Writing & perspectives

Real estate workflows

From Property Data to Professional Judgment: Where AI Changes the Real Estate Workflow

AI changes the route from property evidence to professional advice. The critical question is where judgment, review and accountability sit in the redesigned workflow.

AI is often discussed as though it were a new source of real estate expertise. In practice, its more immediate effect is to alter the route by which evidence becomes analysis and analysis becomes advice.

A valuation, development appraisal, lending recommendation or investment paper is not a single task. It is a workflow: instructions are framed, information is collected, evidence is tested, assumptions are selected, calculations are performed, conclusions are challenged and someone accepts responsibility for the final judgment. AI can contribute at several of those points. Its value, and its risk, depend on which point it enters, what the next person receives and what the eventual decision requires.

A lease summary may be fluent but omit the clause that changes the cash flow. A valuation estimate may be statistically defensible for a common property type but unreliable for an unusual asset. A development model may calculate every scenario correctly while leaving the critical planning assumption unchallenged.

The useful question is therefore not simply whether AI was used. It is where the workflow changed, what judgment moved with that change and what evidence remains available to the professional who must rely on the result.

Begin with the decision, not the tool

The same AI capability can have very different consequences in different workflows. Extracting lease dates for an internal screening exercise is not equivalent to producing inputs for a lending decision. Generating a first draft of a market commentary is not equivalent to selecting the yield on which an acquisition recommendation depends.

I would begin by defining the work in decision terms. What will the output inform? Who will use it? What would a material error look like? Which assumptions are matters of professional judgment rather than data processing? The answers determine the evidence, review and escalation the workflow needs.

This approach is consistent with the RICS professional standard on the responsible use of AI. It focuses on AI that has a material impact on the delivery of surveying services and requires professional judgment to be applied to decisions about the reliability of material AI outputs. The standard treats judgment as more than a final glance: it includes knowledge, skill, reasoning, experience and professional scepticism.

Human involvement is not automatically an effective control. A meta-analysis of 106 experimental studies found that human–AI combinations did not, on average, outperform the better of the human or AI acting alone, with losses particularly evident in decision tasks. The design of the combination has to be tested rather than assumed to be beneficial.

Data work: greater reach, new questions of provenance

Real estate workflows draw on dispersed and uneven information: transaction evidence, leases, planning records, title documents, technical reports, operating data, market research and local knowledge. AI can make this evidence easier to search, classify and compare. It may identify clauses across a document set, standardise descriptions, match transactions to a subject property or flag missing fields before analysis begins.

This changes the amount of information a professional can examine. But extraction does not settle whether the source is authoritative, current or relevant. Nor does it establish that the absence of a result means the underlying fact is absent.

A dependable handoff from data collection to analysis should therefore preserve provenance. The reviewer needs to know which document supported an extracted fact, which version was used, what was unreadable and what the system did not find. Where information has been transformed, the original should remain available. Exceptions should be visible rather than quietly converted into blanks or averages.

RICS valuation standards place established importance on appropriate data, records, investigations and the nature and sources of information relied upon. AI changes how those obligations may be discharged; it does not remove the need to understand the evidence. A faster evidence pipeline is useful only if the professional can still trace a material input back to its source and challenge its relevance.

Valuation: assistance is not an opinion of value

Valuation provides a clear illustration of the different layers in a workflow. AI may retrieve comparable transactions, adjust data into a common format, identify unusual observations, estimate a range or draft parts of a report. An automated valuation model may also provide a useful benchmark for a sufficiently represented market segment.

The opinion of value, however, depends on more than producing a number. The valuer must consider whether the evidence reflects the asset, valuation date, market and agreed basis of value. An apparently close comparable may contain a non-standard lease, an unusual purchaser or a physical constraint that is not represented in the data. An inspection may reveal a condition or configuration that a model has not observed.

The current RICS consultation draft on AI in real estate valuation places AI within the valuation process while retaining clear professional boundaries. It is draft guidance rather than an operative standard. Its practical direction is nevertheless useful: the important control is not a generic approval step after the model has run. It is a review designed around the ways the particular method can fail, including data coverage, outliers, selected comparables, material extractions and the suitability of the output for the instruction.

My research on testing LLM components from structured knowledge approaches a related problem: failures should be tested against defined expectations rather than judged only by whether an output appears reasonable. In valuation, that suggests test cases derived from the work itself, including unusual leases, sparse markets, contradictory evidence and assets beyond the model’s intended scope.

Development appraisal: calculation and judgment are different contributions

A development appraisal brings together land, programme, construction cost, finance, sales or rental evidence, planning obligations and risk. AI can assist with document extraction, scenario production, sensitivity analysis and consistency checks. It may help compare versions of a cost plan, identify movement in assumptions or show which variables drive the result.

Those capabilities can improve the analytical work without determining whether the project is attractive. The selection of a programme, the interpretation of planning risk, the credibility of an exit value and the allowance for abnormal costs remain contextual judgments. Generating more scenarios is not the same as selecting a defensible base case.

The workflow should separate calculation from assumption ownership. Each material assumption needs a source, an owner and a reason for its use. AI-generated scenarios should show what changed and what remained fixed. Escalation should be triggered where an input falls outside an agreed range, conflicts with another source or materially changes the decision. This allows AI to widen the analysis without disguising responsibility for the assumptions that drive it.

Lending and underwriting: models within a chain of reliance

In lending, property analysis connects to credit policy, borrower information, cash-flow assessment, covenant structure and collateral. AI may support document review, property segmentation, fraud detection, cash-flow modelling or collateral monitoring. A model output can therefore influence several downstream decisions even when it is not presented as the final recommendation.

This creates a dependency problem. If an extracted rent roll feeds a valuation, which feeds a loan-to-value calculation, which feeds an approval threshold, an early error can acquire the appearance of precision as it moves through the process. Controls should follow that chain rather than examine each component in isolation.

The PRA’s model-risk principles for banks provide a useful parallel. They emphasise model identification, governance, development and validation, independent review and mitigating actions. The underlying lesson extends beyond regulated models: organisations need to know what models they rely on, their intended uses, their limitations, their owners and their material dependencies.

Professional review is particularly important around overrides and exceptions. An override may represent valuable expert knowledge, or it may conceal a systematic weakness. The reason, effect and approval should be recorded. Repeated exceptions can then become evidence about where the model or workflow needs to change.

Investment analysis and due diligence: omissions matter as much as errors

Investment and due-diligence teams often face a different problem: too much information under time pressure. AI can search data rooms, summarise reports, compare assumptions and draft issue lists or investment papers. These are attractive uses because they may release time for interpretation and challenge.

Yet the central risk is not always an incorrect statement. It may be an omitted issue, an untested relationship between documents or a weak assumption presented as settled. A concise summary can be accurate in what it includes and still be unsafe to rely upon.

A good review process should therefore test coverage as well as correctness. What document classes were expected? Which were missing? Were adverse findings carried into the financial analysis? Did a planning condition affect the programme? Did the lease interpretation flow into the cash-flow model? Did the investment committee receive the uncertainty as well as the central case?

This is where the handoff between specialists matters. AI may help connect evidence across documents, but a professional must decide whether the connection is material and what it means for the proposed decision. The workflow needs a route for ambiguity to be escalated rather than compressed into a confident answer.

Design the workflow around evidence and challenge

Across these applications, I would look for five features.

First, the purpose and boundary of each AI contribution should be explicit. Second, material inputs and outputs should remain traceable to evidence. Third, review should be matched to known failure modes, not reduced to a generic human approval. Fourth, exceptions, overrides and disagreements should be recorded and used to improve the process. Fifth, accountability should be attached to named decisions: who judged the output reliable, who selected the assumption and who approved reliance on the final work.

The aim is not to preserve every traditional step. Some activities may be automated, combined or removed. But when a workflow changes, the organisation should be able to explain what has improved and demonstrate that the resulting work remains dependable for its intended use.

AI can expand the evidence considered, accelerate analysis and make alternative scenarios easier to examine. Professional value arises when those capabilities are organised around a sound decision. The real opportunity is therefore not automation followed by a signature. It is a better-designed relationship between data, models, evidence and professional judgment.