I’m standing in front of a tactical gear maker’s team at an AI lunch-and-learn. One of the questions they’d sent in was whether AI could help with planning: forecasting, seasonal hiring, whether to make a hire permanent. So I’ve just put their forecasting spreadsheet on the screen.
Not a demo file. The real one. The sheet that decides how many knives they order from overseas, months before anyone buys one, for a company that also sells on Amazon. Their inventory planner built it, kept it alive for years, and sent it to me with a video walking through how it works.
Before the session, I’d handed all of it to AI: the sheet, the video, my notes from talking with the owners. Then I asked it to read everything.
It’s dozens of tabs and over a thousand rows. No person reads all of that at once. The AI did.
Here’s what it found.
Row 2 forecasts negative one knife. It’s been sitting there for months.
Most of their vendors have no lead time entered. The reorder trigger is meant to be buffer plus lead time. For those vendors, it was buffer plus nothing.
The excess inventory isn’t in finished knives. It’s in the parts.
And two years of tabs had been photocopied forward, month after month, each one a slightly different version of the truth.
The room laughs at the negative knife. Then it gets quiet.
I said the thing I most needed them to hear. This isn’t a criticism of the spreadsheet. It’s a genuinely sophisticated model. The business grew beyond what a spreadsheet can handle.
The AI didn’t do their planner’s job. It did the looking.
Back to their question
That brought us back to what they’d asked, whether AI can do the planning. Here’s what I told them. AI can run every model you can describe. It’ll tell you that you need 10,000 knives. Change one assumption and it’ll tell you 100,000, just as fast and just as sure of itself.
Somebody has to be responsible for the final number.
I went into that room thinking the hard part of forecasting was the math. It isn’t. AI will hand you a number in seconds, any number you ask for. What’s scarce is a person who understands where it came from, and is willing to put their name on it.
AI can run the model. A person owns the number.
Build it in the sheet first
That afternoon I sat down with their planner and owners to rebuild the forecast. The rule we landed on is the most useful thing I can give you.
Build the model in a spreadsheet first. Then hand it to AI.
If you can’t understand it in the sheet, AI becomes a black box, and you’ll have no idea where it got the number.
The model doesn’t need to be fancy. Here’s the shape, with made-up numbers.
Start with a baseline: what you’d sell if nothing changed. Say a product sold about 600 units in each of the last three Novembers.
Then add one column for every thing that changes the answer, each with its own effect:
- A deal you’re running that week: plus 150.
- A new variation that will steal some sales: minus 60.
- A price increase, a new ad campaign, a competitor going out of stock: one column each.
Add them up and you have a forecast you can explain line by line. When it’s wrong, and it will be, you can see which column was wrong.
Then the second question, the one most sheets skip: not just how many, but when. If your supplier needs 90 days from purchase order to units checked in at Amazon, your reorder point is 90 days of expected sales plus your buffer. A blank lead time doesn’t mean zero days. It means the sheet can’t warn you.
Once you understand that model, give it to AI. That’s when it gets powerful. It can put confidence ranges on the number: what to order to be 80% sure you won’t stock out, and what it takes to be 98% sure. It can run 20 scenarios before lunch. And because you built the model, you can check its work.
What AI runs, what you own
On Amazon, this split shows up everywhere money gets committed.
Reorder quantities. With the Seller Labs MCP Server, an AI tool like Claude can pull your own sales history and FBA inventory directly, and run the reorder math for every SKU at once. That’s the machine half. The human half is the question no sales history answers: is that spike real, or one big order? A single large order, or one great deal week, lands in a sales-velocity window and makes a trend out of an accident. The AI sees the spike. You know whether it’ll happen again.
Ad budgets. AI can model what happens to profit if you push a campaign harder, or pull back on one that’s barely breaking even. It can’t decide whether you’re spending to launch, to defend a ranking, or to make money on every sale. That’s strategy, and it’s yours.
Your P&L. AI can assemble the numbers from fees, ad spend and costs faster than any bookkeeper. Before that number goes to your partner, your lender or your investors, a person reads it and stands behind it.
The pattern is the same each time. AI does the looking, the math, and the what-ifs. A person makes the call and answers for it.
The same math, pointed at people
The question they actually asked was about labor. It turns out to be the same model.
How many people you need is your demand divided by what one person can handle. Your seasonality is already sitting in years of sales history. If orders double every fourth quarter, the staffing need does too, whether those people prep inventory, answer buyers, or run ads.
The hard part is the same as with knives: trend or spike? If demand is up because of a one-time order, you hire seasonal help. If it’s a trend, you make the hire permanent. AI can lay out both scenarios side by side.
Someone in the room asked what changes with five new salespeople. AI can model it: what they might close, what that does to orders, what that does to staffing. It can’t tell you what they’ll actually close.
And it can’t make the call that comes after. A model has never had to tell someone their season is over.
If you’re an agency
Everything above gets sharper when the account isn’t yours.
When your report says the client should order 3,000 units, or raise the ad budget by a third, the client doesn’t care which AI produced it. They see the account manager’s name.
So the account manager owns the number. Concretely:
- Every forecast and budget you send has a named owner who can explain each line of it.
- Show the model, not just the answer. A client who can see the baseline and the columns can push back on the right one.
- Never forward an AI number you haven’t checked. That’s how a confident mistake becomes your mistake.
The name on the number
That spreadsheet was never the problem. It got that company where it is. It’s just bigger than any pair of eyes now, and AI can carry the looking.
The judgment was always the valuable part. They just never had time to spend on it.
AI will give you a number in seconds. It will never have to explain it to your supplier, your partner, or the person you can’t bring back after the holidays.
Make sure the number has a person’s name on it. Ideally, one who understands how it was built.
