I would build around a problem where the AI can actually take action, not just give recommendations. There are plenty of tools that can summarize, write, or answer questions now. The more interesting opportunity is taking a messy workflow and turning it into something that gets completed with minimal human involvement. I would probably start with one very specific business problem, make the AI extremely reliable at solving it, and expand from there.