Where general-purpose AI goes wrong on a brand's numbers
- Textbook definitions. Gross sales for net revenue, refunds for returns, a generic contribution margin instead of yours.
- The wrong source. The store platform's cost field instead of the landed-cost file, store refunds instead of the returns system.
- No seasonality. A quarter times four, month over month on a seasonal business, last week compared to the week before a holiday.
- No sense of normal. Every move is news, so the one that matters is buried among the ones that do not.
- Advice nobody asked for. "Consider reviewing your return policy" in place of the cause.
Same question, two answers
Q1 net revenue was $4.1M, which puts you on pace for $16.4M this year.
Multiplied a quarter by four. For a seasonal brand, Q1 is never the run rate.
Trailing twelve months is $21.8M, up 17%. Q1 is typically 19% of your year, so it points to about $22M, in line with plan. Margin is the watch item: product cost from the landed-cost file is up 90 basis points, which platform cost fields don't show.
TTM first, your own seasonality curve, and cost from the file you actually trust.
More on this one: run rate vs trailing twelve months.
What closes the gap
Four things, written down once and used on every answer: your definitions, your sources of truth, your judgment (the thresholds and seasonal patterns your best analyst carries in their head) and your voice. Then the model is pointed at the recurring work, like the weekly recap and the close pack, and checked against the old way until the numbers match.
Writing those down is the first half of every operations audit. You can start on your own with the free definitions worksheet.
Related
Questions
Is this a different AI model?
No. It uses the same leading models, such as Claude. The difference is everything around the model: your definitions, your sources of truth, your judgment calls and your workflows, written down so it uses them every time.
Could our team do this with custom instructions or a project?
Partly, and it is a good start. The hard part is knowing what to write down, connecting the model to the right sources, and checking its answers against the old way until they match. That is the work of the audit.
Is our data used to train the public model?
Not with business accounts on the major providers, which do not train on customer data by default. Everything runs in accounts you own, under your provider's terms.
Where does general-purpose AI do fine on its own?
Drafting, summarizing, brainstorming and one-off questions where the definitions do not matter. The gap opens on recurring numbers your team and your board will act on.