Calibrating Copilot Ep. 5: Which Copilot Builds a PowerPoint You Would Actually Present?
Almost every knowledge worker has to create PowerPoint presentations, which made this one of our most practical episodes yet. For Episode 5 of Calibrating Copilot, Dino Caputo and I asked several versions of Copilot to create a polished, customer-facing deck from an existing website.
The test
We used our own EnableUC website as the source and kept the prompt deliberately short: using enableuc.com, create a customer-facing PowerPoint presentation of no more than 12 slides that explains all of our services, uses the branding, visual style, colors, imagery, and logo from the website, and includes a talk track in the slide notes for a 15-minute presentation.
Out of curiosity, I asked ChatGPT to improve the prompt. It produced a 464-word monster. In my one test with it, the longer prompt did not make a meaningful difference.
We tested Copilot Chat, Microsoft 365 Copilot, Copilot inside PowerPoint (using both Auto and the newly available Grok, or SpaceX, models), and Cowork at different effort levels and models. Note that the Grok models require tenant admin consent and, for now, are only available inside the Word, PowerPoint, and Excel apps, not in Copilot Chat or Cowork.
Can you tell which deck cost the most?
When we lined up title slides side by side, it was surprisingly difficult to tell which came from the free Copilot Chat and which consumed AI credits. Copilot Chat produced a genuinely professional-looking deck in about four minutes, at no cost.
Microsoft 365 Copilot was the biggest disappointment. Its deck had 11-point fonts, excessive white space, and generic stock-style images. We both agreed it was the weakest of the group, and we would not present it to a customer.
Same model, same prompt, very different results
Dino and I both ran Copilot in PowerPoint with the Grok model, and the results were dramatically different. Dino's deck alternated layouts, included speaker notes automatically, and ended with a strong call to action. Mine was darker, more uniform, needed a second prompt to add speaker notes (about 20 minutes in total), and closed with a lame 'Let's talk.' Since Copilot in PowerPoint does not consume AI credits, running it two or three times and picking the best result is a reasonable strategy.
Cowork showed the same variability. Cowork Medium on Auto used 345 credits and produced a deck with great visual variety. Cowork Medium with a specific model explicitly selected used more than three times the credits and produced something bland and repetitive. Cowork Max, using an Opus model, took 53 minutes and about $17 in credits, and produced the most sophisticated layouts, including a Venn diagram with differentiated text.
Branding, logos, and images
Only Microsoft 365 Copilot, Cowork, and Cowork Max used our official logo. The others invented one, capitalizing the lowercase e in enableUC or adding a space.
Cowork was the only option that used real imagery from the website, including a photo of me. It reported that it could not find a photo of Dino, even though one is on our About page.
The Grok deck had one instance of overlapping text.
Look closely at AI-generated images. One showed a presenter who was three times the size of his audience; another had a meeting participant staring at a blank laptop screen.
None of the decks used a Slide Master. That means changing a logo or background requires editing every slide, and some bullets were built as separate text boxes rather than bulleted lists. AI-generated decks get you started quickly, but they can be harder to edit than a deck built by someone who knows PowerPoint well.
Speaker notes
All of the speaker notes were generally clear and concise. The Grok notes were overly salesy, promising 30% savings and implying we are available 24/7 (Dino does like to sleep). Cowork's notes were cluttered with source references that would need to be cleaned up. Still, any of these talk tracks is better than winging it.
The scores
Note that accuracy here does not mean hallucination; every option accurately reflected the website. It reflects how much rework you would need. Copilot Chat scored two for accuracy and three for completeness, impressive for a free tool. Microsoft 365 Copilot scored one. Copilot in PowerPoint with Grok scored three. Cowork Max (Opus) was our winner, with Cowork Medium on Auto a very close second. Even at 53 minutes and $17, Cowork Max produced in under an hour what would take a skilled person a day or two, including the talk track.
A few more lessons
If the free Copilot Chat is all you have, absolutely use it. It is a great way to get started.
If you have a Microsoft 365 Copilot license, start in Copilot in PowerPoint rather than the M365 Copilot app.
Run the same prompt more than once; results vary enormously.
Even if you build the deck yourself, ask AI to write the talk track and tell it how much time you have.
Copilot should give you a receipt after every task: which models it used, how long it took, and how many AI credits it consumed. Today, /cost shows credits, but Cowork will not tell you which model it used.
AI reads your website differently than people do. Copilot spotted that one page called Dino an 11-time MVP while another said 13, and it found orphaned pages describing older services. We have since fixed both.
Give each person the Copilot that fits
This episode found that the free option was surprisingly capable, the paid Microsoft 365 Copilot app was the weakest for this task, and the most expensive option was not automatically the best. Knowing which Copilot fits which work is exactly how organizations avoid paying for capability that goes unused.
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.
