What I Learned From Automating Work with GPT-4o for Two Weeks
At work, I ran a project to cut down repetitive tasks using the GPT-4o API. To cut to the chase, we saved about 70% of the time, but it also created plenty of new issues, so I'll share a balanced take.
1. Reporting automation: I built a pipeline that reads Excel data and generates summary slides. With a well-crafted prompt, it definitely helps with drafting. However, if a number gets entered incorrectly, it sometimes just goes along with it silently, so human review is absolutely necessary.
2. Chatbot operations: The customer-facing chatbot took a fair amount of initial investment, but surprisingly, it runs well. The false positive rate is around 10%, so a human just needs to reclassify tickets.
3. Coding support: My teammates use GitHub Copilot, and you still can't blindly trust the code it suggests. That said, it's convenient for running unit tests.
My takeaway: Back-office work definitely decreases, but new tasks like quality control and hallucination filtering appear in its place. It feels less like work disappearing and more like work shifting. Still, things like checking repetitive notifications are now delegated to AI, and I just handle the exceptions.
Hope this helps anyone considering adoption.
10 answers
Agreed, 70% savings is amazing.
Do you have a source? Did you actually run it?
Oh, I didn't know this lol
Well, I think verifying the numbers is the most important thing, isn't it?
I tried something similar too, and it really saves time.
I tried something similar at my company, but in the end, human review was still necessary, so it actually made more work. Especially if numbers go into Excel wrong, it's a big deal—and checking all of that takes so much time... T_T
It's a good idea to add a number verification step to the prompt. If you ask the LLM to 'cross-check these numbers against the original,' it catches errors. Also, fixing a template when creating slides makes it much more reliable.