Most AI real-estate tools get evaluated on the wrong axis: does the render look good? That is a design question, not the first engineering question for an output that may inform an investment committee or lender.
Four questions separate a usable early-stage planning tool from a demo.
1. Does the same input produce the same output?
Run the same site boundary and requirements through the system twice. If the answer changes materially without an input change, the system should disclose that behaviour and its cause. A planning option should be reproducible from its inputs, or the workflow should identify the controlled source of variation.
2. Is the output checked against something real?
A layout can be geometrically valid and still fail against the site rule that governs it. The useful test is not whether a drawing looks plausible; it is whether the stated unit mix, setback and floor-count assumptions are checked against the applicable, current source material.
3. Are limits disclosed before they matter?
Every early-stage system has a boundary: unusual site geometry, a rule it does not encode or a condition it cannot observe from available data. A credible workflow names those limits and makes the uncertainty visible in the output.
4. Where does professional sign-off occur?
Early planning output should remain clearly labelled as schematic until the responsible qualified professional completes the required review. The test is whether that hand-off is built into the workflow, rather than hidden in a footer after the result has already been treated as final.
What this means in practice
These questions cannot be answered from one polished image. Test repeatability, trace at least one site assumption to its governing source, ask where the system is known to fail and inspect where professional review becomes mandatory. That is the standard a planning aid should be designed to meet.
This is an evaluation framework, not a claim about a specific competing product or a compliance certification.