AI Demand Planning Software: Built-In Add-Ons vs a ChatGPT Spreadsheet

For most small product businesses a tested spreadsheet forecast is enough to start with, and AI demand planning software is worth buying only if it beats that spreadsheet on your own last year's sales. Run the same backtest on both, compare the errors, and let the numbers decide.

The trap is buying on a demo. Every forecasting tool looks clever on a demo catalogue. The question that matters is whether it would have predicted what you actually sold, and you can answer it for free with a year of sales data and an afternoon.

Follow me on Instagram@sagnikteaches

What the add-ons are, and what they do not say

Several inventory and manufacturing systems now add forecasting to the system you already use for stock. Two that small product businesses meet are worth naming, with the caveat that neither page gave a price for its AI feature in November 2026.

Connect on LinkedInSagnik Bhattacharya

• Katana AI Replenishment. Katana's feature page describes a 12-month forecast built from more than a year of sales history, stock, open orders, lead times and costs, turned into a draft purchase or manufacturing order you approve, with days of stock remaining for each scenario. Its help centre calls the feature a beta and suggests at least six months of sales history. Neither page said which plans include it or what it costs.
• Cin7 ForesightAI. Cin7's pricing page marks ForesightAI Forecasting as an add-on on every plan, with no price shown. Its product page says it forecasts up to 24 months ahead, suggests purchase orders and safety stock, works inside Cin7 Core and Omni, and is priced on gross merchandise value. Cin7's own plans start at $349 a month.

Treat both as unverified until you have the price, the plan requirement and a trial in writing. No accuracy figure for either product appears here, because neither publishes one that I could check, and you should be wary of anyone who quotes a percentage improvement without showing the dataset.

Subscribe on YouTube@sagnikteaches

Built-in add-ons have a real advantage: the forecast sits next to the stock, the supplier and the purchase order, so acting on it is one click. A spreadsheet forecast has to be copied into the system by hand. That convenience is worth money, but only if the forecast is good.

The decision: five questions in order

1. Do you have 12 to 24 months of clean sales by product and month? If not, neither option will work well. Fix the data first.
2. Are your stock-outs or overstocks costing real money? If you can count the lost sales or the cash tied up, forecasting has a payback. If you cannot, it may not.
3. Does your current inventory system already have a forecasting add-on? If so, ask for a trial on your history before paying, as set out below.
4. Does the spreadsheet forecast beat a naive guess? If it cannot beat "same as last year", no software will help much with that product.
5. Does software beat the spreadsheet by more than it costs? Only the backtest answers this.

Questions 1 and 2 cost nothing. Question 4 costs an afternoon. Only then do you spend on question 5.

Main worked example: one product, two years, three methods

Take a small business selling garden watering equipment, with one fast-moving product whose monthly sales are strongly seasonal. These figures are invented for illustration, but the pattern (a summer peak and a quiet winter) is typical. Last year's units, then the year being forecast, month by month from January:

Method A copies last year. Method B averages the previous three months, a common default that ignores season. Method C takes last year's month and adds 10 percent, assuming a growth trend you noticed between earlier years. To score each, add up how far each month's forecast was from what happened (ignoring plus or minus) and divide by total actual sales. This measure is often called weighted absolute percentage error, and the lower the better.

Three lessons come straight from the table. A moving average that ignores season is the worst of the three for a seasonal product, which is why a naive tool can disappoint. A simple seasonal method with a growth factor does well, at 4.2 percent. And the growth factor was an assumption; if the true growth had been 3 percent, method C would have overstated demand. A built-in tool earns its fee if it finds the growth and season on its own across hundreds of products, not just for this one.

Now turn error into money. Suppose each unit costs $18 and sells at $35, and an unsold unit costs about $2 a month to hold. The spreadsheet's error of 96 units is about $1,700 of stock mis-bought at cost over the year. If software cut that error by a third, about $575 less stock would be mis-bought each year for this one product, and only part of that is real money: the holding cost on surplus units and the margin lost on missing ones. Multiply by your number of products, and then compare with the add-on's price, which you must obtain.

Build the spreadsheet forecast in an hour

The spreadsheet route needs nothing exotic. You need one sheet with a row per product per month, and three or four formulas. An AI assistant helps by writing the formulas and explaining them; it should not be asked to invent numbers.

Step 1: export and tidy the data

Export orders for the last 24 months by line item. Make a table with product code, month and units sold. Remove returns, test orders and any one-off bulk sale you would not expect again, and mark months when you were out of stock, because zero sales then do not mean zero demand. For tidying help, see SKU clean-up with AI and the older clean messy data tutorial.

Step 2: ask the assistant for the formulas

I have a sheet with columns: Product, Month (first of month), Units.
24 months of data, no customer information.
Build a forecast for the next 12 months for each product using:
(a) same month last year,
(b) same month last year multiplied by the ratio of the last 12 months
    to the 12 months before that,
and a column for the ratio. Show each formula in plain words and
write them for Excel. Do not invent any numbers.

A typical response gives you the ratio formula, for example the sum of the latest twelve months divided by the sum of the previous twelve, then multiplies each month's same-month-last-year figure by it. Check two or three cells by hand against a calculator. If the assistant gives something you cannot follow, ask it to explain each step or to use a simpler method.

Step 3: backtest on last year

Hide the most recent twelve months, forecast them from the twelve before, and compare. That is exactly what the table above did. Repeat for your ten biggest products, and write the error beside each. If you can, add the extra check of how often the forecast was lower than the real sales, because under-forecasting causes stock-outs while over-forecasting ties up cash.

Step 4: add a safety margin

For a product with a forecast of 200 units and a typical error of 10 percent, order or hold 20 extra units for peaks if a stock-out costs more than the holding. For slow, expensive items, accept the stock-out. See how to use AI to set reorder points for turning that into a number per product.

Run the same test on a vendor's tool

If your system has a forecasting add-on, you can run the identical test. The aim is to find out what the tool would have said last year, given only the data it had then.

1. Ask the vendor for a trial or a sandbox, and say you will test it on your own history.
2. Supply the data up to a cut-off date twelve months ago, or ask them to run it with that cut-off.
3. Take the forecast it produces for the next twelve months and compare with what actually sold.
4. Score it with the same error measure as your spreadsheet, on the same products.
5. Ask how it treats stock-outs, promotions, new products and supplier lead times. Record the answers.

If the vendor cannot run on a past cut-off, ask for a live trial of one quarter, and compare forecasts against sales as they happen. A vendor who will not let you test is asking for faith. Also ask about lead times: a forecast of what customers will buy is only half the problem if your supplier takes twelve weeks to deliver.

Handling promotions and stock-outs in the sheet

Two kinds of month break a simple forecast, and fixing them is where a spreadsheet user adds the most value. Here is the garden watering product again, now with two flags in year 1: a half-price sale in May and a stock-out in July when the warehouse ran dry on the 22nd.

The July figure rises because the recorded 240 understates what customers wanted. If you forecast next July from 240, you under-order again and the stock-out repeats. The adjustments are estimates, and you should write the reasoning beside each so that a colleague can challenge it. A forecasting tool that does not ask about these flags may simply learn the distortions, so ask any vendor how it treats promotion months and out-of-stock days.

Sample assistant output, and what to fix in it

Asked for formulas, a chat assistant might reply along these lines. This is an illustrative reply, not a quote from any real session.

Column E (Growth ratio): =SUM(C14:C25)/SUM(C2:C13)
Column F (Forecast): =INDEX(C2:C13, MATCH(MONTH(B26), MONTH(B2:B13), 0)) * E26
Note: use Ctrl+Shift+Enter if your version does not support dynamic arrays.

What to check before trusting it. First, the ratio compares the latest twelve months with the twelve before; that is right only if rows 14 to 25 are the newest year. Second, the lookup matches month number across a 12-row range, which breaks if the range holds more than one year of the same month, or if the data is sorted differently from the assistant's assumption. Third, one ratio for the whole business hides products that are growing at different rates. Ask it for a ratio per product, or at least per product family, and test on the ten biggest products only. Finally, ask it to explain what happens when a month has zero sales and the growth ratio divides by zero; a sound formula needs an IF test for that.

From forecast to purchase order: the lead-time sum

A forecast only helps if it arrives in time to act. Say the same business buys from an overseas supplier with a ten-week lead time and orders in boxes of 50. June to August is the peak, with forecast demand of 231, 264 and 253 units in method C.

1. The order for June stock must land by the end of May, so it is placed in mid-March, ten weeks earlier.
2. Forecast demand for June to August is 748 units. Stock on hand and on order in mid-March is, say, 180.
3. Shortfall is 568 units. Add a margin of 10 percent for forecast error, 75 units, which makes 643, rounded up to 650 for boxes of 50.
4. Cash needed at $18 a unit is $11,700, which is also what you would be sitting on if the summer is poor.

Compare that with an error of 24 percent from method B, which would imply a margin of about 180 units, a further $1,900 or so of cash tied up (105 extra units at $18). This is the money argument for better forecasting: a tighter error lets you hold a smaller margin. It also shows why an add-on that drafts purchase orders with lead times is attractive, since step 1 is where manual planning most often slips.

Presenting the result to a business partner or bank

If a forecast supports a purchasing decision with real cash, say it plainly. A short note makes the reasoning auditable and keeps the forecast honest.

Forecast note, product group: garden watering
Method: same month last year x growth ratio (last 12 vs previous 12 months)
Backtest: forecast of the last 12 months from the 12 before, error 4.2%
Adjustments: May promotion removed (-35), July stock-out added (+60, cautious)
Assumptions: no new competitor, supplier lead time 10 weeks
Risk: a cool summer; margin of 10 percent added to the order
Owner of the forecast: [name], refreshed monthly

Nothing in that note needs software. It does need honesty about the assumptions, and it makes the comparison with a vendor's output straightforward, since you can ask the vendor to explain its adjustments in the same terms.

Where each option wins

Example: a business that should stay on the spreadsheet

A twelve-product bookkeeping-supplies seller with steady sales and one supplier, with a 14 percent monthly error on method A that drops to 8 percent with a growth adjustment, gains little from software. The monthly refresh takes 40 minutes. Paying a subscription would save perhaps 30 minutes a month, and would cost more than that half hour is worth. Stay put.

Example: a business that should test the add-on

A seller of 300 spare parts across several suppliers, each with a different lead time, spends two days a month on reorders and still has recurring stock-outs. The spreadsheet error is above 25 percent on the slower items, and nobody has time to tune 300 rows. Here a trial of the system's AI feature is worth the effort. If it drafts purchase orders using lead times and cuts those two days to half a day, that is 1.5 days a month back, plus fewer lost sales.

How demand planning goes wrong, whichever route

• Stock-outs hide demand. A month at zero because you sold out looks like no demand. Mark those months and treat them as missing, not zero.
• Promotions distort history. A big discount month inflates a forecast for the same month next year. Flag promotion months and adjust.
• One-off orders. A single bulk order from one customer is not a trend. Remove or cap it before forecasting.
• New and renamed products. Merged or renamed codes split the history. Keep a mapping.
• False precision. A forecast of 212.4 units suggests accuracy you do not have. Round, and think in ranges.
• Ignoring lead times. The forecast says June will be busy, but you must order in March. Always work backwards from the supplier's lead time.

For more on how these go wrong in practice, read why AI stock forecasts go wrong. If you want to build the simple seasonal sheet with a chat assistant step by step, ChatGPT demand forecasting from sales history walks through it.

A one-page decision record

Whatever you choose, write it down so the next review starts from facts. An illustrative filled-in version:

Date: [date]
Products tested: top 10 by revenue
Data used: 24 months, returns removed, stock-out months marked
Spreadsheet error (last 12 months): 9% on average, 20% worst product
Vendor tool tested: yes, quarter trial, error 7% average
Cost of tool: [price quoted in writing] a month
Estimated annual value of the difference: [calculation]
Decision: [keep spreadsheet / buy / retest in 6 months]
Review date: [date]

If you would rather work this out with someone, that is what my AI implementation consultation covers, and you can start from the contact page.

Demand planning questions for small product businesses

How much history do I need for a forecast?

At least 12 months of sales by product and month, and ideally 24 so you can see a repeating season and test it. With less than a year, a seasonal method cannot be checked, and any tool, built-in or spreadsheet, is mostly guessing. In that case use a simple moving average and keep safety stock higher.

Do the inventory add-ons publish their prices?

Not on the pages I checked in November 2026. Katana's AI replenishment feature and Cin7's ForesightAI add-on had no price shown, and Katana's page did not say which plans include it. Ask for the price and the plan requirement in writing, and ask whether a trial runs on your own sales history.

Can ChatGPT build the forecast for me?

It can write the spreadsheet formulas, explain results and spot odd months, but a forecast's accuracy comes from your data and the method. Ask it to build the sheet and show its formulas, then check a few cells by hand. Never paste customer details; sales counts by product and month are enough.

What is a good forecast error?

It depends on how steady your sales are. Fast-moving staples can reach single-digit error on a monthly basis, while lumpy or new products are far worse. Judge by comparison, not by a target: the software is worth paying for only if its error on your own past data is meaningfully lower than the spreadsheet's.

Further reads

• How Small Shops Can Use AI to Forecast Stock and Reorder — A shop-sized version of the same job.
• How Bakeries Can Use AI to Predict Demand and Cut Unsold Stock — Forecasting for short-life stock.
• Boutique Inventory Management: AI for Size Runs and Reorders — Forecasting when sizes and colours split demand.
• Cloud Kitchen Inventory: Forecast Prep by Delivery Platform With AI — Daily demand forecasting for a food business.
• Tyre Stock Excel Sheet: Track Sizes and Reorder Points With AI — A stock sheet with reorder points, step by step.
• Ice Cream Inventory: Forecast Tubs by Flavour, Weather and Weekday — Weather-driven demand planning.
• AI Inventory Forecasting for Small Businesses: How It Works — How inventory forecasting works, in plain terms.
• AI Dropshipping Product Finder: Check Demand and Supplier Risk First — A product shortlist is only a start; use this evidence sheet to reject weak demand, thin margin and unreliable suppliers.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Katana AI Replenishment page and help centre article, Cin7 pricing page and ForesightAI page; checked November 2026. Neither vendor listed a price for its AI feature.