Which AI we use
1.1 Today: Google's Gemini models on Google Cloud Vertex AI, to read documents and draft Reports, and Google Document AI for OCR. Vendors, regions and terms are listed at https://www.bhoomiscan.in/subprocessors.
1.2 We do not use Anthropic, OpenAI or any other model provider that is not on that list.
People decide; AI drafts
2.1 No automated legal decisions. The AI Title Report sets out facts and flags risks. It does not decide whether title is good, whether a loan should be made, or whether land should be bought. Those are decisions for you, your advocate and your lender.
2.2 An advocate signs every opinion. A Professional Title Opinion is always written, checked and signed by an Empanelled Advocate. AI output is one input to the advocate's own review.
2.3 Where our team corrects a Report by hand (for example after the AI misread a handwritten deed), the corrected Report replaces the generated one for that scan, and the change is kept on record.
Every finding shows its source
3.1 Each fact in a Report points to the document and page, or the public record, it came from, so that a person can check it.
3.2 Where the AI is unsure, or a document is unreadable or missing, the Report says so plainly instead of guessing. A missing fact is shown as missing, never filled in.
3.3 Safety rules wrap the AI. If our own checks find a serious problem, the Report cannot show a clear result over it. Rules may make a finding stricter than the AI's view, never softer.
Data used by AI
4.1 We send AI providers only what the task needs.
4.2 Our providers do not train their models on the data we send, under the Google Cloud terms we use.
4.3 We do not train or fine-tune models on identifiable Personal Data. We use de-identified Reports for improvement only if you switch on "Help improve BhoomiScan" (off by default), and you can switch it off at any time.
Watching for errors and bias
5.1 We test changes against a fixed set of real and synthetic documents before release, including handwritten and Odia-script deeds, and we do not release a change that makes catastrophic misses worse.
5.2 We compare what the AI produced with what the Report showed, so that safety rules do not hide real risks.
5.3 We watch for errors that fall unevenly, for example on Odia-script names, older Sabik records, or particular districts, and fix the cause.
When AI gets it wrong
6.1 Report an error with "Report a problem" in the product, or at https://www.bhoomiscan.in/grievance. We treat it as a grievance (24 hours to acknowledge, 15 days to resolve).
6.2 A material error caused by our system is corrected and refunded under the Refund and Cancellation Policy.
6.3 We log every AI error incident, its cause and its fix, and review the log monthly.
Model changes
7.1 When we change the main model used for Reports, we update this page and the Sub-processors list, and record the change and its date.
7.2 Institutional Users get at least 30 days' notice of a change of model provider.