Frequently asked questions

An AI agent is software that carries out a defined piece of work the way a trained colleague would. You brief it in plain language, it reads your documents, applies your rules, produces the output, and hands it back for a person to check.

It is not a chatbot you ask questions. It is configured to do one job inside your workflow, well, every time.

An assistant answers when you ask, one question at a time, and forgets the context when you close the window. An agent is set up for a specific role, holds the rules and context of that role, and carries a task from input through to finished output.

The first helps a person type faster. The second takes a piece of work off the desk.

A consumer tool like ChatGPT knows nothing about your office, your templates, or your rules, and whatever your team types into it leaves your control. Our agents run inside your own environment, are configured to your documents and your way of working, and produce output held to your standard.

One is a clever search box. The other is a colleague who knows how you work.

In our conversations with clients, what we hear most is that other providers offer a SaaS platform or product you connect to and then have to make useful yourselves. We work the other way around: we build the AI agents inside your own environment and configure them to how you already work.

It helps that we come from finance rather than pure technology, so the agents are shaped around your workflows from the start. Salama spent twelve years at Vontobel serving family offices; Tomasz spent seventeen years at Microsoft building enterprise AI. You work with the two of us directly, and we stay with you as the agents run.

No. There is no Fiducia AI portal and nothing new to log into. The agents work inside the tools and environment you already use, so your team keeps working the way it works today while the agent does its part in the background.

A piece of work that used to sit on someone's desk arrives already done, ready to check. The analyst opens the draft report instead of building it. The reviewer reads the flagged items instead of hunting for them.

The person keeps the judgment; the agent removes the manual assembly that came before it.

You keep your systems. The agents connect to the tools you already run, such as Microsoft 365 and Teams, your document stores, and your existing files. We fit the agent to your setup rather than asking you to move to ours.

No. We handle the setup, the updates, and the maintenance. You do not need in-house IT or technical staff. The agents run on infrastructure you already have or that we stand up for you, and keeping them current is our job, not yours.

The demand is at the start, a few focused conversations so we understand how you work. After that the agent gives time back rather than taking it. We design the first piece of work to fit into a full week, not to add to it.

The agent is placed inside your own environment, next to where your data already sits, whether that is your own drive, your Microsoft or AWS tenant, or your own hardware. Your data does not move to us for the agent to work. It comes to the data, not the other way around.

Fiducia AI is not a SaaS platform. We deploy the agents directly into your existing infrastructure rather than operating as a shared service that all clients use. For most family offices, that means inside your dedicated Microsoft Azure tenant or AWS account, with secure connections to your on-premise systems where needed.

This architectural choice changes how several of the standard SaaS vendor risks apply to your operation, and it is what makes the answers to confidentiality, data, privilege, and continuity work the way they do. Your data and your workflows never leave your environment.

Yes. For senior family offices and holdings that run their own infrastructure and refuse anything cloud, we offer a sovereign deployment pattern: an open-source AI model running on your own hardware inside your walls. Nothing leaves your perimeter, and no US-incorporated provider is in the chain.

We are direct about the trade-offs (longer setup, additional configuration work, hardware investment on your side) before you commit. This pattern is scoped and priced separately from a standard engagement.

No. The architecture is model-agnostic by design. You can move between Claude, GPT, Qwen3, Apertus, or any future frontier or open-source model with a configuration change rather than a rebuild. The same applies across Microsoft Azure, AWS, and your own hardware. Lock-in is a vendor problem, not a Fiducia AI design choice.

Operational continuity is built into the architecture. Because agents run inside your own environment, they are governed by your existing business continuity plan and your cloud provider's SLAs, the same framework that protects the rest of your IT operations. There is no separate Fiducia AI platform that needs to stay online for your team to keep working.

On our side, we maintain incident response procedures and can quickly restore agent configurations; on your side, your existing BCP already covers the infrastructure agents run on.

Confidentiality is governed by your existing access controls and tenant boundaries, not by ours. The agents run inside your dedicated Azure tenant or AWS account, with secure connections to on-premise systems where needed. No data leaves your tenant. No information is shared between clients. Each engagement is fully separated, so the professional secrecy you owe your own clients is preserved end to end.

The agent runs inside your own security perimeter, protected by the same controls that guard the rest of your systems, and it holds no separate copy of your data outside your walls. An attacker who reached the agent would face your access controls and your encryption, the same barriers they would face reaching any other tool in your environment.

Because we keep no copy on our side, there is no second store to target. We sit inside the perimeter you already defend rather than widening it.

No on both. Your data stays in your environment; we do not keep a copy on our side. Each client's agents and data are fully separated, so nothing you hold is visible to, or reachable by, anyone else we work with.

No. The enterprise AI services we use inside your tenant, such as Azure Foundry or Amazon Bedrock, are governed by contractual no-training and data-handling commitments from your cloud provider. Your data is not used to train third-party models, and is not visible to other clients or to Fiducia AI.

A direct answer, since most providers do not give one. The Azure Switzerland and AWS Zurich regions are physically Swiss. Their corporate parents, Microsoft and Amazon, are US-incorporated and remain within reach of the US CLOUD Act. No private contract overrides US extraterritorial law.

For most family offices the practical risk is low: the underlying principals are not US persons, the data has no US nexus, and with customer-managed encryption keys the cloud provider holds only ciphertext it cannot read without your cooperation.

For senior family offices and holdings whose threat model includes adversarial legal process, governmental compulsion, or geopolitical risk, we deploy the sovereign on-premise pattern instead.

Only if you choose a setup where it does. The agents can run in the region you specify, and for an office operating across more than one jurisdiction we keep each in its own region, so nothing crosses a border you would rather it did not.

No. An AI agent inside your own environment functions like any other authorized internal tool, comparable to secure email between you and your client. It receives content, applies its rules, and returns the output to the same authorized user. Nothing leaves your confidentiality perimeter. Privilege concerns generally arise when privileged material is shared with external consumer AI services, which is the opposite of how Fiducia AI is deployed.

No. An agent does only the work it is configured to do, within limits you set, and it does not expand its own reach. It improves in the sense that it gets better at your way of working as we refine it, not in the sense that it decides to do more than it was asked. It waits to be asked, and it acts inside a defined boundary.

The agent can reach only what you give it access to, governed by your existing permissions. Anything outside that scope stays outside it, and you decide, at setup and at any time after, what it can see and what it cannot.

A person signs off. The agents do the work and prepare the output; a person decides what happens next, and for anything that leaves your office or touches a client, a person approves it first. Critical decisions stay with people, and every action the agent takes is logged.

A person reviews and approves before anything is used, exactly as they would with work drafted by a junior colleague. The agent produces the draft; your team owns the decision to send it. We build that review step into the workflow, and we treat any error as something to correct in the agent's rules so it does not recur.

Mistakes are treated as input, not incidents. Most fall into two categories, and both are fixable: if the agent misunderstood an instruction, we refine the prompt or rule; if the agent lacked context, we add it to its knowledge base. A structured feedback loop captures these corrections so the agent gains experience over time rather than repeating the same mistakes.

Critical decisions stay with humans in the loop, and every action is logged. Agents are tools, not unsupervised autonomous workers.

The agent shows its working. It points back to the source for the figures it reports and flags what does not reconcile rather than smoothing over it, and when it does not have enough to answer, it says so and asks instead of inventing an answer. That behaviour is deliberate, and it is part of why we hold a minimum quality bar on the models we deploy.

Yes. We will not deploy any model below a defined size and capability floor for due-diligence and reporting work. Below that threshold, models hallucinate and miss subtle clauses at rates that are not acceptable for high-stakes use.

This is a technical floor. Commercial considerations do not lower it. It rules out the cheap consumer-grade models some vendors quietly use behind the scenes.

Yes, at any time. The agents run in your environment, so control sits with you. You can pause or stop one whenever you choose, and because your systems and data are yours, nothing you depend on stops working when you do.

That is a question for your own legal and regulatory advisers, and the answer depends on your jurisdiction and your mandate. What we can say is that the agents work in your back office, on your documents, under your review, comparable to any other internal tool your team uses, and we are glad to give your advisers the technical detail they need to make the call.

No. The goal is to invert the 80/20: today most of your team's time goes to admin and a fraction to the relationship. Agents take the admin layer so your people can spend their time on what AI cannot do, which is the trusted human relationship at the core of this business.

She keeps the work that needs a person and loses the manual grind. The agent takes the repetitive assembly, the copying between systems, the first draft of the recurring report, and your assistant moves to the judgment, the relationships, and the exceptions, the parts of the role a person does better than any tool.

Offices that do this well end up with the same people doing higher-value work.

What we hear from clients is that the more useful question is what they want their people spending time on. Most would rather their team focus on the meaningful conversations and the judgment that stay with a person, and have the manual, repeatable work taken over by an AI agent.

Sometimes a new hire is still the right call, for a role that is mostly relationship and judgment; often an agent frees the people you already have to do more of that work. The diagnostic helps you see which fits your situation.

Hand over the work that is manual, repeatable, and rule-based: pulling figures out of documents, cross-checking them, assembling the recurring report, filling the standard form. Keep with a person the work that is relationship, judgment, and discretion: the client call, the difficult decision, the read of a situation.

If you would comfortably brief it to a careful junior, it is a candidate; if it needs your name on it, it stays yours.

Very little, and no technical skills. Your team briefs the agent in plain language, the way they would brief a person, so there is nothing like prompt engineering to learn. Most of the learning runs the other way, the agent learning your templates and standards, which we handle. A short walkthrough is usually enough to start.

You often cannot see it from the inside, because the manual work feels normal until someone maps it. That is what the first step is for: we look at your workflows, your documents, and your tools, and show you where an agent would give back the most time, in what order to build, and where it would not add value. You get an honest read before committing to anything.

Yes. Rather than a generic demo, we show the agent working on your material, your templates, your documents, your kind of case. Seeing it draft from your own work is the point, and it is how you judge whether it is worth it on your reality rather than a staged example.

Yes, and it is one of the strongest uses. The agent learns your templates, your house style, and your past work, and drafts from them, a deed from your precedent, a report in your format, with a person reviewing and finalising. It is faster because it starts from how you already do things, not from a blank page.

Yes. Filling standard internal forms from information the office already holds is exactly the structured, repeatable work an agent handles well, and it often makes a good first piece precisely because it is tedious and rule-based.

Start small, always. The best first piece is one specific, repeatable task that eats time and does not need a person's judgment to produce the draft: a recurring report, a standard form, extraction across a set of documents. One workflow, one clear result, live in weeks, and you expand once you have seen it work.

A first workflow is usually live in weeks. Getting it fluent in your standards, your templates, and the way you like things done takes a little longer and improves with use, as we feed your corrections back in. Think of it as a capable new colleague: useful quickly, and sharper at your house style over the first months.

Yes. The agents can work across offices and jurisdictions, and where data must stay in a given country we keep each office's setup in its own region. One consistent way of working across locations, without moving data across borders you would rather it did not cross.

It runs in three steps: a diagnostic, where we map your workflows and show where agents help most; a pilot, one agent on one workflow, proven on your data; then an ongoing managed service, where we run, maintain, and improve the agents as your needs change.

It is an ongoing service rather than a one-off purchase, because agents are not something you install once. They need maintenance, updating, and refinement to keep performing, and that is the part we carry for you.

We do not put a number on it before the diagnostic, because the right scope depends on the work, and a figure named too early is usually the wrong one. The useful comparison is the cost of building the same capacity by hiring, the senior people you would need to do this work by hand.

An agent delivers that capacity for a predictable managed fee, live in weeks rather than quarters, and we share the specific investment once the diagnostic has shown what the work actually is.

A fair question when you are working with a small firm. The agents run in your environment and do not depend on either of us being present day to day; once built, they keep working. Between us we cover both the commercial and the technical side, the configurations and knowledge are documented rather than held in one head, and we are candid about how we manage continuity.

You do. What we configure and build for your office is yours, and this is written into the engagement so there is no ambiguity about ownership of the work or its outputs.

The agents live in your environment and the work is yours, so you are not left with nothing if we part ways. We are direct about what keeps running on its own and what needs ongoing maintenance to stay current, so you can decide with a clear picture. No lock-in by design.

Yes, privately. This is a discreet market, and we treat our clients' confidentiality the way we would treat yours, so we do not publish names. Named references are shared in conversation, with the client's permission, when you are seriously considering working together, and the anonymized engagements on our Use Cases page give you the shape of the work in the meantime.

Family offices and the holdings and investment companies that sit alongside them. The common pattern is small expert teams carrying a large mandate, still running on Excel sheets, manual reporting, and fragmented systems.

Four core pillars, plus custom workflows: investment due diligence on funds, deals, and portfolio companies; portfolio monitoring and consolidated reporting; board, committee, and family reporting; research and market intelligence. If your team can describe how they do something, we can build an agent that does it alongside them.

Different category. Microsoft Copilot is a productivity assistant embedded in Word, Outlook, and Teams. It uses a small fraction of what the underlying models can do.

We build on Azure Foundry, which gives direct access to the frontier models (Claude, GPT, and others) with the configuration controls required for due-diligence, reporting, document analysis, and agentic workflows. Most family offices running Copilot today are using a small slice of what is available, often without realizing it. Foundry is where the real capability sits, and building on it at scale is exactly what Tomasz did at Microsoft for 17 years before co-founding Fiducia AI.

Yes. For clients with strong sovereignty requirements, we deploy Apertus, the Swiss national large language model built by EPFL and ETH Zurich, with native Swiss German support. This is the strongest sovereignty option for senior family offices, and it pairs naturally with our sovereign on-premise deployment pattern.

Concrete examples from current engagements: a family office in Zurich reduced quarterly due diligence from 120 hours per analyst to 12, in one week. The agent also surfaced five forecasting errors that had gone undetected across multiple analysts and years.

A family office in Basel reduced monthly reporting from one full day to two to three minutes per report, with personalization per family.

We start with a focused pilot on one workflow, typically delivered in weeks. From there we expand based on what we learn together. You see ROI inside the pilot.