AI and valuation
AI in Real Estate Valuation: What Meaningful Human Review Actually Requires
Human review adds value only when the valuer can test the evidence, challenge the method and take responsibility for the resulting opinion.
AI is entering valuation through several routes at once. Automated valuation models estimate values from transaction and property data. Other tools retrieve comparables, extract lease terms, identify market evidence, suggest adjustments, analyse images or draft parts of a report. Some of these uses are conspicuous. Others are embedded inside software that valuers already use.
This matters because “human review” can describe very different things. It might mean checking that a report has been completed, comparing a model’s estimate with an expected range, investigating the evidence behind it, or forming an independent professional judgment. Those activities are not equivalent.
The relevant question is therefore not simply whether a person remains in the process. It is whether the review is capable of identifying a material problem, understanding its significance and changing the result.
Begin with the purpose of the valuation
A review cannot be designed sensibly without knowing what the valuation will be used for. A preliminary portfolio screen, a secured-lending valuation, an acquisition appraisal and expert evidence in a dispute expose different parties to different consequences. The degree of scrutiny should reflect the intended use, the asset, the available evidence and the effect of error.
The same principle applies to the AI contribution. A tool that formats a table creates a different risk from one that selects comparables, estimates a yield or recommends a value. Review should concentrate on the points at which the system can influence the professional conclusion.
RICS now distinguishes incidental uses of AI from uses that have a material impact on surveying services. Its professional standard on the responsible use of AI requires an appropriately qualified, named surveyor to make or supervise a written decision about the reliability of materially significant AI output. This is a useful starting point: identify the output that matters, the decision it informs and the person who is responsible for deciding whether it can be relied upon.
That decision should be more than a statement that the output appears reasonable. It should be supported by evidence appropriate to the role the output plays in the valuation.
Review the evidence before reviewing the answer
A plausible valuation can be produced from weak evidence. The first task is therefore to examine the information entering the process.
For an AI-assisted valuation, this may include transaction records, asking prices, lease data, property characteristics, planning information, geospatial data, market indices and observations from an inspection. The reviewer needs to know where material inputs came from, when they were obtained and what transformations have been applied to them.
Basic questions remain powerful. Does the subject property match its recorded characteristics? Are the transactions genuine, sufficiently recent and comparable? Have incentives, unusual terms or related-party circumstances been identified? Is the model treating an asking price as though it were an achieved transaction? Has an extracted lease term been checked against the source document?
Completeness matters as much as correctness. A tool may reproduce the evidence it has been given while omitting a feature that changes how that evidence should be interpreted. Condition, configuration, covenant strength, lease events, development constraints and the depth of the relevant market can be difficult to represent consistently at scale. Missing information should not become invisible merely because the output is numerically precise.
Data provenance also affects whether the work can later be explained. A reviewer who cannot trace a material figure to its source is poorly placed to defend the resulting judgment.
Challenge the method, not only the number
Checking whether an estimate falls within an expected range is useful, but it is not a sufficient review of the method that produced it.
The reviewer should understand, at a level proportionate to the task, what the tool is doing. Which variables influence the result? How were comparable properties or market observations selected? What assumptions fill gaps in the data? What period and market conditions informed the model? How does it behave when the subject property differs from the cases on which it performs best?
A confidence score should be treated as one item of evidence rather than proof of reliability. The score may describe statistical dispersion, the density of nearby evidence or the model’s own internal calculation. It does not necessarily establish that the data are correct, that the market has remained stable or that the chosen method is suitable for the instruction.
Sensitivity testing can make the reasoning more visible. What happens if a rental assumption, yield, void period, exit value or development cost changes? Does a small change to wording or property characteristics create an unexpectedly large change in the AI output? Does the result remain stable for economically similar cases while responding when a material fact changes?
This distinction between stability and responsiveness is central to my research. Our study of reliability and sensitivity in LLM risk selection found that repeatability alone can be misleading: an output may appear stable because the method is insufficiently responsive to meaningful changes. Valuation teams should similarly avoid equating consistency with correctness.
Look deliberately for boundary cases
AI systems tend to perform best where relevant examples are plentiful and the task resembles the conditions in which the system was developed. Real estate frequently contains the opposite: heterogeneous assets, sparse transactions, changing market conditions and property-specific rights or constraints.
Meaningful review should therefore search for reasons why a case may fall outside the system’s dependable range. An unusual building, a thin local market, a complex lease, a significant refurbishment, a rapidly changing use class or a transaction during a market dislocation may all reduce the relevance of patterns learned elsewhere.
Aggregate performance can conceal these weaknesses. A model that performs acceptably across a large residential portfolio may still produce systematic errors for a particular geography, property type or value band. Testing should be segmented accordingly, and performance should be monitored after deployment as markets and data change.
Bias deserves the same practical treatment. It is not enough to state that a model should be unbiased. Firms need to examine whether errors are concentrated in particular groups of properties or locations, whether historical data encode persistent distortions and whether those distortions could influence lending, investment or client outcomes. Where valuation contributes to decisions about individuals, separate data-protection and fairness obligations may also arise.
Give the reviewer the conditions needed to disagree
A person cannot provide effective oversight merely by being placed at the end of a workflow. The reviewer needs suitable expertise, adequate time, access to the underlying evidence and genuine authority to challenge or override the system.
The design of the workflow matters. If a valuer sees an apparently authoritative estimate before examining the evidence, it can become an anchor for the subsequent analysis. Requiring a reason for an override may discourage challenge if accepting the output is effortless. A high-volume process may turn review into a target-driven confirmation exercise.
Possible safeguards include asking the valuer to form an initial range before seeing the model output, highlighting rather than hiding conflicting evidence, requiring targeted review of low-confidence or out-of-distribution cases, and testing whether reviewers actually detect known errors. None is universally appropriate. The point is to evaluate the behaviour of the combined process, not to assume that adding a person cures weaknesses in the model.
The Information Commissioner’s Office uses the term meaningful human review in a particular data-protection context and warns against treating a human “rubber stamp” as meaningful intervention. That legal context will not apply to every valuation. The operational insight is broader: review must involve interpretation and discretion, not automatic acceptance.
Preserve the reasoning and the intervention
A reliable process needs a record of how the conclusion was reached. Depending on materiality, this could include the tool and version used, material inputs and their sources, the output generated, confidence or warning indicators, the checks performed, adjustments or overrides, and the reasons for the final judgment.
The objective is not to retain every intermediate screen indefinitely. It is to preserve enough evidence to explain the contribution of AI and the professional work applied to it. This becomes particularly important where tools change over time or outputs are held in third-party systems.
Records of disagreement are valuable. If a valuer rejects an AI-selected comparable or changes a model-derived assumption, that intervention may reveal a recurrent weakness. Analysed collectively, overrides and exceptions can improve validation, training and procurement decisions. They should not be treated simply as friction to be eliminated.
Our research on knowledge-generated test suites takes a related approach: expected targets are derived from structured domain knowledge before an LLM component is tested, and human reviewers certify the test stimuli rather than asking another model to judge the result. The wider principle is that assurance evidence should not depend entirely on the system being assessed.
Keep accountability with the professional judgment
An automated estimate and a professional valuation are not the same thing. RICS explains that an AVM output in isolation is not an opinion of value under Red Book Global Standards. Once a member adopts, reviews and reports that output as a professional opinion, the valuation must comply with the applicable standards. The International Valuation Standards Council likewise retains the need for professional judgment when models are used.
Technology can extend the evidence available to a valuer and make parts of the process faster and more consistent. It does not remove the need to decide which evidence matters, whether the method fits the case, how uncertainty should be represented and what conclusion is supportable for the stated purpose.
Meaningful human review is therefore not a ceremonial control added after automation. It is a substantive part of the valuation method. Its quality should be tested in the same way as other important components of the workflow: by examining what the reviewer can see, what they are expected to do, which errors they detect, when they disagree and whether the final decision is better as a result.