credits. It created an attractive dark-themed app with an introduction, a timer, eight scenarios, a running score, feedback that cited the relevant section of the policy, and the option to retake the training. We gave it five stars for accuracy and four for completeness, mainly because others added a bonus print feature.
Microsoft 365 Copilot with Opus 5.5: excellent
I went back to M365 Copilot and explicitly selected Anthropic's Opus 5.5 model. This time it planned a single, self-contained HTML file, reasoned for 2 minutes and 33 seconds, and delivered an app every bit as good as Cowork's, including a print results button that actually worked. Five out of five.
Cowork Max: beautiful, but we could not share it
Cowork Max built a full React app themed around 'a day in the life' of using AI at Contoso, with 11 questions, four branches, and the most descriptive results dashboard of the group. But it published the app as a Power App, and Dino could not share it with me. Troubleshooting the issue in Cowork burned roughly another 700 credits, almost as much as building the app. Creating it took 31 minutes and $7.90; by the end, the total was about 1,490 credits, around $15, plus an hour and a half of our combined time. Five stars for accuracy, three for completeness.
Microsoft 365 Copilot (Auto), take two
Using the exact same prompt and settings as my failed first attempt, M365 Copilot on Auto this time used Python to generate an HTML app in just 39 seconds. The result was nearly identical to the Opus version, which makes me suspect Auto picked that model. Five out of five.
Dino also tried M365 Copilot with the GPT-6.1 Sol model. In about two minutes it produced a strong app with source notes and a download-your-results option (a CSV clearly labeled as not a compliance certificate), although it did not include a timer or running score. Five stars for accuracy, four for completeness.
Key lessons
If you want a standalone HTML app, say so. Leaving it open invites creativity, which may mean an agent or a Power App you did not want.
Build in Cowork, iterate in Copilot. Dino took a Cowork-created app into M365 Copilot and made changes, including colors, without spending any more credits.
If Copilot gives you a link that is not clickable, you are not losing your mind. Tell it, and it will usually provide a working one.
AI is a meticulous reader. It flagged that the policy's effective date still read 'insert date here' and that two different governance contact emails appeared on different pages.
AI is probabilistic. Because M365 Copilot is all-you-can-eat, trying again is free.
Our recommendation: for interactive training like this, you do not need Cowork. Microsoft 365 Copilot produced excellent results at no additional cost, especially when you specify a standalone HTML app or select a specific model. That is the opposite of Episode 6, where only Cowork could create a usable explainer video.
The right Copilot, used the right way
After seven episodes, the pattern is clear: you need the right Copilot, used the right way, to get value. Spending $15 and an hour and a half on one app may seem minor, but multiply that across a thousand employees using Copilot the wrong way and the impact on budget and productivity is significant.
Visit enableuc.com/calibration for a summary of every Calibrating Copilot episode and the lessons we have learned so far.
That is exactly what our Copilot Calibration service is built to solve. Using your existing Microsoft 365 usage data, we deliver person-by-person recommendations: who should get a Microsoft 365 Copilot license, which licenses should be reassigned, where Cowork makes sense, and where targeted training would pay off. It is fixed-price and requires minimal effort from your team.
