AI and investment
Can AI Improve Real Estate Investment Decisions—or Merely Accelerate the Analysis?
The value of AI in real estate investment depends on whether it improves decision quality, not merely how quickly it produces the first analysis.
AI can shorten parts of real estate investment analysis. It can retrieve lease information, compare transactions, summarise due-diligence material, populate models, produce scenarios and draft investment papers. Yet a faster analysis is not necessarily a better investment decision.
Real estate decisions combine incomplete evidence, forward-looking assumptions and professional judgment. A model may produce an answer quickly while leaving the investment committee no better able to judge whether the rent, exit yield, capital expenditure or financing assumptions are defensible. It may generate more scenarios without identifying which uncertainty matters, or increase the work required to verify them.
My interest is therefore not whether AI can participate in real estate investment analysis. It plainly can. The more useful question is whether the human–AI workflow produces a better decision than the process it replaces, once quality, uncertainty, review effort and accountability are considered together.
Start with the decision, not the demonstration
An impressive demonstration often begins with an available capability: a model can read documents, extract market evidence or generate a discounted cash-flow analysis. An investment case should begin elsewhere—with the decision and the friction that makes it difficult.
Is the problem that analysts spend too long assembling evidence? That relevant risks are identified inconsistently? That assumptions are difficult to compare across assets? That an investment committee receives too much material but too little explanation of what could change the recommendation? Each problem implies a different use for AI and a different test of value.
This matters because “better analysis” is not a single outcome. AI might improve evidence coverage, consistency, speed or the ability to detect an unusual case. It could also improve one dimension while weakening another. A system that accelerates underwriting but anchors reviewers to its first estimate may save time and degrade independent challenge.
The unit of evaluation should be the workflow that produces and governs the recommendation: data acquisition, document review, modelling, professional interpretation, challenge, approval and subsequent monitoring. Evaluating the model alone cannot establish whether that combined process supports better capital allocation.
Define what “better” means before measuring it
A useful pilot needs an explicit account of the improvement being sought. Depending on the decision, that could include:
- more accurate extraction of lease, planning or technical information;
- better-calibrated estimates rather than greater apparent precision;
- earlier identification of material downside cases;
- more consistent comparison of assets and assumptions;
- wider but still decision-relevant scenario coverage;
- clearer traceability from evidence to assumption and conclusion;
- reduced total effort to reach an investment-ready recommendation; or
- faster decisions where timing itself has commercial value.
These measures are not interchangeable. A reduction in drafting time does not establish improved judgment. A more accurate point forecast does not necessarily improve an acquisition decision if the model fails precisely where pricing is most uncertain. A longer risk list may look comprehensive while making it harder to distinguish the risks that could change the decision.
Recent peer-reviewed real estate research illustrates the importance of this distinction. A study using more than 50,000 Oslo housing transactions found that uncertainty information could help buyers avoid the adverse-selection problem created by acting on automated valuation point estimates alone. The benefit came not simply from producing a prediction, but from identifying the cases in which reliance was less defensible. That is a useful principle for investment workflows: knowing when not to trust an answer may be as valuable as improving the average answer.
Account for the whole cost of the workflow
The headline efficiency gain is usually measured at the first output. If a task that took four hours can be drafted in one, the saving appears obvious. But an investment-ready output also requires checking, correction, explanation and approval.
Suppose AI reduces an analyst’s initial work by three hours but adds ninety minutes of verification and thirty minutes of senior review. The gross saving and the usable saving are different. If the senior reviewer is already the constraint, producing more first drafts may increase the queue rather than expand investment capacity. If analysts use the released time to examine better alternatives or challenge assumptions, the same nominal saving may create considerably more value.
The economics should include data preparation, integration, model or vendor costs, validation, exception handling and monitoring. The Bank of England and Financial Conduct Authority’s 2026 survey of AI in financial services is instructive. Firms see analytical and efficiency benefits, yet identify data quality, third-party dependency, model complexity and integration as material concerns. Pre-deployment controls were far more common than continuing monitoring and periodic review. That gap matters as markets, data and working practices move.
For a real estate investor, the commercial test is not “How much faster did the model produce the paper?” It is “What decision-relevant work became possible, what additional review did the method require, and did the resulting process improve the use of capital?”
Treat uncertainty as information, not an inconvenience
Real estate investment analysis is built on variables that cannot be known with precision. Leasing velocity, incentives, operating costs, construction prices, financing terms and exit yields are not merely missing inputs waiting to be discovered. Some are inherently uncertain; others are affected by structural change or by sparse, non-comparable evidence.
AI can help explore those uncertainties. It can test combinations more quickly, identify patterns across a larger evidence base and reveal which assumptions drive the result. But scenario volume is not the same as scenario quality. Generating hundreds of variations adds little if they inherit the same questionable evidence, omit a material mechanism or present improbable combinations with equal prominence.
A defensible workflow should distinguish uncertainty in the evidence, sensitivity of the result to an assumption, and confidence that the model remains appropriate for the asset and market. It should preserve ranges and disagreements rather than collapse them into a confident narrative.
This is also where professional knowledge remains central. An experienced investor may recognise that a comparable is technically similar but exposed to a different occupier market, or that a modelled refurbishment programme is inconsistent with the sequencing of the work. The human contribution should not be described vaguely as “oversight.” It should be specified: challenging comparability, testing causal plausibility, choosing decision thresholds, interpreting an exception or deciding whether the residual uncertainty is acceptable.
Design the investment committee interface deliberately
Investment committees do more than approve a number. They test the proposition, compare it with alternatives and decide which uncertainties the organisation is willing to carry. AI changes that process even when it has no formal vote.
The first output seen by a committee can become an anchor. A polished narrative can create what researchers have called artificial certainty: the impression that a complex future has become more knowable because it has been represented in detail. Conversely, an explanation that exposes the evidence, assumptions and contested steps can make challenge more effective.
The committee therefore needs more than a disclosure that AI was used. It needs information proportionate to the role AI played. That may include which evidence was machine-extracted, which assumptions were generated or challenged by AI, where a professional departed from the system’s output, what tests were applied and which material limitations remain.
RICS’ professional standard on the responsible use of AI reflects this workflow perspective. It addresses data and system governance, output reliability, assurance protocols and client communication while maintaining professional control. NIST’s AI Risk Management Framework similarly calls for intended benefits and costs to be compared with appropriate benchmarks, limits and human oversight to be documented, and systems to be tested before deployment and regularly in operation.
For an investment committee, those principles translate into a practical requirement: the record should make it possible to reconstruct how the recommendation was reached, not merely reproduce the final output.
Run a pilot that can answer the investment question
A useful pilot is bounded enough to learn from. It might test one task—lease abstraction, comparable selection, downside identification or scenario preparation—within one defined decision process. It should compare the proposed human–AI workflow with the current expert process on the same cases and information.
The case set should include difficult and atypical examples, not only clean demonstrations. Measures should cover errors and omissions, total analyst and reviewer effort, traceability, and stability when wording, data or market conditions change. Reviewers should record when AI changes their judgment, when they reject its contribution and when they fail to identify an error.
The pilot should also define stopping and escalation rules. Which outputs require independent re-performance? When does low confidence make the system unsuitable? What change in the model, vendor, data or use case triggers reassessment? Without those rules, a successful experiment can drift into a materially different production workflow without a fresh decision about reliance.
My research on human–AI assessment reinforces the need to validate the combination rather than assume a successful method transfers. Structured reasoning can improve one model and fail to help another. Stability can also mislead when a method is insensitive to meaningful changes. The appropriate response is conditional: specify the model, task, evidence, human role and benchmark, then test the workflow that will be used.
The opportunity is better judgment, not simply cheaper production
AI’s most valuable contribution to real estate investment may not be automating the investment paper. It may be improving the quality of the questions asked: exposing an assumption that drives the return, locating contradictory evidence, identifying a case outside the model’s competence or enabling the committee to examine a credible alternative before capital is committed.
That opportunity is larger than a productivity calculation, but it is also harder to establish. It requires evidence that the combined process makes better use of information and professional judgment. It requires uncertainty to remain visible, review to be meaningful and accountability to stay with the people responsible for the decision.
Speed is valuable when it releases scarce expertise or allows an organisation to act in time. It is not a substitute for decision quality. The appropriate ambition is a workflow in which AI makes analysis more useful, experts make their contribution more explicit, and the investment committee has better grounds for deciding when reliance is justified.