What AI in property management can really do today

Every property trade fair now has "AI" on every second banner, and every second management office considers it nonsense. Both sides are a little right, and both attitudes cost money: one buys too much, the other leaves real relief on the table.

What follows is about what works today, what does not, and how to tell the difference before you have paid for it.

The one distinction that matters

A single dividing line lets you place any feature in thirty seconds:

AI is good at reading something off or putting it into words. It is bad at deciding anything that depends on context the document does not contain.

Reading off means, for example: which amount is on this invoice? Putting into words means turning these bullet points into a polite, formal message. In both cases the answer is already in the material.

Deciding is different. Take the question whether this contractor's invoice is routine maintenance or a repair. The invoice does not say. The answer is in a resolution or an agreement, or it follows from a history only you know. A model that answers here is guessing. The answer sounds plausible and reads fluently, and it comes with the same confidence as in the cases where the model is right. Every section below applies this rule.

What genuinely carries today

Reading invoices is the clearest case. Supplier, invoice number, amount and date are four fields printed on the document. Net of your checking, that saves roughly half a minute per invoice, and at eight hundred invoices a year it adds up to a working day. The details are in AI invoice capture.

Suggesting a title, a short summary and metadata for an uploaded PDF is also reading off. It is useful mainly because it means the files get looked after at all. A document that takes thirty seconds to name otherwise gets named "later", and then never.

For drafting, think of the reply to an owner writing for the fourth time about the same defect. You know what has to go in it. The work is the wording, and that is hard when you are irritated. Revising a draft is quicker here than facing an empty box.

When a new defect report comes in, AI can suggest a category and urgency and look for likely duplicates. The duplicates are the most useful part. Three owners report the same broken garage door in three different wordings, and a full-text search does not catch it.

Finally, AI can answer questions from documents. An owner asks what the assembly decided about the façade, and the answer is in the minutes he is allowed to see. That is a search task with some wording on top. The condition is that the system answers only from the documents that owner may see, names its source, and honestly says "I don't know" when the answer is not in the records.

What does not carry

Fully automatic booking does not, for the reason given above. The cost type is a decision, the distribution key hangs on it, and who pays hangs on the key. A mistake only surfaces in the annual settlement, in front of thirty owners.

Automatically sent replies are technically trivial and a bad idea in practice. A single message in the wrong tone to the wrong owner costs more time than a hundred drafts ever saved, and you hear about it first from the advisory board.

The promise "AI prepares your annual settlement" does not hold either. An annual settlement is a calculation with statutory rules. What makes it reliable is correct apportionment logic, cent-exact rounding and immutability after approval. All of that is built deterministically because it has to be deterministic, and at this point AI does worse than plain arithmetic.

"AI checks your resolutions for legal soundness" would be legal advice. A model can point out the typical formal errors, and as a pointer that is useful. Whether this resolution in this association can be challenged is something it cannot tell you, and a vendor who implies otherwise is selling you a risk.

That leaves forecasts from your data, such as "AI predicts your maintenance needs". A practice with forty buildings does not have the data for it. Statistics needs cases, and you have a two-digit number of them.

The question that makes the difference

Whenever someone demonstrates an AI feature to you, ask: What happens without my confirmation?

If the answer is "nothing", the feature is at worst useless. It suggests something wrong, you discard it, and you have lost ten seconds.

With any other answer, your liability has changed without the contract saying so. You remain responsible for the booking, the settlement and the message. You just did not make them yourself any more and can no longer say exactly why they look the way they do.

That is why all our AI features follow the same pattern: they propose and never carry anything out themselves. Invoice capture reads out, and you do the booking. Triage makes a suggestion that you confirm, and you send any draft yourself. More would be technically possible. We keep it this way because a manager has to be able to stand behind their own acts.

Data protection, short and practical

Before an AI feature touches real data, settle three things.

First, where it is processed. Defect reports, documents and enquiries contain personal data. If they are processed outside the EU you need a solid legal basis; inside the EU it is simpler.

Next, whether there is a data-processing agreement that covers the AI vendor. Your software vendor is a processor, and so is the AI provider behind it. Both have to appear in your records.

And finally, whether content is used for training. The only acceptable answer is no, and it belongs in the contract and not merely on a marketing page.

Our answers are in the privacy notice: processing via a processor in Germany, exclusively on servers in the EU, an Art. 28 GDPR agreement and no use for training. The management company has to switch the features on itself; we do not do it for them.

What it realistically delivers

For a practice with thirty to fifty buildings whose data is already digital, it looks roughly like this:

Together that comes to just under a working day a month. That is respectable, but your practice will change in other places.

More comes from an owner portal that catches phone calls, document storage where nobody has to search any more, and a resolutions register that numbers itself. None of it makes it onto a trade-fair banner, yet together it is worth considerably more.

There is also an order of steps that is not up for negotiation. AI works on structured data and can do nothing with a filing cabinet. Starting with the AI module therefore means buying a feature that cannot run on your data. Where to start instead is in the five-step roadmap.