Real-Time AI Insights: The initial plan was to provide AI insights in near real-time, allowing the tool to function like an active participant in the meeting, notifying users when critical insights or actions were available. However, this approach proved to be prohibitively expensive due to high computational costs.

Periodic AI Insights: Switching to periodic AI insights (delivered at regular intervals during the meeting) seemed like a cost-effective alternative. However, during testing in real-world meetings lasting over two hours, this approach was found to be less effective. Users often found the insights either too delayed to be actionable or disruptive to the meeting flow.

On-Demand AI Insights: Finally, the solution was to provide AI insights on-demand, triggered by user action. This approach proved to be the most efficient and flexible, as it allowed users to fetch insights whenever needed—whether mid-meeting, at the end, or during a follow-up. With a processing time of under 30 seconds, this solution struck the right balance between speed, usability, and cost-effectiveness.