A prestige property manager in Monaco spends a considerable share of their weeks producing documents nobody reads with pleasure : notices, minutes, service charge calls, reminders. And they spend the rest doing what a wealthy resident truly pays them for : being reachable, reassuring, present when a water leak threatens a ten-figure apartment on a Sunday evening. The tragedy of the profession is that the first job devours the time meant for the second. The thesis of this article runs against the prevailing fear : well placed, AI does not dehumanise property management, it re-humanises it. On one strict condition : entrust it with exactly the repetitive and the nocturnal, and with nothing that calls for genuine attention.
The real job : two demands nothing naturally reconciles
Managing a high-end Monaco co-ownership means holding two seemingly contradictory promises. On one side, flawless administrative rigour : legally sound minutes, a service charge call accurate to the penny, a notice sent in the proper form and within the deadlines. On the other, service of the very highest order : availability, tact, the memory of a resident's preferences, the ability to defuse a tension between co-owners who pass one another in the lift every morning.
These two demands compete for the same scarce resource : the manager's attention. Every hour spent transcribing decisions into minutes is an hour not spent anticipating a neighbour dispute or attending to a sensitive resident. Most firms resolve this tension by silently sacrificing service : they reply more slowly, they standardise, they delegate to an overstretched junior. AI allows it to be resolved differently, by removing from the chain what never needed a human brain in the first place.
The useful question is not “what can AI do in my place ?” but “what, in my week, does not deserve my attention, and stops me from giving it where it is expected ?”
The annual general meeting chain, step by step
The AGM is the best example, because it is a long, standardised and repetitive chain of which only a single step demands genuine judgement. Let us break it down honestly.
Upstream : building the agenda (resolutions, quotes to be voted on, items carried over from last year), drafting the notices within the deadlines, assembling the supporting documents (contractors' quotes, contracts, provisional budget, accounts), then routing everything to the right co-owners with the right proxies and mandates. During the meeting : the conduct of proceedings, the attendance roll, the tally of votes according to ownership shares. Downstream, the most time-consuming step of all : drafting the minutes, which must faithfully reflect each resolution, the required majority, the result of the vote and any reservations expressed.
This is where AI fits in, and precisely how. For the preparation, an agent connected to your documents assembles the file : it spots contracts coming up for renewal, reconciles the quotes received with the items to be voted on, flags a missing document before it goes missing. For the minutes, the mechanism deserves to be set out in detail, because this is where many promises come off the rails.
Generating a first draft of the minutes : what that actually involves
Producing minutes automatically does not consist of “asking a chatbot to summarise the meeting”. Technically, three things are needed. First, capturing the decisions : from a transcript of the meeting, or from structured notes taken during the vote, the system identifies each resolution, the applicable majority and the result. Then structuring : this is constrained extraction, not free prose. The model is forced to fill in a template (resolution number, heading, basis for calculating votes, in favour, against, abstention, adopted or rejected), because minutes are a fixed-format document, not an essay. Finally, and this is non-negotiable, letting the manager review and validate.
This last point is not a token precaution. A model can confuse two neighbouring resolutions, misreport a tally, or “smooth over” a reservation voiced by a disgruntled co-owner, a reservation that has precisely a legal value. Good design does not mask this risk : it organises it. The first draft arrives with its sources visible (this sentence of the minutes points back to that moment of the meeting), so that the review is quick and targeted rather than a full rewrite. The manager no longer starts from a blank page on a Friday evening ; they correct, and assume responsibility for, a text already well advanced. The gain is not that the minutes are written by the machine : it is that they are reviewed by an available human rather than rushed by an exhausted one.
Service charge calls and reminders : cold precision in the service of the right tone
Service charge calls concentrate the other half of the repetitive work. Calculation of shares according to ownership tantièmes, allocation between the current budget and works voted on, issuance, tracking of payments, and the most delicate matter of all : the graduated reminders in the event of non-payment.
The calculation itself is deterministic : it does not need a language model, it needs a sound rule and clean data. An important point of honesty : for strictly arithmetical operations, you do not “add AI”, you add reliable software ; AI only comes in where there is text, judgement or language. The reminder, by contrast, is a problem of tone as much as of procedure. Reminding a resident who has forgotten a deadline is not the same as chasing a debtor acting in bad faith ; and in a building where the co-owner in arrears may be dining the following evening with the chair of the management council, clumsiness costs dear.
A well-tuned agent modulates the reminder according to the stage (courteous reminder, firm reminder, formal notice prepared for signature), keeps a complete audit trail of every dispatch (who, when, what amount, what response), and above all never sends the most sensitive level without validation. The reflex to acquire : AI drafts and logs, the human decides to press “send” the moment the relational stakes rise. The machine brings rigour and memory ; the manager brings the deft touch.
The resident concierge : qualify and escalate, never replace
This is where the whole argument is really decided. The residents of a prestige co-ownership are international, demanding, and their requests do not respect office hours. A multilingual conversational agent available 24/7 seems the obvious answer. It only is on one condition : that its mission be to qualify and escalate, never to stand in for the human on what matters.
In concrete terms, the agent handles on its own what is routine and verifiable : concierge opening hours, the procedure for booking the reception room, the status of a parcel, a reminder of a building rule. For everything else, its job is not to answer, it is to prepare the human. When a resident reports at 2 a.m. that water is dripping from the ceiling, the agent does not attempt to handle a claim : it qualifies the urgency, asks for the floor and the nature of the damage, identifies the apartment and the resident's contact details, and escalates to the on-call team with the context already gathered. The human who picks up does not lose ten minutes reconstructing the situation : they act.
This nuance, qualifying and escalating with the context rather than replacing, is what separates a deployment that re-humanises from one that frustrates. A wealthy resident who senses they have been met with a wall of robot to save a salary will feel slighted, and they will be right. The same resident, for whom the AI saves thirty seconds of logistics to put them more quickly in touch with someone who already knows their problem, perceives a superior service. AI here is not the voice of the concierge : it is what allows the human voice to arrive faster and better informed.
How the agent “knows” your building : RAG and tools, without the jargon
For such an agent to be useful and not dangerous, two mechanisms need to be understood. The first is RAG (retrieval-augmented generation) : rather than letting the model invent from what it “believes” it knows, it is forced to draw its answers from your documents (co-ownership rules, minutes, contractors' contracts, building instructions). It does not answer from memory ; it cites your knowledge base. The second is tool-use : to escalate an emergency or create a ticket, the agent calls a precise and auditable function, exactly as an employee would follow a procedure, and not by improvising. This combination of RAG plus tools is what turns a chatty chatbot into a reliable assistant : it knows where it gets its answers from, and what it is allowed to trigger.
Sovereignty and partitioning : non-negotiable with this clientele
Everything above handles data of extreme sensitivity : the identities of wealthy and international residents, their habits, disputes, amounts, sometimes the simple information of whether a person is present or absent from their apartment, which bears on their security. In Monaco, this material is underpinned by professional secrecy, a pillar of the financial centre's reputation, and governed since law n° 1.565 of 3 December 2024, aligned with the European GDPR and the Council of Europe's Convention 108+, under the oversight of the APDP, the Personal Data Protection Authority, which succeeded the CCIN with broadened powers.
The consequence is direct for the architecture. The data must not leave to feed a consumer service hosted outside Europe. The preference goes to sovereign hosting (local players such as Monaco Cloud, Monaco Telecom or Telis, a European private cloud, or on-site installation for the most critical material), and a partitioning by client is imposed : one co-ownership's data never crosses another's, the agent of one building “sees” only that building. Every sensitive action is logged, not as a formality, but because the ability to prove who saw what and triggered what is what distinguishes a serious firm on the day of an incident. For certain exposed residents, this level of discretion is not a comfort : it is one of the reasons they chose Monaco.
What AI cannot do, and why saying so is a mark of seriousness
An honest article must name the limits. An agent may misqualify an ambiguous emergency : the escalation thresholds are therefore set upwards, meaning that in case of doubt the human is disturbed, never the reverse. The extraction of minutes can get a poorly worded resolution wrong : hence the mandatory review before any distribution. On tone, a model can be too familiar or too curt depending on the interlocutor, and it has no awareness of the micro-politics of a management council. Finally, AI has neither mandate nor legal responsibility : it does not sign minutes, does not send a formal notice and does not bind the firm. These limits are not flaws to hide ; they are the lines that draw the division of roles. A supplier who promises to “automate everything” without review exposes you ; experts who show you where to place the human hand protect you.
Where to begin, soberly
The mistake would be to want to plug everything in at once. The right point of entry is not technology but the audit : mapping precisely where your managers' time evaporates (AGM season, service charge cycles, peaks of resident requests), choosing a first use case of high impact and low risk, measuring it, then extending it. The orders of magnitude often quoted (McKinsey's 2024 estimate of 60 to 70 % of administrative tasks potentially automatable, of the order of five hours per week per employee freed up) are worth nothing until they are confirmed on your chain, profession by profession. Worth noting, for eligible entities : a structuring AI project can be co-financed up to 70 % excluding tax by the Blue Fund of the Extended Monaco programme, which changes the equation for a first deployment. This is the approach we apply at Minervia : audit first, sovereign architecture next, human decision always. A good deployment is not recognised by the number of tasks taken away from your teams, but by the time it gives them back for the moments when a resident is waiting for, precisely, a human.