Sallra
What EU-first AI means at Sallra
11 November 2026
EU-first does not mean putting a European label on every part of a technology stack. At Sallra, it means choosing European infrastructure where the architecture supports it, keeping AI access narrow, and being precise about the provider, the model, and where each part of the product runs.
EU-first is a set of technical decisions
An AI feature is a chain, not one service. The application prepares context, an inference provider serves a model, storage holds the working data, and other systems carry the customer conversation. Each link answers a different question about access, location, and control.
Our approach is to make those decisions one by one. Run the Sallra application in Stockholm. Route AI through a provider chosen for this setup. Keep documents private. Give the AI a prepared task instead of access to the systems behind it. EU-first is the direction of the architecture, not a shortcut around explaining the architecture.
The path from quote to draft
When Sallra prepares a follow-up, the application assembles the context for that task. That can include selected quote and customer details, relevant text from a quote PDF or mailbox thread, timing, and the seller’s own instructions. The server sends that task-specific package to Berget and receives a proposed next step or draft.
The response is checked against the structure the product expects before it is used. Berget does not receive direct access to Sallra’s database, storage, or connected mailbox. Sallra reads the permitted source, selects the context for the task, and sends the request. That separation is a core part of the design.
Provider, model, and runtime are different layers
Berget is the inference provider in this flow. The model family served through it is Google Gemma. Sallra’s application runtime is configured for Stockholm. Those facts describe three separate layers: who serves the AI request, who developed the model, and where the Sallra application executes.
The implementation uses an OpenAI-compatible client format, but the requests are sent to Berget, not to OpenAI. Likewise, serving a Google-developed model through Berget does not change the model’s origin. Clear language matters because “European AI” can otherwise hide several technically different claims inside two words.
Narrow access and explicit boundaries
The AI receives content for a defined task, not credentials or a path into customer systems. Quote documents are held in private storage and opened through short-lived links. Access credentials for connected mailboxes are encrypted on the server. Company and owner scopes are applied in the data layer and in application queries.
These controls do different jobs. Private storage controls document access. Encryption protects mailbox credentials at rest. Tenant and owner scopes control which records a signed-in person can reach. A narrow AI request controls what leaves the application for inference. Trust comes from the combination, not from one broad promise.
The seller is part of the system
AI can prepare timing, a next step, and a draft. It does not own the customer relationship. The seller sees the suggestion, reviews the wording, changes it when needed, and decides whether to send. Sallra does not autonomously send the follow-up to the customer.
That human step is not a fallback added after the AI. It is part of the product flow. The model handles preparation at scale. The person who knows the account remains responsible for the message and the decision. For sensitive business communication, control is a feature of the architecture.
Precision matters more than a label
EU-first is not the same as saying that every service, model, support path, or connected system is European. A Stockholm application runtime does not establish the location of every downstream processor. An EU inference provider does not make the underlying model European-owned. We keep those statements separate because buyers should be able to evaluate each one.
It also means we do not turn an infrastructure choice into an absolute compliance claim. GDPR responsibilities depend on the full processing arrangement, contracts, configuration, and how a customer uses the product. The technical foundation can reduce unnecessary complexity and make the data flow easier to explain, but it does not replace that work.
What this gives a Sallra customer
A customer should be able to ask concrete questions and get concrete answers. Where does the application run? Which provider serves the AI request? Who developed the model? What context is sent? Can the AI reach the database? Who approves the final message? Those questions are more useful than a single label.
Our aim is straightforward: use AI where it removes administration from quote follow-up, keep its access limited to the task, place a person at the customer-facing decision, and choose European infrastructure deliberately where we can substantiate it. That is what EU-first means in the way we build Sallra.
Ready to turn quotes into decisions?
See the next step on every quote.