Product
Telling the buy plan what your data doesn't know yet
Your sales history is a complete record of everything that already happened and a total blank on everything you already know is coming.
The three things your sales history can't see
A statistical forecast is an extrapolation of what already happened. That's its strength — it is disciplined about the past in a way memory isn't — and it's also its structural limit: a driver with no historical trace can't be inside it. A planned promotion, a marketing push that just went live, a supply date that moved: none of them has left a mark in the sales record yet, and one of them never will.
This is an information-arrival problem, not a model-quality problem. The information exists; it just exists in your head first, hours or weeks before it exists in your data. Which is why judgmental adjustment is standard practice in business forecasting rather than an admission of failure — Fildes and Goodwin, in their 2007 Foresight review, found promotions and price changes to be the most common reason planners adjust a statistical forecast, precisely because the planner holds something the model cannot have. Large companies build process for routing that information — consensus demand planning, demand sensing — and a Shopify merchant has the same information problem with none of it. It all lands in one person's head, and the only place it can go is the buy plan.
What your history knows, and what only you do.
| What changed | In your sales history? | Who knows it first |
|---|---|---|
| A promotion you agreed to on Friday | No — it hasn't happened yet | You |
| A post that just went live | Not for days — the orders lag the signal | You |
| A supplier date that moved | Never — it's a supply fact, not a sales fact | You |
| Last quarter's steady sell-through | Yes — this is what history is good at | The data |
Why "just override the number" is the wrong fix
The spreadsheet reflex is to type over the quantity. You know the promo is coming, the suggested number looks low, so you replace it and move on.
Two things break. First, the new number carries no reason. A week later — or to whoever opens the file next — it is just a figure someone typed, indistinguishable from a typo. Goodwin found in 2000 that simply requiring a written justification for an adjustment sharply reduced the number of unnecessary adjustments people made; the discipline wasn't in the arithmetic, it was in having to say why. Second, and worse: an override discards the reasoning it replaces. The suggested quantity wasn't a guess, it was the output of constraints — so typing over it silently drops the case-size rounding, the maximum-stock cap, and the units already inbound on an open purchase order. You haven't added your information to the plan. You've replaced a computed number with an uncomputed one.
And adjustments aren't free. Fildes, Goodwin, Lawrence and Nikolopoulos studied judgmental adjustments across four supply-chain companies and tens of thousands of forecast-and-outcome pairs, and the picture was uneven: adjustments helped on average in most of the companies, but larger adjustments tended to improve accuracy while small tinkering often damaged it — and upward adjustments were far less likely to help than downward ones. Optimism is the expensive direction.
The fix isn't to stop adjusting. It's to change what you adjust — from the answer to the assumption underneath it.
State the change in plain English, watch the plan recompute
In the Smart Buying Table, every SKU gets an AI-suggested quantity you can check, adjust, and turn into purchase orders — by hand or by conversation. The conversational path is the one that takes an assumption instead of a number: tell Atlas how reality differs from the data and the whole table recomputes.
There are two shapes to that. Demand-side, you restate how the product is moving — "The mug sells faster than usual." Constraint-side, you restate what you can spend — "I have $1,000 to spend" — and fit-to-budget defers the least-urgent items and shows exactly what got cut. That second one is the thing an override structurally cannot do, because a number typed into one row knows nothing about the other rows competing for the same money. Upstream in Forecasting the same move runs as a what-if: ask "What if sales trend up 20%?" and scenarios apply instantly — the forecast and the reorder needs update in front of you, no spreadsheet copies.
| You say | What recomputes |
|---|---|
| “The mug sells faster than usual.” | The whole table recomputes (Smart Buying Table) |
| “I have $1,000 to spend.” | Fit-to-budget defers the least-urgent items and shows exactly what got cut (Smart Buying Table) |
| “What if sales trend up 20%?” | The forecast and the reorder needs update together, instantly (Forecasting) |
| “Exclude the weeks we were out of stock.” | Stockout days drop out, so the forecast reflects real demand (Forecasting) |
Recomputing beats overwriting because of what's in the calculation. Every suggestion is computed from your sales velocity, lead times, and incoming stock, then rounded to your case sizes and capped at your maximum stock levels. Restate one input and the machinery re-runs with every other constraint still intact. The same logic runs backwards, too: days you were out of stock can be excluded, so the forecast reflects real demand rather than the days you couldn't sell.
Nothing commits until you create the order
Stating an assumption is only reasonable if being wrong about it is cheap. Every change shows a was→now chip with per-row undo, and nothing is committed until you create the order. The chip is the written justification Goodwin's finding points at, except you don't have to write it — the statement you made is the label on the change, and the previous value is still sitting right there beside the new one. So you can say the mug is moving faster, look at what that did to the rest of the table, and take it back if the plan it produces isn't one you'd sign.
The Smart Buying Table page shows two illustrative examples of this. Say the mug sells faster than usual, and a row's quantity moves from 84 to 112 with an Undo beside it. Say you have $1,000 to spend, and the order total moves from $1,276 to $844 with the reason attached — Deferred: Travel Tumbler ($432). Both are illustrative sample figures from the product page, not measured results — but the shape is the point: the old value, the new value, and what changed, all still reversible until you create the order.
When it is, the plan hands off: select the rows you're ready to buy and Atlas splits your selection into one PO per supplier and asks which location receives it — or add items to the Purchase List and order once you've finished reviewing. The thing that reaches your supplier is still a computed quantity. It just got computed from what you know, instead of only from what already happened.