The scene: I almost undercharged by 2,000 yuan last month
I had AI calculate three pricing tiers for a client — and it almost cost me 2,000 yuan. I only caught it after running the numbers by hand.
That was the moment I finally got it: when AI does the math, the gap in those numbers might be half a month's grocery budget.
What happened + who's already using it
Last week OpenAI released their new model's score on the IMO (International Mathematical Olympiad) — gold-medal level. That same day, mathematician Terence Tao responded online: he acknowledged AI is great at "hard problems with standard answers," but still a mile off from actual research math.
Translated for the rest of us: AI can now solve exam and calculation problems. But ask it to "come up with ideas" or make open-ended judgment calls, and it'll confidently make stuff up.
For non-research folks like us, though, "exam math" is already impressive. Meaning: if you give AI a clear, well-defined problem (like "calculate the total price given this discount rate and tiered structure"), it basically won't mess up. But ask "what price should my business charge?" and it'll fabricate.
My indie designer friend Ah May started letting AI calculate per-project labor costs last week. She said "saves me 3 hours of bookkeeping a week" — but she also admits the final version she sends to clients, she always runs by hand once more.
What you can try today
Cost: Free to ~50 USD/month. ChatGPT and Claude both have free tiers, or 20 USD/月的 paid versions.
Time: 30 minutes. List the number-justification work you've been stuck on (quotes, profit, conversion rates), pick the simplest one, and toss it to AI.
Tech barrier: Copy-paste + calculator skills. You don't need any technical jargon.
First step: Open ChatGPT or Claude and type: "I have 3 quote tiers, costs are 1000, 2000, 3000, and I want to mark them up 30%, 25%, 20% respectively — make me a table." Then verify by hand before sending it out.
Key habit: Always recheck AI's numbers before they leave your hands. This isn't distrust — it's self-protection.
If you fall into one of these three camps
Just starting out (no clients yet): Don't bother with AI bookkeeping right now. Your data is too sparse — Excel beats it. Just clearly record your 1-2 real income entries.
Got 1-2 clients (monthly revenue 10-30k): Sweet spot for AI bookkeeping. Volume isn't huge, but every number matters to what you actually take home. Outsource the "calculate" part to AI, focus on client conversations. Reconcile weekly: AI calculates + you review.
Scaling up (team of 3-10): AI bookkeeping hits a new wall — it doesn't know your team's "hidden costs" (like one employee's time allocation across projects). At this point, switch to a dedicated small tool like 金数据 (Jinshuju) + AI assist, or hire a part-time accountant. AI here fits as "reconciliation assistant," not "bookkeeper."
Tao's right: AI is still a mile off from doing research independently. But for side-hustle players like us, that mile is just "it can calculate, but I need to glance over." You can't skip that step — but at least you're not staring at a calculator anymore.