Artificial Intelligence (AI) can now deliver results that appear remarkably complete. It can summarise leases, review rent rolls, draft memos for investment committees, and explain changes in occupancy rates in concise language. This creates a new problem for real estate companies: a document can look ready long before its use is safe.
For an asset manager, controller or acquisition expert, the challenge is not whether an answer is easy to read, but whether the person receiving it can defend the work. Where do the figures come from? Were they taken from the current or an older rent roll? Does 'occupancy' mean leased, physically occupied, or economically occupied? Were the calculations independently verified? What happens if the system is uncertain?
These questions are gaining importance as AI moves beyond mere drafting to autonomously perform longer workflows. An example of this is the routine request: 'Send the Monday operations report.' For a human, this instruction involves a great deal of tacit knowledge. The report must use the correct reporting period, the right property list, and accounting data. It may need to flag certain variances but ignore others. Some information might be accessible to the asset management team but withheld from other recipients. The finished document may also require review before it leaves the company.
The Need for Traceability
The request sounds like a single task, but in practice, it is a series of decisions. This distinction is relevant because each individual step can appear plausible, while the final result is flawed. An AI system can retrieve a document, summarise it, insert figures into a table, and prepare an email. However, if it selects an outdated file or sends the report to the wrong distribution list, the fact that each individual step worked correctly offers little comfort.
The commercial real estate industry is particularly unforgiving of these types of errors. A report can be professionally written yet contain an incorrect entity assignment, an inaccurate date, or a formula in the wrong cell. A lease summary might accurately cite the document but overlook the most important provision for the owner. An underwriting memo can appear coherent, despite starting from a false assumption.
Therefore, the industry must distinguish between generating an answer and creating accountable work products. The latter requires more than a powerful AI model; it demands documentation of the sources used, the calculations verified, changes from the previous period, remaining uncertainties, and the decisions that still lie with an individual. Trust is not a confidence score next to an answer.
Targeted Applications and Verification Mechanisms
A system that states it is '92 per cent confident' does not tell an investment committee whether the underlying Net Operating Income (NOI) has been reconciled with the accounting system. It does not inform an asset manager if the latest lease amendment has been incorporated. It does not reveal to a controller who approved an exception. Trust is built through evidence.
- —Verification of calculations against the source workbook.
- —Confirmation that data sets are current.
- —Identification of discrepancies and pausing on conflicting information.
- —Retention of human approval for critical actions, rather than allowing the system to make the decision.
The way forward is not to have a single AI model handle all tasks. Instead, the work can be broken down into smaller, more clearly defined steps. One system can identify the correct source documents. Another can compare spreadsheet cells or recalculate a figure. Yet another can test whether the request was based on a flawed assumption. A separate audit can reconcile the final result with the accounting system, an approved template, or another reference source. If the evidence is contradictory or requires judgment, the work is passed to an individual.
Specialised systems are not necessarily more intelligent than the largest general-purpose models, but they can perform a narrower task that can be tested more consistently. A tool that compares spreadsheet calculations, for example, can be evaluated based on a clear right or wrong answer. The same applies to checking whether a document is current, whether a user has permission to view it, or whether a required approval has been obtained.
Our research shows how much these accompanying checks can influence performance. In one benchmark, changing information retrieval and linking increased recall precision from approximately 72 per cent to 92 per cent. The frequency of retrieving outdated or unauthorised information dropped from about 41 per 1,000 retrievals to less than one. The underlying AI model did not change, but the improvements resulted from providing better context, assigning each step to the appropriate tool, and independently verifying the result. These results illustrate a broader point: better work does not necessarily come from buying a larger model, but can be achieved by breaking down work into defined steps, using the right tool for each step, and verifying the final result against a proven company reference. Arunabh Dastidar is the co-founder and CEO of the real estate investment platform Leni.
