How Nemma looks after your data.

The separation happens in two places. The database itself rejects all access across shops, whichever screen or query is asking. And in the server code, which has wider rights than a user, the shop is always applied as a filter.

We test the separation automatically, against a separate test environment and never against real customer data. And an automatic test fails if a new type of data is created without the separation.

Within your shop, you decide who may do what. By default only the owner can change settings, refund and cancel, and edit the AI's Skills.

Nemma only looks up orders by the email the customer writes from. If the order's email is different, nothing is handed over - neither to the AI nor to the sender. The answer is the same whether the order does not exist or belongs to someone else.

Because the check sits in the tools the AI uses, it cannot be bypassed with a clever phrase in an email. Nothing about the order reaches the AI until the sender's email has been checked against the order.

One narrow exception: if the customer bought with a different email, a human on your team can confirm that the order belongs to the customer. That applies only to that one order in that one case. The reply is never sent automatically, and the address on the order cannot be changed.

Nemma can propose a refund, but never carry it out on its own. The same applies to an exchange for a cheaper product, where the price difference is paid back. Nothing happens until a human has approved it.

It is not just anyone on the team. By default only the owner may move money. The owner can give a support agent permission, and that permission is checked both in the app and in the database.

A refunded amount can never exceed what the order was paid. If an action dies halfway through the call to Shopify, Nemma never retries on its own, so the same amount cannot be paid back twice. A human settles it.

Without you, Nemma on Auto-send can only cancel an order that has not shipped, is less than a day old and is DKK 2,500 or less, and change the address on an order that has not shipped when the order's own email matches the sender. Everything else waits for you.

Some people try to hide instructions in an email to get the AI to do something it should not. This is called prompt injection. So Nemma reads the content of an email as data and never as commands.

But we do not let security rest on the AI seeing through everything. The most important safeguard is what it can do at all: there is no tool that can send an email to another address, give a discount, create an order or send data out of the system. A reply always goes to the customer's own email in the case, and the recipient cannot be changed by the AI or by an email.

The worst a fully hijacked model can do is propose an action on the sender's own order - and refunds and exchanges still need a human.

Before Nemma sends on its own, the draft is checked for links and email addresses that do not belong to your shop or the customer. If the check finds anything, the reply ends up as a draft for you to approve. The same applies when the sender's email cannot be confirmed: then nothing is sent automatically.

What is processed in Nemma is not used to train or improve AI models - neither by us nor by our AI provider. The provider, Anthropic, is set up without data retention for training purposes, and this is stated in the data processing agreement and the terms of service.

Under its own terms, the provider may carry out limited monitoring for misuse. That is not training, and it is also stated in the agreement.

We never write the content of an email or access keys to our logs. Your Shopify access is stored encrypted and is never shown in plain text, not even to us.

The database and hosting are in the EU. Resend and Anthropic process data in the USA, on the basis of the European Commission's standard contractual clauses and, where the provider is certified, the EU-U.S. Data Privacy Framework.

Your own Shopify is your own system and not our sub-processor. Payment goes through Stripe, which does not process your customers' personal data.

Data is encrypted in transit and at rest. We give at least 30 days' notice before adding or replacing a provider, and you can object.

Every reply that Nemma has helped write ends with a short line in the customer's language saying that the reply was made with the help of an AI assistant. That applies both to replies you have approved or edited and to replies Nemma sends on its own, on email and in chat.

The line is added when the reply is sent, after you have finished editing. So it cannot be edited away, and there is no setting that turns it off. It supports the requirement in Article 50 of the EU AI Act to disclose when someone is communicating with an AI system.

Your fixed acknowledgement and out-of-office reply are your own text, not written by AI, and so do not carry the line.