Trust & technology
AI & models
AI works from the report's evidence to assist with writing and checking. It does not take responsibility for the appraisal.
Search the public guides. Your search stays in this browser.
On this page
Two tasks, with the appraiser in control.
Narrative assistance
The first-pass workflow drafts supported empty narrative sections from available report context. Existing appraiser text is preserved. The appraiser reviews the draft and decides what to retain, revise or discard.
Report checks
An AI check can surface inconsistencies and missing support for further investigation. It can miss issues or flag something incorrectly. A completed check does not validate a value conclusion or replace the firm's professional review.
The configured model and provider.
Scroll horizontally to read the full table.
| Layer | Current implementation | What it means |
|---|---|---|
| Narrative and report-check provider | OpenAI API | Requests originate in the authorized cloud service when a user requests assistance. |
| Configured model default | gpt-6-luna | Default in the reviewed service configuration. The deployed model can be configured separately. |
| Content screening | Google Cloud Model Armor | Screens request and response text; it does not determine appraisal accuracy. |
| Room scanning | Apple RoomPlan on supported devices | A separate device capability for geometry capture, rather than the narrative model. |
Model information reflects the implementation reviewed on October 5, 2026, rather than a live deployment readout. The service records the model used with each AI request. Confirm the active model and processing terms for a firm evaluation; model names and availability can change.
What the narrative model receives.
Requests can include subject and comparable details, narrative text, valuation inputs, inspection observations, source notes, assignment context, a text summary of floor-plan information and photo captions.
OpenAI narrative requests exclude photo image bytes, designated contact/access fields and account identity fields such as the user's name, email and user ID. This is not a claim that report text is anonymous: addresses and information entered in narrative fields may identify people or properties.
Google Cloud Model Armor screens broader report text, which can include the designated contact/access fields. The model-input exclusions do not mean those fields are excluded from every processing service.
OpenAI response storage is disabled for these requests. That setting is not a promise of zero provider retention. Consult the privacy policy for the processing and provider-retention disclosures.
Screening surrounds generation.
- AuthorizeCheck the account, workspace, seat and permitted action.
- Screen inputCheck the selected report context with Model Armor.
- Generate & screenRequest model output, then screen the response.
- ReviewReturn assistance for the appraiser to inspect.
Request records support traceability of who requested assistance, which report it concerned, the model used and the returned output. Internal prompts, screening configurations and evaluation fixtures are not part of the public documentation.