How we stopped our AI from making up property prices

Language models have a well-documented tendency to produce confident answers even when the underlying evidence is weak. In most contexts this is annoying. In property valuation it is a real problem: if someone makes an offer, accepts an offer, or sets a listing price based on a figure an AI invented, the consequences are financial.

This post explains the specific design choices we made to reduce that risk in Home's valuation system.

The core problem

A standard large language model asked "what is this house worth?" will produce an answer. It has no built-in concept of "I do not have enough information to answer reliably." Left unconstrained, it will fill the gap with something plausible-sounding. The result looks like a valuation. It is not.

The same problem exists in simpler algorithmic tools. An automated valuation model trained on sold prices will return a figure even when the training data for that postcode is three years old and covers only a handful of transactions. The output has a number and possibly a confidence band, but the confidence band often understates the real uncertainty.

What we did

Multiple independent models, not one. The valuation council runs GPT-4o-Vision, an algorithmic comparables engine, and a cross-model reconciliation step involving additional models including Gemini and Grok. Each produces an output independently. The spread between outputs is itself informative: when the models agree closely, that is some evidence the question has a clear answer. When they diverge, that is a signal the property is unusual or the data is thin. We go into how this council is put together in inside the multi-LLM valuation council.

Explicit confidence intervals. We do not collapse the output to a single number. The result is always a range, and the width of the range reflects the spread of underlying inputs. A narrow range (say, £300,000 to £320,000) means the models found good comparable evidence and substantial agreement. A wide range (say, £280,000 to £360,000) means the property is harder to pin down and you should treat the estimate accordingly.

Hard floors on data recency. The comparables engine has a minimum data-quality threshold. Below a certain number of recent sales in the area, it flags the estimate as low-confidence rather than proceeding as if the data were adequate. We would rather tell you we are uncertain than tell you a number we are not confident in.

No invented facts. The Q&A side of the system (for buyer questions) is configured to decline rather than guess when a question falls outside the data available. An AI that says "I do not have that information" is more useful than one that generates a plausible-sounding answer to a question about flood risk or planning history.

What this means in practice

You will sometimes find that Home gives you a wider range than a tool that does not acknowledge its own uncertainty. That is intentional. A wide range is honest. A false precision is worse than a useful bracket.

For properties where the data is limited, we recommend using the estimate as one input alongside an agent's view and, where appropriate, a formal survey. See online valuation vs agent valuation for how the two should work together, and what AI can and cannot tell you about the value of your house for the limits worth knowing before you rely on any figure.

Explore price estimates at house price data.

Valuations are estimates only. Home does not guarantee accuracy and these figures should not be relied on as formal appraisals.

Further reading: the ONS UK House Price Index.