When integrating AI throughout a company, a common trap is to automate everything you possibly can. This can be a big problem because LLMs give you middle-of-distribution output.
If you prompt well with good context and use a frontier model, the output will be competent and follow best practices. But it will be average.
You won't often get the equivalent of one-in-a-million insights. You won't be "thinking differently". That's not how LLMs are trained.
The problem is, if you're a startup, that's what you need to succeed.
Being average at your core differentiator means you're doomed.
Luckily there are many aspects of your company that are totally fine to be middle of the road. Tedious tasks for example. Or even areas that you don't have expertise in and would otherwise be bad at. Commonplace best practices are fine for these things, and continuous evals ensure things keep working as expected.
Just make sure you nail the key things with expert human judgement. If you can hire great people, automate the repetitive work with AI, and free them up to be making tasteful out-of-distribution decisions all day, that's a winning combination.