Give an AI a sparse property record and ask for a compelling investment story. It may return something impressively fluent. The missing lease amendment has not, alas, materialised through literary effort.
A German qualitative study submitted to arXiv on 11 September 2026 offers a useful prompt for reflection. Based on eleven semi-structured interviews, it reports uneven adoption of generative AI in real estate marketing, with property-description writing the most established use case. Its authors identify integration with listings and documents, data availability and compliance as important constraints. This is a small qualitative preprint, not a representative market survey or proof of productivity gains. [1]
Fluency is not a control environment
Our interpretation is straightforward: the next useful demonstration of property AI should begin with evidence and end with an authorised decision. The beautifully rewritten description can come along, provided it brings its working.
Suppose the rent roll, the approved underwriting workbook and the latest lease amendment disagree. A credible assistant should surface the conflict, identify the source versions and ask the appropriate person to resolve it. Averaging the numbers would be mathematically neat and commercially absurd.
For a sales executive, the benefit is a cleaner conversation with the buyer. For the analyst, it is a reproducible calculation. For the IT leader, it is knowing which identity accessed which document and under whose authority. The same workflow must satisfy all three.
Measure the work that reaches a decision
We would evaluate a property assistant with a small, deliberately awkward test set before admiring its prose:
- Can it distinguish a signed term from a draft negotiation position?
- Does it expose a missing source rather than inventing a convenient fact?
- Does it notice when the approved model has changed beneath a proposal?
- Can a reviewer trace each material claim back to the evidence they are entitled to see?
- Does it stop when its authority ends?
Track unsupported-claim rates, reviewer corrections and time to an approved output against a defined manual baseline. Record the model, prompt and evidence versions. Compare like tasks and include the time spent checking the result. A faster first draft is not automatically a faster completed deal.
These are proposed evaluation measures, not results we claim to have achieved. The distinction matters. Even a monkey with excellent academic manners should not award itself a control group.
The Rainmaker connection
Rainmaker’s current staging product includes versioned proposal drafts, reviewed finite edits, evidence and model-version checks, plus separate platform and tenant AI controls. Tenant model overrides require their own credentials; live provider execution still requires configuration and verification. Those controls are a foundation for measuring useful assistance, not a claim that every workflow is already autonomous.
Our product direction is to put that discipline inside the work of winning and executing a deal. Prepare the next action, show the supporting evidence, let the responsible person decide, and capture the result without asking them to enter it again for a dashboard.
Ask to see a proposal after its underlying model changes. That is a more revealing demo than asking an assistant to describe the lobby in five different tones.
Source and scope
[1] Victor Kolominsky-Rabas, Leopold Müller, Felicia Perpina and Niklas Kühl, Generative AI Use Cases In Real Estate Marketing: Adoption and Constraints in Germany, arXiv preprint 2609.12684v1, submitted 11 September 2026; accessed 5 October 2026. https://arxiv.org/abs/2609.12684
From the Canopy · Fresh perspectives. Sharper deals. This article is original commentary on recent research, not a claim of a new event today or an endorsement by the researchers.