Subprocessors

1. Description

This list names the service providers that we engage to provide Waterglass Waves and that process personal data on our behalf in doing so. It is Annex 3 of the Data Processing Agreement for Waterglass Waves and is incorporated by § 10 of the Product Terms for Waterglass Waves. Which data the platform processes in detail is set out in the Privacy Notice for Waterglass Waves. The date of the last change to this list is stated at the end of this page.

Services you connect to Waves yourself, as well as tool servers and websites an AI agent retrieves on your instruction, are not part of this list. You determine those recipients, not us (§ 8 of the Product Terms for Waterglass Waves).

2. Infrastructure and operation

Hetzner Online GmbH, Germany. Computing power, network, object storage and encrypted backups, operated in data centres in Germany. All platform data is held here, that is account and organisation data, billing data, and the content you process in Waves.

To keep the services reachable, a provider of TLS certificates and a public name service are also involved. Both receive technical connection data only, no content and no account data.

3. Payment processing

Stripe. Payment processing, invoices and the determination of VAT. The data processed are your name, your email address, your billing address, a VAT identification number you provide, and invoices and payment transactions. You enter card details exclusively with Stripe. They do not reach our systems.

4. Email delivery

Resend (Resend, Inc.), processing in Ireland. Account emails only, that is confirmation links, sign-in links, password links and the notice of a password change. The data processed are the recipient address, your name and the name of your organisation. The operational notice to us about a new registration additionally contains the IP address of the registration.

5. AI model providers

You determine which provider processes the content of a run through your choice of model for each AI agent. The console states the place of processing for every model. One point matters for understanding this list: the recipient is whoever operates the deployment we call, which is not always the company that developed the model. Most models are available in a version operated inside the European Union, and for those the model call does not leave the Union. Two things below are transmitted regardless of that choice, and they are named separately at the end of this section: the conversion of knowledge-base documents, and the web search tool.

Recipients with processing inside the European Union. Amazon Web Services EMEA SARL, processing in Sweden, for the European versions of models by Anthropic, DeepSeek and Z.ai. Google Cloud EMEA Limited, processing in the European Union, for the European versions of the Google models. Microsoft Ireland Operations Limited, processing in the European Union, for the OpenAI models. Mistral AI, France, for the Mistral models. Scaleway SAS, France, processing in France, for the models offered under the Waterglass name.

Recipient outside the European Union, and only if you select one of its models: Anthropic (United States), processing in the United States, for those of its models for which no European deployment is available to us.

Providers may retain the transmitted content for a limited period in order to detect misuse of their services, as a rule no longer than 30 days, and that can include inspection by their own personnel. It differs by provider: some store nothing at all for the models we call, others store only what their systems have flagged. Where a provider's systems flag content as suspected misuse, it can be kept longer and examined more closely. This retention is the provider's and serves the security of their service, not any purpose of ours, and it happens whether or not we keep the content ourselves. Whether it can be switched off differs by provider and is not always within our control.

Regardless of the language model selected, we transmit documents you upload to a knowledge base, as well as your search queries within it, to Mistral AI (France), in order to convert them into a vector representation usable for search.

Also regardless of the language model selected, the web search tool transmits the search query to Google Cloud EMEA Limited. The search runs as a query to Google Search using a Google model, which is why the recipient here does not follow from your choice of model. For this feature Google does not undertake to process the data inside the European Union: its data residency terms expressly exclude grounding with Google Search, so the query and the context sent with it may be processed outside the Union even when the model the agent otherwise uses is one operated inside it. According to Google's terms for the feature, Google also stores logs containing the query derived from the prompt, together with that context, for up to three days and not associated with you or your users, and uses them to debug the systems behind the search. That storage cannot be switched off while the feature is in use. The tool is inactive unless you allow it for an individual agent.

6. Changes to this list

We announce the addition of a new subprocessor or the replacement of an existing one on this page at least 14 days before it takes effect, and we inform the affected customers by email. You may object to the change within that period for good cause relating to data protection. Where the objection cannot be resolved by agreement, you may terminate the contract with effect from the date the change takes effect. Where a change has to be made immediately in order to avert an acute security or availability risk, we inform you without delay afterwards.