AI Supply Chain BriefChatGPT for demand planning: 6 prompts and a forecast-accuracy example worked through
Short answerUse ChatGPT around the forecast, not as the forecast: sanity-check a statistical forecast against history, calculate and explain accuracy and bias, draft the assumptions memo and S&OP commentary, and build first-pass forecasts for new items by analogy. Paste exported, anonymized numbers, make it show every formula, and keep the final forecast in your planning system.
Fictional exampleExample Fluid Systems Co., its SKUs and every number on this page are made up for illustration. Don't paste real customer names, programs or confidential demand data into an AI tool; anonymize first and follow the data-safety checklist.
This guide is about the planner's own workflow with a general AI assistant. If you're choosing forecasting software, start with Best AI tools for demand planning (2026). For safety stock, reorder points and excess stock, see ChatGPT prompts for inventory management; this page doesn't repeat them.
What is ChatGPT good and bad at in demand planning?
Good at: explaining a forecast and its error in plain English, structuring an assumptions memo or S&OP commentary, drafting the questions to ask sales, and small calculations when it shows every step.
Bad at: running a statistical forecast across thousands of SKUs, anything that needs live data from your ERP or planning system, and knowing about promotions, customer wins or losses you haven't told it about. Researchers testing ChatGPT on spreadsheet tasks titled their paper "ChatGPT and Excel – trust, but verify", and that is the right rule here too. For what companies get from machine-learning forecasting software, see the sourced examples in the demand forecasting use case.
Prompt 1: Forecast sanity check
Use it before the forecast goes to S&OP, to catch the obvious problems a statistical model can produce.
You are a demand planning analyst. Below are 24 months of history and the next 6 months of forecast for each SKU. For each SKU, flag: (1) forecast months above the highest or below the lowest month in the last 24 months, (2) a step change of more than [X]% between the last actual month and the first forecast month, (3) seasonality that appears in history but not in the forecast, or the reverse. Show the numbers behind each flag. Don't change the forecast; list the questions I should answer before I accept it. [PASTE ANONYMIZED TABLE: SKU, month, actual, forecast]
Prompt 2: Forecast accuracy and bias, worked through
Example Fluid Systems Co. (fictional) reviews three SKUs over three months. Error = forecast − actual, so a positive error means the forecast was too high.
| SKU | Month | Actual | Forecast | Error (F − A) | Abs % error |
|---|---|---|---|---|---|
| A (high volume) | Jul | 1,000 | 950 | −50 | 5.0% |
| A | Aug | 1,100 | 1,000 | −100 | 9.1% |
| A | Sep | 900 | 1,000 | +100 | 11.1% |
| B (mid volume) | Jul | 200 | 240 | +40 | 20.0% |
| B | Aug | 180 | 230 | +50 | 27.8% |
| B | Sep | 220 | 250 | +30 | 13.6% |
| C (slow mover) | Jul | 5 | 10 | +5 | 100% |
| C | Aug | 2 | 10 | +8 | 400% |
| C | Sep | 8 | 10 | +2 | 25% |
MAPE, WAPE and bias in one table
- MAPE (mean absolute percentage error) = the average of the absolute % errors.
- WAPE (weighted absolute percentage error) = Σ |F − A| ÷ Σ actual.
- Bias = Σ (F − A) ÷ Σ actual. Positive means over-forecast.
| Measure | All 3 SKUs | SKU A | SKU B | SKU C |
|---|---|---|---|---|
| MAPE | 68.0% | 8.4% | 20.5% | 175% |
| WAPE | 10.7% (385 ÷ 3,615) | 8.3% | 20.0% | 100% |
| Bias | +2.4% (85 ÷ 3,615) | −1.7% | +20.0% | +100% |
What the three numbers tell you:
- MAPE says the forecast is terrible (68%). WAPE says it's about 11% off. The difference is SKU C: 15 units in total, but one month with a 400% error. Drop SKU C and MAPE falls to 14.4%. Report WAPE for the portfolio and look at slow movers separately.
- Overall bias looks harmless (+2.4%), but SKU B is over-forecast every month (+20%). That's the real action: excess inventory is building on B. Bias by SKU matters more than the total. For what to do with that stock, see the excess and obsolete prompt in the inventory guide.
- SKU A looks fine on every measure, with errors in both directions that roughly cancel out.
Why MAPE can mislead (low volumes and zeros)
Percentage errors divide by the actual, so small actuals produce huge percentages. Forecasting: Principles and Practice by Hyndman and Athanasopoulos notes that percentage-based measures have "the disadvantage of being infinite or undefined if yt = 0 for any t in the period of interest, and having extreme values if any yt is close to zero" (FPP3, chapter 5). If any SKU had a zero-demand month, MAPE couldn't be calculated at all.
The prompt that produces the tables above:
You are a demand planning analyst. For the table below (SKU, month, actual, forecast), calculate per SKU and in total: MAPE, WAPE (Σ|F−A| ÷ Σ actual) and bias (Σ(F−A) ÷ Σ actual), where positive bias means over-forecast. Show the formulas and the sums. Flag any SKU with zero actuals and exclude it from MAPE. Then write 4 plain-English sentences for the S&OP meeting: overall accuracy (use WAPE), which SKU has persistent bias and in which direction, which SKU distorts MAPE and why, and one question for sales about each problem SKU. Don't invent causes. [PASTE ANONYMIZED TABLE: SKU, month, actual, forecast]
Prompt 3: Explain the error to sales and finance
Turn the numbers into something a sales director will read. A shorter Excel version is in Copilot prompts for procurement.
Using the accuracy and bias results below, write a 150-word note for sales and finance leaders who aren't planners. Explain in one sentence each what WAPE and bias mean, say which SKUs were over- or under-forecast and by how much in units, and what that means for inventory and service. End with two specific questions for sales. No jargon beyond WAPE and bias, and don't speculate about causes. [PASTE RESULTS TABLE]
Prompt 4: New product forecast by analogy
You are a demand planner forecasting a new item with no history: [DESCRIBE THE NEW ITEM, PRICE POINT, CHANNEL, LAUNCH DATE]. Here are the first 12 months of sales for [2-3] similar past launches: [PASTE ANONYMIZED TABLE]. For each analog, say how similar it is and why. Build a month-by-month first-year forecast as a range (low, base, high) based on the analogs, showing how you scaled them. List the three assumptions that would change the forecast most, and what data would confirm each.
Prompt 5: Promotion lift questions
Use it before you accept a promotional lift from sales.
Sales expects a [X]% lift on [SKU GROUP] during a [PROMOTION TYPE] from [DATES]. Write the 8 questions a demand planner should ask before accepting this lift: baseline, past results of similar promotions, cannibalization of other SKUs, pull-forward and the dip after, customer inventory, supply constraints, and who owns the risk if the lift doesn't happen. For each question, say what answer would make you raise or lower the lift.
Prompt 6: Forecast assumptions memo
Draft a one-page forecast assumptions memo for this month's S&OP meeting from the notes below. Sections: what changed since last month, key assumptions (market, customers, pricing, promotions, supply), the 3 biggest risks and opportunities with rough unit impact, decisions needed, and owners. Keep it under 350 words. Mark anything in my notes that is unclear or contradictory instead of filling the gap. [PASTE ANONYMIZED NOTES]
How do you check the AI's planning math?
- Recompute one SKU by hand. In the example, check the WAPE sums (385 and 3,615) and one MAPE row.
- Look for zero or tiny actuals before you trust any percentage-based measure.
- Confirm the sign convention. This page uses F − A, so positive bias means over-forecast. Some companies report A − F; match yours.
- Check units. Cases, eaches and pallets mixed in one table will give confident, wrong answers.
More on planning: the inventory optimization use case, the planning topic page and Brief stories tagged #forecasting.
FAQ
How do I use ChatGPT for demand planning?
Use it around your forecast rather than as the forecast: sanity-check statistical forecasts against history, calculate and explain accuracy and bias, draft S&OP commentary and assumption memos, and build first-pass new-item forecasts from analogs. Paste anonymized exports and make it show every formula.
How do you calculate forecast accuracy and bias?
A common pair is WAPE for accuracy, the sum of absolute errors divided by the sum of actuals, and bias, the sum of forecast minus actual divided by the sum of actuals. In the example on this page, WAPE is 385 divided by 3,615, or 10.7%, and bias is 85 divided by 3,615, or +2.4%.
What's the difference between forecast bias and MAPE?
MAPE measures how big the errors are, ignoring direction. Bias measures whether you forecast too high or too low on average. A forecast can have a modest error and still be biased, like SKU B in the example, which was over-forecast every month.
What is the MAPE formula in supply chain?
MAPE is the average of the absolute percentage errors: for each period, take the absolute value of forecast minus actual, divide by the actual, multiply by 100, then average. It breaks when an actual is zero.
Why is MAPE misleading for slow-moving items?
Because it divides by the actual. Small actuals produce huge percentage errors, and a zero actual makes MAPE undefined, as Forecasting: Principles and Practice points out. In the example, one slow mover pushes MAPE to 68% while WAPE is 10.7%.
Is AI actually useful for demand forecasting, or is it hype?
For statistical forecasting across many items, use planning software built for it; the tools guide and the demand forecasting use case cover that. For a planner's analysis and communication, such as explaining error, preparing S&OP and questioning assumptions, a general AI assistant is useful today if you check its math.
