Your Spreadsheet Is a Rear-View Mirror
Inventory is a cash-flow killer. Order too much, and your money is stuck on a shelf, collecting dust. Order too little, and you're telling paying customers to go away. For a small business, the gap between those two mistakes is razor-thin. Most operators rely on a simple spreadsheet with a static reorder point—when stock for TSHIRT-BLUE-L hits 10 units, order more. This is purely reactive. It’s like driving by only looking in the rear-view mirror.
That spreadsheet has no idea that blue t-shirts are trending on TikTok. It doesn't know a holiday weekend is coming up. It can't see the sales velocity for that SKU has doubled in the last three weeks. It just sees a number and a rule. This is a recipe for constant stockouts or bloated, cash-draining overstock. This tutorial for AI-based inventory management will show you how to build a simple, automated forecasting system. You'll stop guessing and start predicting.
How to Build Your Predictive Engine, Step by Step
We're going to build a system that reads your recent sales data, feeds it to an AI model to forecast future demand, and then updates a spreadsheet with reorder recommendations. No data science degree required. Just a few affordable tools and a willingness to connect them.
Step 1: Get Your Data In Order
This is the absolute, non-negotiable first step. AI cannot predict anything from a messy, inconsistent dataset. Your data hygiene is 80% of the battle.
This is where most people give up.
You need a clean, simple record of what you've sold. A Google Sheets or Airtable base is perfect for this. At a minimum, you need three columns:
-
SKU: A unique identifier for each product variant (e.g., TSHIRT-BLUE-L). Be ruthlessly consistent with your naming. -
SaleDate: The date the item was sold (e.g., 2026-10-23). -
QuantitySold: How many units were sold in that transaction.
If your e-commerce platform (like Shopify) or Point-of-Sale system can export a clean CSV of this, you are way ahead. If not, you need to start tracking this manually. It's tedious, but a few weeks of data is all you need to get a useful baseline.
Step 2: Forget Training a Model. Use an API.
I see so many guides suggesting you should use complex libraries like TensorFlow or Facebook's Prophet to build a forecasting model. Honestly, for a small business, that's a complete waste of time. It's an engineering project in itself. Instead, we're going to use a large language model API, like OpenAI's GPT-4o or Anthropic's Claude 3 Sonnet.
Why? Because they are shockingly good at time-series forecasting when given clear instructions and clean data. We can treat the AI as a rented data scientist that we pay fractions of a penny per request. It’s the ultimate shortcut.
Step 3: The Master Prompt for Forecasting
The core logic of our system lives in the prompt. This is where we instruct the AI on its role, give it the data, and tell it exactly what we want back. You can copy and paste this directly into your automation workflow.
You are an expert supply chain analyst specializing in inventory forecasting for small e-commerce businesses.
Analyze the following historical sales data, provided in CSV format with columns: SKU,SaleDate,QuantitySold.
{{PASTE_YOUR_CSV_DATA_HERE}}
Based on this data, provide a sales forecast for each unique SKU for the next 30 days. Consider any trends, weekly patterns, or seasonality you can infer from the data provided.
Your output MUST be in a clean, parsable JSON array format. For each SKU, provide the 'sku', the 'forecasted_30_day_sales' as an integer, and a 'confidence_score' (low, medium, high) based on the volume and consistency of the historical data for that SKU.
Example output format:
[{"sku": "TSHIRT-BLUE-L", "forecasted_30_day_sales": 85, "confidence_score": "high"}, {"sku": "MUG-LOGO-RED", "forecasted_30_day_sales": 22, "confidence_score": "medium"}]
Step 4: Automate It with Make (formerly Integromat)
Manually running this process is pointless. The goal is automation. We'll use Make.com (formerly Integromat) to build a workflow, which they call a Scenario. It will run on a schedule without you lifting a finger.
-
Trigger: Schedule. Start your Scenario with the
Schedulemodule. Set it to run once a week, maybe every Sunday night at 10 PM. - Get Data: Google Sheets > Search Rows. This module will connect to your sales data spreadsheet and pull the raw information. Set it to grab all rows from the last 90 days.
-
Aggregate to CSV: Tools > Text Aggregator. This is a key step. The OpenAI API works best when you feed it a clean block of text. We need to convert the individual rows from Google Sheets into a single CSV-formatted string. The Text Aggregator module will loop through every row it received and build this for you. You'll map the columns like this:
{{1.SKU}},{{1.SaleDate}},{{1.QuantitySold}}and set the row separator to a new line. -
Call AI: OpenAI > Create a Completion. Here's where the magic happens. Select the
Create a Chat Completionaction. Choose your model (I recommendgpt-4o-minifor its speed and low cost). In the prompt section, paste the master prompt from Step 3. Crucially, in the spot where it says{{PASTE_YOUR_CSV_DATA_HERE}}, you will insert the text output from the Text Aggregator module. - Parse Forecast: JSON > Parse JSON. The AI will respond with a text string that looks like JSON. This module turns that string into structured data that Make.com can use in subsequent steps.
-
Update Inventory Sheet: Google Sheets > Update a Row. The final step. This module will loop through the JSON output from the AI. For each SKU forecast, it will find that SKU in your master inventory sheet and update a column called
Forecasted 30-Day Saleswith the new number. You can then have a simple formula in another column:IF(CurrentStock < ForecastedSales, "REORDER", "OK").
The first time you see that sheet update itself with an accurate forecast is a fantastic feeling. It's a small taste of running a business on autopilot.
Is this setup actually affordable?
Yes. It’s almost free.
Let's break down the costs. Make.com has a free tier that includes 1,000 operations per month. The workflow we just designed uses about 6 operations per run (plus one for each row it has to aggregate and update). Even with 100 SKUs, you'll likely stay under the free limit if you run it weekly. If you need more, the Core plan is $9/month. I think it's one of the best values in software, period.
What about the OpenAI API? Let's do the math for a store with 50 SKUs and 90 days of sales data. The input data might be around 3,000 tokens. The AI's JSON response might be 500 tokens. Using gpt-4o-mini, which costs a tiny $0.15 per million input tokens and $0.60 per million output tokens, a single forecast run costs about $0.00075. That is not a typo. You could run this every single day and you wouldn't even spend a dollar a month on the AI part. The value is frankly ridiculous.
What most guides get wrong about AI inventory management
Building the system is the easy part. Using it correctly requires you to avoid a few common traps.
First, the old saying holds true: garbage in, garbage out. If your SKUs are a mess (tshirt-blue, T-Shirt Blue, blue-t-shirt-large all for the same item), the AI will treat them as three different products and your forecast will be useless. Clean your data. Standardize your naming convention. This is the most important part.
Second, the AI only knows what you tell it. It can't read your mind or your marketing calendar. If you're about to launch a massive 50% off sale, you have to add that context into the prompt. A simple sentence like, "Note: The upcoming 30-day period includes our annual Black Friday sale. Historically, demand for promotional items increases by 500%." will transform a mediocre forecast into a brilliant one.
Finally, this is not a "set it and forget it" machine. It's a decision-support tool. My biggest screw-up with an early version of this system involved a client who sold imported goods. The system worked perfectly for months, then started sending urgent "REORDER NOW" alerts for half their catalog in July. The problem? Their main supplier was in Italy and shut down for the entire month of August. The AI, having no knowledge of European holiday schedules, just saw inventory dropping without replenishment and panicked. It was my fault for not building that context in. You, the human operator, still have to apply that final layer of real-world knowledge.
Adjacent reading: AI meeting tools coverage.
You can absolutely build this yourself using the steps above. It’s a great weekend project. But if you want to skip the trial-and-error and deploy a pre-configured version of this entire system in an afternoon, we've packaged it as a deployable blueprint. You can find it at deepusecase.com/vault.
Originally published at deepusecase.com
Top comments (0)