Property valuation was one of the first professions to use statistics at scale, long before anyone spoke of artificial intelligence. Today, automated valuation models and tools that can read hundreds of pages of documents promise to speed up analysis further. They are genuinely useful, provided one knows exactly where their reliability ends.
Hedonic models: a Swiss tradition
In Switzerland, automated valuation relies mainly on the hedonic method. The idea is to break a property down into characteristics (floor area, volume, year of construction, condition, finish, location, noise, access to public transport and so on) and to estimate statistically, from recent sales, how much each contributes to the price. A new property is then valued by adding up the value of its attributes.
Several players in the Swiss market offer such models. IAZI, co-founded in Zurich in 1994, helped introduce the method to the country; banks, insurers and pension funds use it to appraise properties before granting a mortgage. Wüest Partner developed the model used by UBS together with the bank in 1999. Fahrländer Partner (FPRE) also provides hedonic valuations for residential and commercial property. More recently, firms such as PriceHubble, founded in Zurich in 2016, have built on machine learning and richer location data.
According to UBS’s own description, its model draws on more than 20,000 arm’s-length transactions a year, which the bank estimates at about half of Swiss property sales, and on around 70 criteria. The bank itself states that the estimate may differ from the actual market value by up to 20%.
Where the models fall short
The limits of hedonic models are well known, and providers acknowledge them:
- Unusual properties. A prestige villa, a listed building, a house on a very large plot or a building with an unusual use has few comparables. The model extrapolates, and the error grows.
- Scarce or dated data. In a thin or illiquid local market, recent transactions are lacking; a model fed with old data describes yesterday’s market.
- Hidden features. An awkward easement, an exceptional view, the true state of wear or the potential for additional floor space do not always appear among the variables.
- Built for standard housing. These models are designed for condominium flats and detached houses. For an income-producing building, value depends mostly on rents, leases and costs, which calls for an income capitalisation or discounted cash flow approach.
As an example, take a flat valued by a model at CHF 1.2 million: a 10% margin of uncertainty already represents CHF 120,000, enough to change how a purchase is financed.
Generative AI and document due diligence
The other front is document review. An acquisition involves land register extracts, leases, condominium regulations, technical reports and building permits. Language models can now summarise these documents, extract lease expiry dates, indexation clauses or easements, and flag inconsistencies between them.
The Swiss land register shows what is at stake. Under Article 970 of the Civil Code, anyone may obtain, without showing an interest, a property’s description, the owner’s name, the form of ownership, the acquisition date, and the easements and land charges; anything further requires a legitimate interest. No one can claim to be unaware of an entry in the register. Misreading an easement is therefore not a detail but a legal risk borne by the buyer.
Hallucinations and confidentiality: the two main risks
Language models can produce answers that sound plausible but are wrong. In June 2023 a federal court in New York fined lawyers USD 5,000 for filing a brief that cited case law invented by ChatGPT (Mata v. Avianca). Specialist tools are not immune either: a 2024 study by Stanford’s RegLab found that leading legal research tools produced such errors in roughly 17% to 33% of the queries tested, fewer than general-purpose models but far more than their vendors’ claims suggested.
The second risk is confidentiality. An acquisition file contains personal data (tenants, sellers) and commercially sensitive information. Before handing it to an online service, one needs to know where the data is processed, whether it is used to train the model, and whether the transfer complies with data protection law.
Human judgement as the keystone
The profession is starting to formalise these safeguards. RICS, the international professional body for surveyors and valuers, brought a standard on the responsible use of AI into effect on 9 March 2026. It confirms that the professional remains accountable for their advice whatever tools are used, and requires, among other things, internal policies, due diligence when selecting tools, and a risk register.
Our view: AI is an excellent first filter. It frees up time for what matters: the issues it flags and those it cannot see. But an automated estimate is not a valuation report, and a lease summary is not a reading of the lease.
Key takeaways
Hedonic models are well established in Switzerland for standard housing, with a margin of error that their own providers put at up to 20%. They become less reliable for unusual properties, illiquid markets and income-producing buildings. Generative AI speeds up document due diligence, but hallucinations and confidentiality concerns call for systematic checking against the sources. Final responsibility for a valuation remains human, something professional standards are now starting to spell out.
Sources
This article is for information only and is based on the sources cited as at the date of publication. It is neither legal or tax advice nor an investment recommendation.