I've been working with AI agent training where you connect LLMs with data, search engines, etc. I get some results that I didn't want but I may need to learn how best to ask for the data. Just as I had to learn 10-15 years ago on how best to search in Google.
That aside, I'm very impressed with how well it can consume data and then provide answers. Or how a AI mgr agent can be given a team of AI agents, and it decides what resources it needs to utilize in order to complete a task. Automated workflows have been around for a while but you have to hardcode the logic. This uses it's own decision making just as a person would do.
Downsides, complex processing isn't fast such as a Google search but it's still night and day compared to people processing. If you asked Marketing AI agents to research, propose 3 marketing campaigns ideas, decide on most effective campaign, and then build out a detailed campaign for the winner, it might take 4-5 minutes and cost upwards of a $1 of LLMs processing. Obviously not as quick and free as a google search but imagine how long that would take with people and how much you pay in salary.