PIING
Reading Security and privacyAI and your order data
Security and privacy
- Data and privacy overview
- What data Piing stores
- GDPR and your customers
- AI and your order data
- Sub-processors
- Uninstall and data deletion
In this guide
- What is sent to the model
- What is never sent
- Training and retention
- Protection against prompt injection
- Where the results go
Search or browse documentation
Security and privacy 4 / 6
AI and your order data
What Piing: Intelligence sends to the AI model. Only your order and product metadata is sent, the provider does not train on it, and retention is short.
Piing: Intelligence reads a customer's order history and produces structured attributes, such as a preferred category or a likely size, which are written back to your platform as app owned metafields. This article explains exactly what reaches the AI model and what does not.
What is sent to the model
For each customer who has enough order history, Piing builds a prompt from order and product information:
- Product titles, variant names, product types and vendors from the customer's orders.
- Line item prices and quantities, and the date each order was placed.
- A summary of your store's catalogue, such as the product types and tags you sell, so the model uses your real taxonomy.
- Aggregated behavioural signals derived from browsing, where the storefront pixel is in use, such as a browse to buy ratio. These are summary figures, not a raw event log.
That is the full extent of it. The analysis works on what was bought and how it relates to your catalogue, not on who the customer is.
What is never sent
The following are not transmitted to the model:
- Customer names.
- Email addresses.
- Postal addresses.
- Phone numbers.
- Payment or card details.
The customer name and contact details are not part of the prompt. Piing uses a pseudonymous request reference outside the prompt so each result can be matched to the correct Shopify customer during the active run.
Training and retention
The model is provided by Anthropic as a sub-processor. Anthropic does not use the inputs or outputs of these requests to train its models, and the data is deleted on a short retention window. This is set out in the Data Processing Addendum, which also lists Anthropic and the other sub-processors.
Protection against prompt injection
Order text comes from the outside world, so a product title or note could in theory try to smuggle in an instruction. Piing treats all order and catalogue text strictly as data to analyse, not as instructions to follow, and the instruction to do so is kept separate from the untrusted content. This is a deliberate guardrail so that text inside an order cannot change what the model does.
Where the results go
The structured attributes the model returns are scored against the configured confidence threshold and written to app-owned metafields under the namespace $app:piing-intelligence. They live in Shopify, where you can read, export, confirm or delete them. Piing does not retain a separate copy of the customer's order history as the persistent Intelligence output.
Tip
The free preview automatically analyses up to five recent customers with at least one order. It does not need a Shopify tag and consumes no billable allowance.
For a wider view of what Piing stores, see What data Piing stores. For the full sub-processor list, see Sub-processors.
If your security team needs more detail on the AI path, contact us.
Related guides
Was this guide helpful?
Next guide in Security and privacy
Sub-processors
The third parties Piing relies on to deliver its services, what each one does, and where to find the current list.
Still need a hand? Contact us