I think there's a space for something that wraps an llm (especially multimodal) to do something that's halfway to agentic. Yes you could do it yourself but it's not worth it to you to figure out prompts etc, especially when someone has already optimized it. Plus, it could go from 100 clicks and 10 minutes in front of chatgpt to zero clicks, automated ingest, and get an email when the results is baked.

A good example I saw recently was stripping ads from podcasts.