Ecommerce AI should stay local when the answer only becomes useful after it sees sensitive business data: customer history, margin notes, supplier terms, stock rules, internal documents or operational decisions.
A public chatbot can help with generic wording. It can draft a product description or clean up a paragraph. But once the question contains information you would not send outside the business by email, the setup deserves more thought.
The point is not to make every AI project heavy. The point is to know which questions are ordinary and which questions should stay close to the data.
What makes a question sensitive?
A question becomes sensitive when the useful answer depends on private context.
For a webshop, that context can be commercial, personal or operational. It may be a customer's buying rhythm, the real margin on a product, a supplier agreement, a private stock rule, a complaint history or a document that explains how the business makes decisions.
The AI does not need to be dangerous for the risk to be real. Sometimes the risk is simpler: the team pastes more context than it meant to paste, or a useful workflow slowly becomes a habit before anyone has decided where the data is allowed to go.
That is usually where local AI starts to make sense.
Cloud AI still has a place
Keeping some AI local does not mean every cloud tool is wrong.
Cloud models are useful for broad tasks: rewriting a public text, summarizing non-sensitive notes, creating variations, checking tone, or helping a team think through a generic process.
The line is crossed when the model needs the private material to do the job. If the answer is only good because the prompt includes customer records, internal documents or commercial terms, the question is no longer generic.
At that point, the operator question changes from "which model is best?" to "where should this data be while the model reads it?"
Local does not have to mean training on secrets
Many teams hear local AI and imagine a model being trained on every private document in the company. That is not the only way, and often not the first way.
A practical setup can keep documents as documents. The model receives only the relevant pieces when a question is asked. That pattern is often called retrieval, but the operator version is simpler: the AI reads from an approved shelf instead of memorizing the whole room.
That matters because business knowledge changes. Prices change, product rules change, stock rules change, and internal policies get updated. If the documents remain the source, the team can inspect and correct the source.
The answer becomes easier to trust because it points back to the material it used.
What to test before choosing a setup
Start with one real question, not a technology choice.
A useful test is: what would the team ask, what data would the answer need, who is allowed to see that data, and what would happen if the answer were wrong?
Then compare setups. Try the question with a generic cloud model, a generic local model and a controlled setup that has access to the right business context. The answer is not always the most impressive sentence. It is the one the operator can use without losing control of the data underneath it.
That is the thinking behind Butterstreet's local AI work. We start with the question and the data boundary. The model comes after that.