Is your accountant actually doing a good job with your books? It’s kinda impossible to know.
But LLM context windows are now large enough that you can feed your entire balance sheet and P&L into the LLM and ask it to point out things that might be errors (or that could be optimized). With the right prompting, it produces results that are… surprisingly good.
So I hacked together a little tool around it, combining some LLM checks with some deterministic checks of a few things that are usually signs of a problem in QuickBooks. I’m pretty happy with it.
My main critique so far is that some of the LLM output feels a little bit like a horoscope: specific enough to feel tailored, but vague enough to be true for anyone. In my testing, it always seems to want you to be more profitable (which, in its defense, is probably a nearly-universally-true statement about businesses.)
On the backend it’s Python and Flask, and it uses a mix of Claude (Anthropic) and OpenAI for different parts of the prompt.
I’ve always been shocked at how manual the job of doing the books is, and I’m very bullish on stuff like this to both (a) improve quality, and (b) actually make the job more enjoyable.
I incidentally happen to be a founder of an accounting firm, but this was a nights-and-weekends hackathon-style project made by me and a person on our marketing team--don’t tell the engineers :)
Trying this out (perhaps obviously) requires a QuickBooks Online account. If you don’t have one, you can watch a short screen recording of it working here: https://youtu.be/Xvqnzr82XVc
Very interested in your feedback. I'll be checking this all day to answer any questions.
But LLM context windows are now large enough that you can feed your entire balance sheet and P&L into the LLM and ask it to point out things that might be errors (or that could be optimized). With the right prompting, it produces results that are… surprisingly good.
So I hacked together a little tool around it, combining some LLM checks with some deterministic checks of a few things that are usually signs of a problem in QuickBooks. I’m pretty happy with it.
My main critique so far is that some of the LLM output feels a little bit like a horoscope: specific enough to feel tailored, but vague enough to be true for anyone. In my testing, it always seems to want you to be more profitable (which, in its defense, is probably a nearly-universally-true statement about businesses.)
On the backend it’s Python and Flask, and it uses a mix of Claude (Anthropic) and OpenAI for different parts of the prompt.
I’ve always been shocked at how manual the job of doing the books is, and I’m very bullish on stuff like this to both (a) improve quality, and (b) actually make the job more enjoyable.
I incidentally happen to be a founder of an accounting firm, but this was a nights-and-weekends hackathon-style project made by me and a person on our marketing team--don’t tell the engineers :)
Trying this out (perhaps obviously) requires a QuickBooks Online account. If you don’t have one, you can watch a short screen recording of it working here: https://youtu.be/Xvqnzr82XVc
Very interested in your feedback. I'll be checking this all day to answer any questions.