AI in property reports: when it genuinely helps and when it makes up the facts

In 2026, putting ‘AI-powered’ on the cover of almost any product boosts sales. Property reports are no exception: AI analysis, AI valuation, AI risk detection. We work with this technology every day, which is why we can tell you what the marketing leaves out: the important question is not whether a report uses AI, but what it uses it for. One use adds real value; another can introduce fabricated information into the biggest financial decision of your life.

What is a ‘hallucination’, and why does it matter here?

Generative language models do not consult reality: they generate the most plausible text in response to a prompt. When they know the answer, they get it right; when they do not, they produce something that sounds right, complete with specific figures, case numbers, dates and flawless technical vocabulary. This is known as a hallucination. Its most dangerous feature is that it cannot be distinguished at a glance from a real fact: falsehood is written just as convincingly as truth. In a casual chat, that may be a curiosity. In a property purchase report, where a ‘€45,000 special assessment’ or ‘cracks in the façade’ could shape a six-figure negotiation, it is a serious problem.

Two architectures: the difference you should insist on knowing

Put simply, an automated report can be built in two ways. In a deterministic architecture, every item of data comes from a direct query to an official source: intersecting the plot with MITECO's flood layer returns ‘affected’ or ‘not affected’, and that result remains the same if the check is run a thousand times. If AI is involved, it comes later, to draft, explain or translate the technical language. The data and the wording are separate, and the data is auditable.

In a generative architecture, the model helps produce the content: it interprets a façade photograph, ‘infers’ building characteristics and fills sections for which no searchable source exists. The report appears more complete because nothing prevents it from doing so. Where a deterministic system would honestly say ‘data unavailable’, a generative one may write three convincing paragraphs. Paradoxically, apparent completeness is the warning sign.

How to spot generated data in a report: four clues

• Data with no source you can check. If the report makes claims found in no public register, such as the structural system, the condition of waste pipes or owners' association resolutions, ask where they came from. ‘Our model estimates it’ is a legitimate answer only if presented as an estimate, not as fact.
• Suspicious precision. Case numbers, meeting dates and exact amounts in sections where no public source exists. Hallucinations tend towards detail: a specific invention is more convincing than a vague one.
• An overly coherent narrative. If different sections elegantly ‘confirm’ one another like a novel, with the photograph showing exactly what the document predicted, be wary. Real data from independent sources usually joins up imperfectly, not like a script.
• The familiar-property test. This is the decisive one: request a report for a home you know in detail, such as your own or your parents', and count how many specific claims are verifiably true. Five minutes of testing is worth more than any marketing page.

Where AI does add value — a great deal of it

This is not an argument against technology; that would be absurd coming from a technology platform. AI excels downstream of the data: explaining a T100 return period in plain English, summarising a long document you provide, translating a report for an overseas buyer or answering questions about verified results. In each case, the model works from facts established by other means, containing the risk of hallucination. The principle we follow at Hogarfax, and recommend you demand from any service, fits in one sentence: AI may explain the data; it must not be the source of the data.

The one question that sums it all up

For any ‘AI-powered’ report, ask one question: did this specific item come from a source or from an inference? A reputable service will answer item by item because its architecture separates retrieved information from generated text, and because every claim is accompanied by its source and date. A service that does not distinguish between the two is really asking you to take a leap of faith. Decisions involving hundreds of thousands of euros are not made on faith; they are made on sources.