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 are a little right, and both positions cost money — one by buying too much, the other by leaving real relief on the table.

This article aims at the sober middle: what works today, what does not, and how to tell the difference before you have paid for it.

The one distinction that explains everything

There is exactly one dividing line, and once you know it you can place any feature in thirty seconds:

AI is good at reading something off or phrasing something. It is bad at deciding something that depends on context not present in the document.

Reading off: this invoice has an amount on it — which? Phrasing: turn these notes into a polite message in the formal register. In both, the answer is already contained in the material.

Deciding: is this contractor invoice maintenance or repair? That answer is not on the document. It is in a resolution, in an agreement, or it follows from a history only you know. A model answering here is guessing — plausibly, fluently, and with exactly the same confidence it shows when it is right.

Everything below is just that rule applied.

What genuinely carries today

Reading invoices. The clearest case. Supplier, invoice number, amount, date — four fields printed on the document. Saves roughly half a minute per invoice, net of checking. At eight hundred invoices a year that is a working day. In detail: AI invoice capture.

Describing documents. Proposing a title, a short summary and metadata for an uploaded PDF. That is reading off. The value is not in the suggestion itself but in the filing actually getting maintained — documents whose naming costs thirty seconds get named "later", and then never.

Drafting text. The reply to an owner writing about the same defect for the fourth time. You know what has to be in it; the work is the phrasing, and phrasing is unpleasant when you are irritated. A draft you revise beats an empty field.

Pre-sorting what comes in. Proposing a category and urgency for a new defect report, and finding possible duplicates. The duplicate is the real value: three owners report the same broken garage door in three different wordings, and a full-text search will not find that.

Answering questions from documents. An owner asks what was resolved about the façade. The answer is in minutes they are entitled to read. That is a search task with phrasing — provided the system answers only from the documents that owner may see, cites the source, and says honestly "I don't know" when the answer is not in the material.

What does not carry

Fully automatic booking. See above: the cost type is a decision, the distribution key hangs off it, and who pays hangs off the key. The error surfaces in the annual statement, in front of thirty owners.

Automatically sent replies. Technically trivial, practically a bad idea. The one message that goes to the wrong owner in the wrong tone costs more time than a hundred saved drafts — and you hear about it first from the advisory board.

"AI produces your annual statement." An annual statement is an arithmetic task with statutory requirements, not a writing task. What makes it dependable is correct apportionment logic, cent-exact rounding and immutability after approval — all things you build deterministically because they have to be deterministic. AI is not better than arithmetic here; it is worse.

"AI checks your resolutions for legal soundness." That is legal advice. A model can name the typical formal defects, which is useful as a prompt. It cannot tell you whether this resolution is challengeable in this community, and a vendor implying otherwise is selling you a risk.

Forecasts from your data. "AI predicts your maintenance needs." For a practice with forty buildings there is no data basis for that. Statistics need cases, and your cases number in the tens.

The question that makes the difference

For every AI feature demonstrated to you, ask one thing:

What happens without my confirmation?

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

If the answer is anything else, your liability position has changed without it being written down anywhere. You remain responsible for the booking, the statement and the message; you simply did not make them yourself any more and can no longer say exactly why they look the way they do.

Which is why our own pattern across every AI feature is the same: propose, never execute. Invoice capture reads out, you book. Triage proposes, you confirm. The draft sits there, you send it. That is not a technical limitation but the position that a manager has to be able to stand behind their own acts.

Data protection, short and practical

Three things to settle before an AI feature touches real data:

Where is it processed? Defect reports, documents and enquiries contain personal data. Processing outside the EU needs a solid legal basis; inside is simpler.

Is there 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.

Is content used for training? The only acceptable answer is no, and it belongs in the contract, not on a marketing page.

Ours is in the privacy notice: processing via a processor in Germany, exclusively on servers in the EU, an Art. 28 GDPR agreement, no use for training. Activation is the management company's decision, not ours.

What it realistically delivers

For a practice with thirty to fifty buildings, once the data is already digital:

Together: a short working day per month. Respectable, but not the place where your practice changes.

The larger effects are elsewhere — in the owner portal that intercepts phone calls, in document storage that ends searches, in a register of resolutions that numbers itself. Unspectacular, no banner material, and worth considerably more in total.

And the order is not negotiable: AI works on structured data. Pointed at a filing cabinet it can do nothing. Starting with the AI module means buying a feature that cannot run on your data.

Where to start instead: the honest roadmap.