The New Surveys Agent Is Here — But Can You Trust Its Questions?

The new Surveys Agent has arrived in the Frontier Program for Microsoft 365 Copilot licensed users to try out, and it's expected to roll out to General Availability by mid-September.

The UX that the team making this has created is really nice. When you ask the agent to create a survey, it builds one in Forms for you, and then you can open it alongside your Copilot chat just like a Copilot Page.

But be careful to check its output. For example, in the image, I fed it the PowerPoint deck I used for a quick client training on Copilot in Teams and it included a useful "Which topics did you find most useful" question. However, on the left in chat it seemed to understand which topics were actually in the deck, but on the right, once added to Forms they had turned into generic best guesses for this type of training.

There are two core assistive use cases for AI. First, we use AI to bring greater efficiency to things we are doing every single day that are part of our core expertise - like using summarization to act as an aide memoire for meetings for projects we are deeply involved in. Second, we use AI to allow us to do things efficiently that are outside of our day-to-day expertise, where the use of Copilot or another agent avoids us asking for outside help or spending time remembering how to do something we do infrequently.

For many users a tool like a Surveys Agent will fall into category 2. It will speed up the process of creating surveys or allow us to execute on gathering feedback for things that would have previously just sat in the "that's a great idea" list. But this category also comes with, in my view, a lot more danger from mis-grounding as we are less familiar both with the content and with the expectations.

Where is the correct balance between saving time and checking every detail? As the utility of AI tools moves into new places, like what we're seeing with the Surveys Agents, this is a moving target. And gaining that productivity boost without sending out nonsense is perhaps the greatest challenge of the AI age.

First posted on Linkedin on 08/20/2025 -> View Linkedin Post here