Contents
In short
- There are two kinds of AI in StockAI: the forecasting models, trained inside StockAI on your account's data, and a language model (DeepSeek by default, or the provider your company chooses), for the assistant and the explanatory texts.
- Forecasts never leave StockAI and are never mixed with other companies'.
- The assistant only reads: it cannot create, change or delete anything in your account.
- AI makes mistakes. The assistant's figures are checked against your data, but the decision is always yours.
1.The forecasting models
To forecast demand, StockAI trains statistical and machine-learning models (such as LightGBM, XGBoost, Prophet, ARIMA, ETS, Croston and LSTM networks) on the sales history you upload.
- They are trained on our own server, not on a third-party service: your history is not sent to anyone to be forecast.
- They are trained only on your account's data. We do not mix data from different companies or use yours for another customer's forecasts.
- The trained models are stored in your account and deleted with it.
2.The language model
Some features use a language model from DeepSeek (by default, a provider based in China) or from the provider your company has chosen in Settings, to write text:
| Feature | When it is used | What is sent |
|---|---|---|
| Assistant (in the app and over WhatsApp) | When you write it a question | Your question, the last lines of the conversation, your first name, your role and your company's name, and the account data needed to answer: products, quantities, semáforo, cover, suppliers, lead times, costs and purchase orders |
| Questions about documents | When you ask about documents you uploaded to the assistant | Your question and the passages of those documents related to it |
| Purchasing Panel summary | Automatically when that screen opens | The day's summary: products at risk and their quantities, overstock, amounts and suppliers |
| Forecast and inventory explanations | When you ask for the explanation | The figures of the product or inventory being explained |
| File quality diagnosis | When the quality of an uploaded history is reviewed | A summary of the file (number of rows, dates, quality score) and the warnings for up to 10 products with their codes. The file itself is not sent. |
We never send passwords, API keys or credentials. The connection to the provider is encrypted (HTTPS).
If the service has no provider configured, nothing is sent: the assistant says it is unavailable and the other screens show rule-based text.
Your company's administrators can choose their own provider (OpenAI, Anthropic, Google, Groq, OpenRouter, DeepSeek or a compatible service) with their own key. The data then goes to that provider, under its terms, and not to DeepSeek. If that provider fails, we do not use another one on our own: we tell you what failed and show rule-based text. We store the key encrypted and never show it again.
3.The assistant only reads
The assistant can look things up in your account, but it cannot act: it does not approve purchase orders, register receptions, change data or delete anything, neither in the app nor over WhatsApp. When an answer involves doing something, it tells you which screen to do it on, and you do it.
The same holds for the MCP server you can connect to your own AI assistant with an API key: it only has read tools.
4.How we check what it says
Before showing you an assistant reply, StockAI compares every figure in it with the data the assistant had in front of it. If one does not match, it asks the assistant to correct the reply once; if a figure still cannot be verified, the reply points it out to you instead of hiding it.
That check is done in the assistant. The summaries and explanations on other screens are generated with instructions to use only the figures they receive, but they do not go through that review: compare them with the numbers on the screen.
5.Limits: AI makes mistakes
A language model can misread a question, leave out something important or write something that sounds certain and is not. A forecast can be wrong because demand changes for reasons that are not in your data.
That is why everything AI produces in StockAI is support for deciding, not an instruction. Check the figures and the orders before acting. As the terms say, the purchasing decision is yours.
6.Training third-party models
We do not use your data to train language models, nor do we share it for that purpose. The AI provider (DeepSeek by default, or the one your company chooses) receives the data to answer each request; what it does with them is governed by its own terms. [DEEPSEEK TERMS ON USING DATA FOR TRAINING — check the current contract or API terms and summarise them here]
7.How to avoid AI
- If you do not write to the assistant or ask for explanations, those features send nothing.
- The Purchasing Panel summary is requested on its own, when that screen opens. If you do not want that data to reach the AI provider, do not open the Purchasing Panel: inventory, forecasts and orders can all be used from their own screens.
- Today AI cannot be turned off for a single account from the app. If you need your account not to use it at all, write to us at contacto@stockai.es before uploading data.
Without AI you still have forecasts, the semáforo, purchase recommendations, orders and alerts: all of that is computed inside StockAI without the language model.
Questions about this document? Write to us at contacto@stockai.es.