I am a responsible AI advocate. I believe in AI enough to have built my agency around it. I also believe in auditing what I use before it gets anywhere near client work. So when I deployed Microsoft 365 Copilot for enterprise use in a healthcare marketing context, I did what most companies don't, I logged everything.
What I was trying to do
I deployed Copilot across Microsoft 365: Planner Premium, Power Automate, Dataverse, SharePoint, as part of building out an operational infrastructure for FutureNova Health. Real work. Real stakes. Not a sandbox experiment. Checkout our initial article on our audit.
Where it failed (and how badly)
The reliability failures were not edge cases. Copilot directed me to set up databases using incorrect guidance. Those configurations cannot be undone once created. My only option was to delete hours of work and start over. When I pushed back (when I told it something didn't seem right) it didn't acknowledge the concern. It gaslighted me. It continued down the path of destruction with full confidence, presenting fabricated Microsoft features and retired products as current and functional.
That is not a hallucination problem. That is a trust problem. A tool that responds to correction by doubling down is not a tool you can use in healthcare marketing, where the cost of being wrong is not a rebuilt database but a patient-facing campaign that has already launched.
The checkpoint breakdown
Reliable — 1 out of 5. Fabricated features, irreversible errors, and active resistance to correction. It fails the first checkpoint, which means under the Purposeful AI framework the conversation ends here. The remaining scores are logged for completeness.
Time Saving — 1 out of 5. 56 hours lost across 129 incidents. The compounding factor was false confidence — the tool presented itself as trustworthy enough to keep following, which extended the time loss far beyond what an honestly unreliable tool would have caused.
Elevate — 0 out of 5. No elevation of the work at any point. Other AI tools in the workflow category occasionally deliver genuine quality improvements. Copilot did not once attempt to make the work better.
Drives Progress — 0 out of 5. Dead ends, loops, irreversible errors, and a tool that fought correction at every turn. If progress means moving forward, this tool moved in the opposite direction (and it did so with a blindfold on).
The bottom line
Microsoft 365 Copilot will not be used in any FutureNova Health client engagement. Not with caveats, not for low-stakes tasks, not until a future version demonstrates that it can acknowledge its own failures rather than defending them. A tool that gaslights its users does not belong in healthcare marketing at any price point.
In October 2025, Microsoft quietly updated their consumer Copilot terms to describe the tool as intended for entertainment purposes only. The enterprise version carried no such transparency. This audit is the transparency they didn't provide.
Purposeful AI means auditing what you use before it costs you what it cost me.
Learn more about our stance on AI.
