I run OTSATE, a small lemongrass sauce brand in Malaysia. I’m the founder, the cook and the person who does the numbers, so everything below is one operator’s experience, not a benchmark.
Most of my admin now runs through Claude Code, an AI agent that works inside a terminal and can read my files and write to my spreadsheets. I’m going to describe one task it does well, one moment it got badly wrong, and the one column it will never fill in.
The task it now owns: Monday numbers
Every Monday I give it four things. Sales last week. What I spent on ingredients and packaging. Cash in the bank today. The decision I’m making this week.
It writes the row into a Google Sheet and works out the rest: revenue per day, food cost as a percentage of sales, and how many weeks of fixed costs the week’s sales would cover. A small script it wrote reads my bank statement PDFs and sorts each transaction into a category. Every category it guesses gets flagged until I confirm it, because it cannot tell a lemongrass supplier from a water bill without being told once.
The honest reason this works: I avoided my own trackers for months. Typing a bad week into a cell felt like a verdict on me. Saying “sales were low” out loud to a terminal and watching the row app still get recorded. That is the whole point.
What I stopped doing is opening the spreadsheet to type. What I do with that time is read it. Ten minutes, every Monday, and I actually look.
Also Read: I used AI to rebuild my websites without knowing how to code
The moment it was confidently wrong
In July I asked it to cost a new product: Kam Heong sauce in 100g retort pouches, five for RM45 (about US$10.80). I gave it the recipe and asked for the contribution margin.
It came back with a clean, well-formatted answer. Contribution of 12 to 19 per cent. Reprice to RM60 to RM65 (about US$14.40 to US$15.60) for five, or don’t launch.
I nearly believed it. The table looked like something a consultant would charge for. Then I read the cost lines.
It had priced every ingredient at what you’d see on a Malaysian supermarket shelf or an online grocer. Dried shrimp at RM90 (about US$21.60) a kilo. I pay RM37.50 (about US$9) at the wet market. Oyster sauce at RM21 (about US$5.04) a kilo. I pay RM4 (about US$0.96) in bulk. Fresh chilli at RM14 (about US$3.36). I pay RM5.63 (about US$1.35). Every single line was high, most by double or more, and the errors compounded through the recipe. It also added a per-pouch fee for outsourced retort processing. I own the retort machine.
With my real prices the contribution was 61 per cent. That’s all. The model wasn’t lying. It never asked where I shop, and I never told it, and it filled the gap with the most visible number on the internet.
In late September it did a smaller version of the same thing. It suggested I fry my own garlic instead of buying it fried, to bring food cost down. On paper, correct. But I don’t draw a salary. My labour appears nowhere in the P&L. Buying fried garlic at RM5 (about US$1.20) per 100g is how I force my own time into the ingredient cost so the pricing stays honest. The tool optimised the number it could see and ignored the one it couldn’t.
What changed about how I check its work
I gave it a standing rule, saved in the memory file it reads at the start of every session: never estimate my costs. Every line in a costing is either a price I supplied or marked OPEN. No recommendation on top of an OPEN line. If the answer depends on something it doesn’t know, it has to say so and ask.
The second rule is on me. I read the inputs before the conclusion now. A margin figure is only as true as the ugliest line underneath it, and the tool will format an ugly line just as nicely as a true one.
Also Read: How AI became my sparring partner
The column it can’t fill
The Monday sheet has a column called “the decision this week”. It is always blank until I write in it.
In August I cancelled a December launch for a nyonya achar I had been testing in five-kilo batches. The costing was fine. The product was good. I cancelled it because every jar needed days of sun and hours of my hands, and when I pictured making a thousand of them I did not want that life. No spreadsheet carries that information.
If I let the tool decide, it would pick whichever option lines it had. The costing story above is exactly why that’s dangerous: the lines it has are not always the real lines. And it has no stake. It will not be the one at the stove in December.
What shifted with people
My husband sits as an informal advisor on the business. The Monday sheet has a column for the one question I’m bringing to him that week. Our conversations got shorter and more specific, because the numbers arrive before the feelings do. The friction is that he can now see a bad week in a cell, same as me. We are both still learning to talk about that.
If you run something small
Three things from doing this.
Give the tool your real inputs or it will invent plausible ones. Plausible is worse than blank.
Let it do the typing and keep the reading for yourself. The reading is where the business is.
And write down the decisions it is not allowed to make. Mine fit in one column. Yours probably do too.
—
Editor’s note: e27 aims to foster thought leadership by publishing views from the community. You can also share your perspective by submitting an article, video, podcast, or infographic.
The views expressed in this article are those of the author and do not necessarily reflect the official policy or position of e27.
Join us on WhatsApp, Instagram, Facebook, X, and LinkedIn to stay connected.
The post The AI told me to raise my prices by 40 percent but it had never been to a wet market appeared first on e27.