Just finished one of my favorite LLM experiments! I wanted to test out how LLMs could do with auditing expense report line items, and I got so much more out of than I anticipated. Finding at least 4 use cases for LLMs.

Idk genie 3 might be the biggest ai announcement this week even with gpt5.

Clear the schedules gpt5 tomorrow.

Genie 3 wow I’m going to be one shot in the future.

Bob Ross but creating benchmarks/tests for ai and finding out new use cases.

Opus 4.1 like opus 4.0 didn't do great with this auditing expense task. nice report though.

ChatGPT you are more limited to the built in tools but with Claude Code you are only limited by your tech knowledge. Both require an imagination though to think outside the box. New tools for new solutions! very exciting.

still lots of potential here for workflows. One use case I'm thinking about is with retrieving data from an ERP system, with ChatGPT agent you would work with sharepoint/gdrive. With Claude Code I think you would be able to retrieve the data by having it run an existing script.

Because of the local setup (I'm sure with MCPs or w/e else I don't know about can make it work with cloud files like sharepoint/google drive), this tool feels like it would empower the individual more than a team. ChatGPT Agent feels way more user friendly to set up for a team.

It's not clear to me how Claude Code "reads" pdfs, but with Chatgpt Agent it will either try to use a python script to parse pdf info or it will open it up in an image. The image part is really helpful with those pesky pdfs that don't play nice.

Claude Code saved python scripts to my folder that it used to accomplish tasks while ChatGPT Agent will run code in it's virtual environment. The "out of the box" product feels like Claude Code is more for local files and ChatGPT Agent for cloud files.

I ran it through both tests (the 15 docs correct combo and the 21 docs with variances) and it found the correct errors. Claude Code was definitely more fun to watch work but it also required me to approve a couple things. still it was relatively quick for this easy task

lately I've been seeing a lot of use cases similar to this with Claude Code! So I decided to give it a try with the same tests I did with ChatGPT Agent! It's cool to see the differences between the two. Claude Code did just as well as ChatGPT Agent! 1/x

I’ve found codex to be very useful in some coding projects. For me having to go and confirm a pr was a bit tedious especially for smaller tweaks and changes (and had to learn what GitHub was lol). Something that can go start to finish sounds really good.

OpenAI really nails down that easy to use product though.

Still think there’s that initial hurdle that non tech people need to get over to access Claude code but if they do it’s a pretty dang powerful tool. I can see why I’m seeing so many tech entrepreneurs post about using it to do some accounting like work for themselves.

Was testing Claude code out with some of things I’ve been demoing ChatGPT agent with and it was really fun to see it work lol. Seems like Claude code for local files (without the additional setup) and ChatGPT agent for cloud files.

With the ChatGPT interface, it is kind of weird they don’t date and timestamp messages for scheduled tasks. Guess it’s just a needed thing in the prompt to first respond with the date and time?.

RT @TheRealTechCPA: A big distinction I can imagine playing out in the near-term in Accounting but true for other similar “white collar” wo….

Might be a very short term problem, but it is still an initial hurdle for this mass implementation in work. Future workflows you probably just have other AI workflows to audit the previous work after figuring out where the human reviewal needs to be.

For now, saving files to folders automatically (claude code) can cause some laziness where a user won't double check the files. I think that's where chatgpt agent providing the files in a chat is good. extra step of downloading and saving to file storage allows the user to audit.