1. The Standard Recipe and Where It Stops
Generic retrieval-augmented generation follows a standard recipe. Documents are cut into chunks. Each chunk is turned into a vector. When someone asks a question, the system finds the chunks closest in meaning and hands them to a language model to answer from.
For a company knowledge base full of help articles, the recipe works well enough. A family office archive is a different object. It holds trust deeds, amendments, side letters, tax opinions, valuations, entity charts, board minutes, insurance schedules and years of correspondence. The documents refer to each other, supersede each other and contradict each other on purpose. A nearest-match search treats them as a pile of paragraphs.
2. Six Places Generic Retrieval Fails Private Wealth
Which version is current. A trust amended three times exists as four documents. The closest match to a question may be the original, which is also the most clearly worded. A system without a model of supersession will quote it with confidence.
Which entity is meant. A family may hold assets through a dozen vehicles whose names differ by one word. Semantic similarity cannot tell them apart. A mistake here is a misstatement of who owns what.
Who may see the answer. A principal, a spouse, an adult child, a trustee and an accountant each hold a different slice of the archive. Retrieval that ignores the boundaries of a person's access will place the wrong paragraph in front of the wrong reader.
What the answer rests on. An executive who is about to sign needs to see the clause, its page and its date. A fluent paragraph with no traceable source is a liability in a setting where decisions are reviewed afterwards.
What follows from the answer. Reading a document is rarely the goal. Someone must then draft the letter, move the meeting, notify the adviser and set the reminder. Retrieval returns text. It does not act.
Where the system runs. A retrieval pipeline that calls out to a hosted model or a hosted vector store sends the most confidential material in the family to a third party, one question at a time.
3. What a Bespoke Intelligence Layer Does Instead
SARION Cortex is our local cognitive engine and agentic orchestration layer. It is built into SARION Appliances or deployed onto secure infrastructure the client owns. It is configured for each client, not sold as a fixed feature list, and that is the point of it.
- Local vector retrieval. The family's archive is indexed on the appliance, covering governance papers, contracts, blueprints and estate documents, with fast semantic retrieval and no dependency on an outside service.
- Retrieval shaped to the archive. The index is organised around how this family's documents actually relate: which entities exist, which instruments are current, which matters belong together. That structure is agreed with the client, not guessed by a general-purpose pipeline.
- Access that follows the person. Context is separated by family member and by role, with sharing that the family chooses explicitly.
- Orchestration with tools. Cortex coordinates calendar, communications and analytical tasks, so a question can end in an action the principal has approved.
- Air-gap capable. The whole layer can run isolated from the internet with no external telemetry.
- Presence across the property. Cortex connects residential touchpoints, private study terminals, yacht staterooms and boardroom enclaves, with audio and data staying inside private walls.
We do not publish benchmark comparisons between Cortex and generic retrieval, because the right test is the client's own archive and the client's own questions.
4. A Worked Example (Illustrative)
A principal asks: what do we need to do before the Lisbon property transfers into the new holding company?
A generic pipeline retrieves the five paragraphs most similar to the words in the question. They may come from a superseded draft, mention the wrong holding company, and include a section the principal's adviser was never meant to see. The model assembles a plausible answer from them.
A bespoke layer is configured to treat the question as a task. It identifies the entities involved, works from the instruments currently in force, respects who is entitled to see each, lists the obligations with their sources, and proposes the follow-up actions for the principal to approve: a note to counsel, a calendar hold, a reminder before the deadline.
The difference in outcome comes from structure and process, not from a bigger model.
5. Executive Operations Follow the Same Logic
The same pattern applies beyond the family archive. A chief of staff coordinating a principal's week, a legal team preparing a transaction, a security lead managing a residence: each works across documents, people and deadlines with confidentiality attached. Retrieval is the first step. Orchestration, permissions and traceability are the work.
6. Keeping It Accurate Over Time
An archive changes every month. New instruments arrive, entities are created and wound up, people join and leave the circle of trust. A layer that is configured once will drift.
Cortex is commissioned with a continuing engineering relationship that covers periodic re-indexing, agent tuning, security patches and support. We treat the layer as part of the family's infrastructure, which needs maintenance in the way a building does.
7. Questions to Ask Before You Commission Any AI Layer
- Where does retrieval run, and does any part of it leave the premises?
- How does the system know which version of a document is in force?
- How does it separate what each person may see?
- Can every answer be traced to a clause, a page and a date?
- What happens when the answer should be an action rather than a paragraph?
- Who maintains the layer, and how often is it re-evaluated against your own questions?
A system that answers all six is closer to a private office function than a search box. To discuss how this would be configured for your archive, request a confidential consultation through our Cortex page.