AI Consulting for Family Offices

Architecture, then contract — in that order. We deploy models self-hosted on your infrastructure or in a private cloud tenancy you control, so documents, prompts, and generated answers are processed inside your perimeter and never sent to an external AI provider. Where a component genuinely cannot run locally, we say so explicitly and show you precisely what data would cross the boundary before anything is built, so the office can accept or reject that trade-off knowingly.

Contractually, this is backed by GDPR-grade data processing agreements, confidentiality undertakings covering every individual on the project, named-person access lists, and deletion of any material we held on request at the end of the engagement. As a German company we treat European data-protection standards as the baseline for every client, wherever they are based.

No. That is the single clearest line in our family office work. Self-hosted deployment means there is no third-party provider receiving your content in the first place, so there is nothing to be retained or trained on. Where a private-cloud arrangement is used instead, we configure it so that customer content is excluded from provider training, and we document the setting rather than asking you to take it on trust.

The same applies to us: we do not reuse client documents, models fine-tuned on client data, or extracted datasets in any other engagement.

It is proportionate precisely because you are four people — a small team feels the return on removing reading and reconciliation work far more sharply than a large one. What changes is the shape of the engagement, not the quality of the engineering: one narrow workflow, an interface designed around named individuals rather than roles, and no administration overhead.

The practical test we apply is whether a two-person team could run the system unaided after handover. If a proposed design would require someone to become its part-time administrator, it is the wrong design and we rework it.

Yes, and it is usually the second project rather than the first, because it repays the effort every reporting cycle. We work from whatever each source realistically provides — statement files, portal exports, custodian reports, fund letters, and the office’s own spreadsheets — and build a normalization layer that maps them to one consistent set of positions, entities, and currencies. Where a source offers an interface, we use it; where it offers only PDFs, we extract from the PDFs.

Two things we insist on: every consolidated figure traces back to the document it came from, and mismatches are surfaced as exceptions for a human to resolve rather than silently reconciled. The goal is a view the CIO trusts enough to act on, not a dashboard that looks tidy.

No, and we would not build a system that claims to. Judgment about a sponsor, a family’s appetite for a sector, or the quality of a management team is exactly what a family office exists to exercise. What AI does well is the part before judgment: reading everything that arrived, extracting the comparable facts, checking them against the criteria the office has already stated, and ranking what deserves attention.

So the division is deliberate — AI narrows, humans decide. Screening output always shows its reasoning and links back to the source pages, and nothing is ever rejected automatically without being visible in a review list. The measurable outcome is that the team spends its reading hours on the opportunities worth reading.

Most offices hold decades of reasoning in three fragile places: one long-serving employee, an email archive, and a folder structure only its creator can navigate. We structure that material into a private, searchable record — decisions, the rationale behind them, the documents that supported them, and the entities involved — and put a self-hosted assistant in front of it that answers with citations to the original files.

The effect on succession is practical rather than sentimental: the next generation, or a newly hired CIO, can ask why a holding was taken, what was agreed with a manager in a given year, or how a structure was set up — and get an answer with the source attached, instead of an apology. Access can be scoped per family member and per entity, so widening the readership does not mean opening everything to everyone.

It keeps working, and it stays yours. We deliver the source code, the model artifacts and configuration, the data pipelines, and operating documentation — deployed on infrastructure you own or control. There is no proprietary platform to remain subscribed to, no hosted component of ours in the critical path, and no licence that lapses.

Handover is treated as a deliverable, not a courtesy at the end: documentation written for a non-technical operator, a walkthrough with the people who will run it, and, if you wish, an introduction for whichever internal person or external IT provider takes over maintenance. We are happy to continue on a support arrangement, but the system must not depend on it.

As a matter of policy, we do not name our clients. We do not publish logos, we do not describe engagements in a way that would allow an office to be identified, and we do not ask for a reference or a case study as a condition of working together — the projects published on this site are there because those clients specifically agreed. If you would prefer a mutual non-disclosure agreement in place before the first substantive conversation, that is a normal request and we will sign one.

Internally, family office projects run with a deliberately small, named team, access limited to the individuals who need it, and no client material stored outside the agreed environment.

A named, deliberately small team — typically a lead consultant and one or two engineers — all bound by confidentiality undertakings, listed by name in the engagement documentation, and changed only with your knowledge. We do not rotate anonymous staff through the project and we do not subcontract family office work.

In many engagements the answer is narrower still: where a self-hosted deployment can be developed against synthetic or redacted material, our engineers work on the system without ever handling live family documents, and the real archive is only ever processed inside your environment by the system itself.

It starts with a confidential conversation about where the office is losing time or visibility — usually deal screening, document reading, or consolidated reporting. From there we scope a fixed-price proof of concept with a defined outcome: a working tool on your real material, in your environment, that the office can evaluate internally before deciding anything further.

Pricing depends on the complexity of the workflow, the state of the source material, and the deployment environment, and is agreed in advance for each stage rather than billed open-endedly. Every stage has an off-ramp. Contact us to arrange the initial discussion.