Trust, privacy, and compliance

How the Privacy Vault protects sensitive data

The Privacy Vault is the data-protection layer at the edge of the Unless platform. It is not a single feature but a set of twelve numbered privacy measures that work together to keep sensitive identifiers away from AI models. The core idea is simple: your data goes in, but sensitive identifiers stay home. This article explains how the Vault filters, tokenizes, and governs personal data, and what that means for your compliance posture.

What the Privacy Vault does

When customer data enters the platform, the Privacy Vault filters and tokenizes it before anything reaches a model. The model sees scrubbed input and generates output. The Vault re-attaches identifiers only when the response needs them, for example to address a customer by name. This happens inside the controlled Unless environment, so raw personal data never reaches the foundation model provider.

The Vault operates at the gateway, which means protection happens before any model call. Sensitive identifiers such as names, emails, phone numbers, and account numbers are detected and removed by default. Where personal context is required, the data is tokenized, and the model only sees tokens. The Vault holds the mapping back to the real identifier, and only authorized components can de-tokenize.

Filtering and tokenization in practice

PII filtering and tokenization are two distinct but complementary mechanisms.

PII filtering detects and removes personal data before it reaches the model. This happens at ingestion, when content enters the platform, and again at runtime, before any model call. You can configure a whitelist of exceptions for fields the model needs to see in plain text. The guidance is to add only fields the model actually needs.

Tokenization goes a step further. Sensitive identifiers are replaced with safe placeholders, and the agent works with those tokens. The Vault stores the mapping back to the real identifier. When the underlying value changes or is deleted, tokens orphan automatically, so old references cannot be reused. This means a data erasure request can be fully processed without leaving stale references behind.

Data residency and sub-processors

All personal data on the platform stays inside the EU/EEA by design. Unless does not rely on adequacy decisions or transfer impact assessments, because the data never leaves the EU. Core infrastructure runs on AWS in Ireland, with auxiliary workloads in EU regions of Microsoft Azure and Google Cloud. Platform sub-processors are EU entities of these three providers, all processing inside the EU. The full list lives in your DPA.

Business tools used for Unless’s own administration may sit outside the EU, but those tools never touch end-user data. Only financial administration uses such tools, and they contain no end-user data.

Consent and data subject rights

Consent management is configurable. You can choose between two modes: consent not required, or consent required, with sub-options for cookieless experiences. A cookieless experience does not require tracking and can show before consent for visitors who have not opted in yet.

Data subject requests are built in. You can process access, correction, and erasure requests from end users through the dashboard. For erasure, you initiate as the controller, the Unless team verifies with a human in the loop, and the platform deletes everything related to the identifier across Living Memory, Conversations, and integrations.

What the Vault does not do

Unless does not train models on customer content, end-user conversations, or any other customer data. Living Knowledge improves through approved content and Team Feedback, never by feeding raw conversations into model training. Prompts and outputs go through model providers under contracts that forbid reuse for general training.

The Vault also does not give the agent unlimited access. You control which fields the agent can see at each moment, and you can change that decision at any time. Source selection, topic-level separation, audience rules, and role-based access are all configured in the dashboard.

Compliance documentation

For procurement and DPO reviews, Unless provides a standard DPA, a sub-processor list, a security addendum, data flow diagrams, and a description of technical and organizational measures. These live in the Compliance Center, and your account manager can send signed copies. Most security or DPIA reviews close inside two weeks.

The architecture is configurable for GDPR, DORA, ISO 42001, EU AI Act, and entities supervised by BaFin and AFM. Per-decision audit trails and transparency are built into the platform, not bolted on.

The Privacy Vault is the reason you can deploy an AI agent without becoming personally liable for every model input and output. It filters, tokenizes, and isolates sensitive data at the edge, keeps everything inside the EU, and gives your DPO the audit trail they need. The controls regulators ask about live where the people responsible for them work, and they are all exportable.