Should AI automate PR reviews or enhance them?

I've recently watched some webinars and demos from the GitHub Copilot team, and I noticed a focus on automating PR (pull request) reviews, especially for the enterprise version. It makes me wonder: if PR reviews can be fully automated, why maintain a manual code review policy at all? Wouldn't it make more sense to integrate these checks earlier in the development process, perhaps directly in the IDE with plugins, linters, or similar tools?

In my view, PR reviews are valuable not just for catching errors but for getting a second opinion, having discussions, and sharing knowledge. Things that are difficult to automate.

The real value of PR reviews is the conversation: second opinions, design discussion, and knowledge sharing. That is much harder to automate than finding bugs.

However, the idea of using Copilot in the same window during a review to chat about PR changes and flag areas of code that need attention does seem interesting.

GitHub Copilot offers a lot of great features, but this focus on PR automation feels a bit strange to me. What are your thoughts? Do you encounter a lot of problems with PR reviews in your company?