Real estate AI assurance
How Should AI-Assisted Real Estate Decisions Be Assured?
Assurance should establish whether the whole decision workflow is dependable, not merely whether an AI model performs well in isolation.
Artificial intelligence is beginning to contribute to real estate work at several points: retrieving evidence, classifying documents, selecting comparables, forecasting cash flows, identifying risks, drafting analysis and challenging assumptions. The important question is not simply whether these tools are accurate. It is whether the decision process in which they operate is dependable for its intended purpose.
That distinction matters. An accurate model can be embedded in a poor workflow. Relevant evidence may never reach it. A plausible output may be given more weight than it deserves. A nominal human review may confirm rather than challenge the machine’s framing. Material changes in data, models or market conditions may go unnoticed after deployment.
For real estate firms, AI assurance should therefore address the combined workflow: data, model, professional, review and decision. It should provide evidence that the arrangement works under the conditions in which people will actually rely on it.
Start with the decision, not the tool
The same AI output can carry very different significance in different settings. A model-generated rent range used to explore scenarios is not equivalent to a figure incorporated into a secured-lending valuation. A document summary used to orient an experienced reviewer is not equivalent to an extraction that determines whether a lease obligation enters an investment model.
The first assurance question is consequently: what decision does the workflow support?
That requires a clear account of the purpose, intended users, materiality and consequences of error. It should identify which outputs are informative, which affect the professional conclusion and which could cause an action to be taken. It should also establish what the AI is not being used to decide.
This is consistent with the risk-based approach in the NIST AI Risk Management Framework. Its mapping function asks organisations to define context, intended use, human oversight and the possible impacts of the system. RICS takes a similarly practical approach. Its professional standard on responsible AI applies where AI may have a material impact on surveying services, rather than treating every use of software as equivalent.
Clear scoping prevents assurance from becoming a generic technology checklist. It allows controls and evidence to be proportionate to the decision.
Map how judgment is produced
Professional decisions rarely emerge from one model output. Consider an investment appraisal. Data may be extracted from leases, compared with market evidence, transformed into cash flows, subjected to assumptions about voids and expenditure, and interpreted against an investor’s objectives. AI could contribute to any of these stages. A professional might review some outputs, amend others and make the final recommendation.
Assurance needs to make those contributions visible. A useful workflow map identifies:
- the source and quality of the evidence entering each stage
- the task performed by AI and the output it produces
- where an AI output becomes an input to later analysis
- the professional judgments required
- the checks, challenge and escalation points
- the person accountable for the resulting decision.
This often reveals risks that a model-level assessment will miss. An extraction tool may perform well overall but consistently omit unusual break clauses. A forecasting model may be technically valid but receive inputs whose definitions vary between teams. A reviewer may see an attractive headline figure without being shown the assumptions or contradictory evidence that produced it.
The object of assurance is not only the component. It is the chain through which evidence becomes judgment.
Test performance under decision-relevant conditions
A broad claim that an AI system is accurate is not enough. The evaluation should reproduce the task, data and operating conditions that matter in practice.
For a real estate workflow, this may require test cases spanning property types, locations, document quality, market states and levels of ambiguity. Difficult and unusual cases are particularly important because these are often where professional expertise contributes most. Performance should be separated by meaningful categories rather than compressed into one average.
The test also needs a defensible reference. Some outputs, such as a date or floor area in a document, can be checked against a source. Others involve professional interpretation and may not have one indisputable answer. In those cases, assurance should record how the expert reference was produced, the extent of disagreement and the limits of the comparison.
Relevant measures depend on the role of the system. They might include factual accuracy, omissions, calibration of confidence, sensitivity to changed assumptions, repeatability, robustness to document variation and the frequency with which material errors survive review. Total effort matters too. A system that creates a faster first draft but requires complete re-performance may not improve the workflow.
NIST expressly links measurement to deployment context and calls for validation, uncertainty, benchmarks and continuing monitoring. My research on model-dependent reasoning effects reinforces that discipline: a method that helps one model on a defined task may not transfer to another. In real estate, a model validated on one market, asset class or task may not generalise to another.
Make human review meaningful
“Human in the loop” is not evidence of assurance. The relevant question is what the person is expected and equipped to do.
Review can involve verifying evidence, challenging assumptions, interpreting uncertainty, identifying missing considerations or deciding whether an output is suitable for a particular purpose. Each activity requires different information. A reviewer cannot verify a claim without access to its source. They cannot challenge a modelled conclusion if the important assumptions and limitations are hidden. Nor can they reliably detect omissions from a fluent summary if the underlying evidence is difficult to inspect.
The sequence of the workflow also matters. Showing an AI recommendation before a professional forms an initial view can anchor the subsequent analysis. Requiring an independent assessment may produce a stronger comparison, but it can also remove much of the intended efficiency. These are design choices to test rather than assumptions to make.
Research on human–AI decision-making reinforces the point that complementarity is not automatic. Combining a person and a model does not guarantee performance better than either alone. Assurance should therefore compare the proposed combined workflow with credible alternatives: current professional practice, AI alone where appropriate, and differently designed review arrangements.
The RICS position is especially clear for valuation. Red Book Global Standards retain the need for professional judgment when automated models or AI support the work. The current RICS consultation draft on AI in real estate valuation expresses the practical implication: an AI output is an input to the valuation, not the conclusion. The guidance is still being developed, but the underlying professional boundary is not new.
Preserve evidence of the reasoning
A dependable decision should be capable of reconstruction. This does not mean retaining every transient machine interaction. It means preserving the evidence needed to understand how a material AI contribution affected the result.
Depending on the workflow, the record may need to identify:
- the system and version used
- the material inputs, data sources and instructions
- the relevant output and any expressed uncertainty
- checks performed and exceptions identified
- whether the professional accepted, adjusted or rejected the output
- the reasoning supporting the final conclusion.
This is not separate from professional practice. Red Book already requires records of key inputs, calculations, investigations and analyses. RICS’s consultation draft on AI in real estate valuation applies that principle to material AI-supported work and warns that third-party outputs may later change or become inaccessible.
Traceability supports several purposes at once. It enables quality review, explains a conclusion to the client, supports investigation when something goes wrong and creates evidence for improving the workflow. It also prevents apparent precision or polished language from concealing a weak evidential basis.
Assign accountability to the work
AI can disperse responsibility if organisations focus on ownership of the technology rather than ownership of the decision. Procurement may select the tool, a data team may configure it, professionals may use its outputs and a committee may approve the policy. None of this establishes who is accountable for a particular piece of work.
Accountability should be specified at both levels. Organisational governance should identify who owns the system, data, validation, monitoring, change control and incident response. The assignment-level workflow should identify the professional responsible for determining whether the output can be relied upon in the circumstances.
Third-party tools require particular care. Vendor documentation can support due diligence, but it cannot establish performance in every client’s workflow. The Bank of England and FCA’s 2026 survey of AI in UK financial services found far greater reliance on vendor assurance for third-party systems than deep internal knowledge, while oversight of third-party AI was the most frequently reported governance challenge. This is evidence from financial services rather than real estate, but firms using embedded or externally hosted models face a comparable evidential problem: responsibility cannot be outsourced with the software.
Treat assurance as continuing evidence
A successful pre-deployment test is not a permanent conclusion. Models are updated. Data changes. Market regimes shift. Staff find new uses. The surrounding workflow evolves.
Continuing assurance should define the changes that require reassessment, the indicators that will be monitored and the conditions that trigger restriction, escalation or withdrawal. Useful signals might include changes in error patterns, override rates, missing-data frequency, reviewer effort, drift in inputs, complaints or unexpected concentrations of outputs.
The evidence should remain tied to the stated purpose. If a tool tested for research support begins to influence final recommendations, the original assurance does not automatically follow it into the new role.
A better assurance question
The wrong question is: can the AI be trusted?
Trust is not a fixed property of a tool. The more useful question is whether there is sufficient evidence to rely on a defined human–AI workflow for a defined real estate decision, subject to stated limitations and controls.
That formulation keeps professional judgment central without treating it as a universal safeguard. It recognises that AI can improve the reach, consistency or speed of analysis, while requiring those benefits to be demonstrated in the combined process. And it directs assurance towards the result that matters: not a model that appears impressive, but a decision process that remains evidence-based, reviewable and accountable.