Every commercial real estate platform now has an AI story, and most of them demo well. The distance between a good demo and something an asset manager will put in front of an investment committee is wider than the marketing suggests.
We build reporting software for CRE, so we spend our time in the part of the stack where AI either earns trust or loses it, which is the numbers. These are the five problems that come up on nearly every project.
1. The data lives in systems that disagree with each other
A single portfolio might run Yardi for property accounting, MRI for a group of assets acquired three years ago, and a set of Excel workbooks holding valuations, mortgage detail and the budget assumptions nobody ever moved into either system.
Before a model can say anything useful, somebody has to decide which source wins when two of them conflict. That is not a technical decision, it is a commercial one. Whether a property’s occupancy comes from the rent roll or the operating statement is a reporting policy question, and no amount of AI removes the need to answer it. Skip that step and you have automated an argument rather than a report.
2. Language models are fluent when they are wrong
An assistant that reads a rent roll and tells you portfolio occupancy is 91.4 percent sounds exactly the same whether it is right or off by two hundred basis points. In most software an error announces itself. In this kind, it arrives well written.
For finance teams this is the central objection and it is a fair one. Confidence is not accuracy, and a reader cannot tell the difference from the sentence alone. Any serious deployment has to show its work rather than ask to be believed.
3. Traceability is treated as a feature instead of a foundation
If you cannot click a figure and see the transactions behind it, nothing else about the tool matters much. An auditor will not accept that the model calculated it, and neither will an investor who has spotted a variance you did not mention.
This is the part most often left until last, and it is the hardest to retrofit. Traceability is a property of how the data was assembled in the first place. If the aggregation step threw away the link back to the source record, no amount of clever interface work will put it back.
4. Real portfolio data is far messier than demo data
The examples in a vendor demo are tidy. Actual portfolios are not. What we see instead:
- Account codes that changed partway through the fiscal year.
- Two properties with nearly identical names, one of which was sold.
- Recoveries posted to different accounts depending on who did the posting.
- Budgets that exist in three versions, only one of which was approved.
A model given this material without a mapping layer will produce something plausible and quietly wrong. The unglamorous work of mapping accounts, properties and periods is what makes the interesting work possible, and it does not go away.
5. Nobody has decided who owns the answer
When an AI assisted figure turns out to be wrong in a published report, who is accountable? The analyst who accepted the suggestion, the head of reporting who signed off the section, or the vendor whose tool produced it?
Most organisations have not written this down, and the question surfaces at the worst possible moment. Decide it before the tool goes live. In practice that means a named reviewer per section and a record of who approved what, which is a governance problem more than a software one.
Where this leaves us
None of this is an argument against AI in commercial real estate. The teams getting real value from it are the ones treating it as a drafting and investigation layer sitting on top of data they already trust, rather than as a shortcut around the work of making the data trustworthy.
Get the aggregation and the audit trail right and AI becomes genuinely useful for the thing it is good at, which is helping a person ask better questions of numbers that already tie out.
What are the biggest challenges you’ve encountered integrating AI into your CRE workflow?