Data infrastructure for the agentic era
Deploy AI securely while keeping data sovereignty and governance, with the multimodal Data Lake + security dome pattern.
43 pages
9 sections
EN language
PDF format
The central idea, before the detail.
Deploying AI in your company raises four questions: how do you control who accesses which data? How much will you spend on tokens? How do you keep hallucinations out? Where does your data end up? This whitepaper sets out a model for deploying AI safely, designed for the European regulatory framework, with sovereignty over your data and governance of every access.
9 sections · 43 pages of technical detail.
- - Introduction Why most AI agent projects never get past the pilot: neither the data nor its governance is ready for agents. page 3
- - The thesis An AI is only as good as the knowledge it operates on; the difference between pilot and production lies in the data and security infrastructure. page 5
- 01 An agent without context is expensive and ineffective What happens to an agent that has to rebuild context on its own: it consumes more resources and fills the gaps with inference. page 6
- 02 From raw data to sourced knowledge The first two layers: a data lake with data-level lineage, including unstructured content, and a semantic layer that describes it as a business graph. page 10
- 03 The dome: atomic security for the age of agents Why role-based control does not work for agents and how the cell works: every access is decided by agent, task, data and context, and is recorded. page 18
- 04 The complete architecture How the three layers operate as one system, with a single query interface for people and for agents from any provider. page 27
- 05 Day one The value available from the first data set, what the first agent that knows the business adds, and why each additional agent costs less. page 32
- 06 Why Europe, why now How the architecture responds to GDPR, DORA, the EU AI Act, ISO 27001 and ENS, and why data sovereignty requires your own infrastructure and open source. page 36
- - Conclusion The document's three ideas: truth is in the data, memory emerges from well-organized data, and security is inseparable from the data. page 42
Frequently asked questions
- Why do AI agent projects fail when moving to production?
- Because neither the data nor its governance is ready for agents. Business data is scattered, unrelated and undescribed, and permissions designed for people do not work for an autonomous system. The agent has to rebuild context on its own, which makes it expensive and unreliable.
- What is data-level lineage?
- It is the record, for each individual piece of data, of its source system, the date it entered and the transformations it has gone through. It is recorded per field and per record, not per batch or per file. It lets you walk any answer back to the specific data that supports it.
- What is the security dome, and how does it decide each access?
- It is the security layer that wraps all the knowledge and is the only way to reach the data. Every access is decided at the moment it happens by evaluating four dimensions: which agent is asking, for which task, about which specific data and in which context. Every decision is recorded.
- Why does role-based access control not work for AI agents?
- Because role-based control relies on a person's judgment: they know not to use data outside its purpose even when they have access. An agent does not apply judgment, it executes. It also runs thousands of queries per minute and combines sources, and roles are static while its activity changes with every task.
- How is unstructured content, such as contracts or calls, protected?
- The control reaches inside the document. Depending on the policy that applies to each agent and task, the parts the task does not justify are hidden before the content is delivered. An agent analyzing the tone of calls receives the transcript without account numbers or balances.
- Why should an agent be defined by task and context?
- Because security needs something to evaluate each access against. A generic agent that "helps with whatever is needed" has no bounded task: almost any access becomes defensible and none can be guaranteed. With specific agents, every access is decided deterministically against the policy. They also perform better.
- What is the headless model, and why does it affect data sovereignty?
- It is the architecture in which a platform exposes data through APIs for agents to consume, usually billed by usage. If the data an organization generated lives on the provider's infrastructure, it has to pay for its own agents to read it, and that data is subject to the provider's jurisdiction.
- How much do you have to build before getting value?
- You do not need everything in place. From the first data set that is properly ingested, related and described, the knowledge can already be queried and answers business questions that used to require manual work. The first agent costs the most; each later agent only needs its question, its task and its policies defined.
- What does the whitepaper include?
- It is 43 pages in six chapters, with an introduction and a conclusion. It explains the three layers of the architecture (data lake with lineage, semantic layer and security dome), how to start getting value and the response to five frameworks: GDPR, DORA, the EU AI Act, ISO 27001 and ENS. It defines terms such as lineage, corpus and ontology.
Data infrastructure for the agentic era.
The safest strategy to make your AI feel trained on your data.
Received.
We have received your request — the team will get the document to you shortly.
Download now Condensed edition — 14 pagesWant to go through it with your case in front of us?
Request a technical meeting