The general theme: users don't want to use BigCorp's chatbot, they want to use ChatGPT.

I have consulted on a number AI Projects. The issue is less whether they're useful, and more, people just want to use ChatGPT, etc. Major AI providers gradually have consumed the use cases once reserved for bespoke apps. In search, consumers wanted to use Google, not your search engine; consumers want to use ChatGPT, not your bot.

Since most companies are bad at building AI/agents (and search), user trust gravitates to a few known brands.

Honestly this has been a pattern for decades in my domain (search) and it seems to extend to the AI space as well.

You can be successful with positioning and expectation setting. Positioning to a niche. Expectation setting that you're not going to drive 10x company value by sprinking AI on it. But you can build something useful for a subset of users. Companies need to start with low ambitions, not big ones.

You can also be successful just using AI for non-chat things. Enriching content or using LLMs in backend processes that aren't user-focused chat.

I've seen AI successful used in big companies for improving matching algorithms and writing custom email alerts (as opposed to a single template).

I think the good uses of AI are places we don't even realize AI is being used.

For some definition of "large" and some definition of "impact" within an Australian context, the law firms(and legal departments) are using GenAI to produce and review legal documents. There are redundancies and less hiring with management apparently asking their staff to "use AI" for more productivity. That means less interns and less juniors and businesses pushing more and more work on experienced staff, that they can retain, and mandate that they use GenAI to do more with less.

Bunch of (big/medium/small) companies I know are using Cursor, Codex, Claude Code (just noticed: 4Cs) to write/read code and deploy it to production.

Isn't this a definition of AI projects in production?

Yes but only with people in the middle. You can't operationalize most of this properly.

So tasks like coding, summarizing, personas for editing, finding excerpts from long pdfs, answering questions based on documents feeding, it's generally useful to some extent.

Yes, many people are in production of ai that can bypass plagiarism and do you essay. I’ll notify when it’s done.