How it works
Every finding in a Scrivenza case review traces back to a specific, documented rule — not an unexplained AI judgement. This page explains how that's actually built, and how your data is handled along the way.
Rules you can read, not a black box
The compliance checks behind a case review aren't one freeform instruction to an AI model. They're built from a set of separately authored rule sections — gambling, undeclared credit commitments, document integrity, housing costs, and so on — each with its own written description of what it checks and why. Gambling detection, for example, is its own section rather than folded into general spending because the Financial Ombudsman Service has repeatedly held mortgage advisers responsible for missing gambling patterns that affordability or vulnerability assessments should have caught.
You choose what's checked — genuinely
Every case lets the broker or AR firm choose which categories apply before running the review. Unchecking a category isn't a display filter on the report — that rule is left out of what's sent for analysis entirely, so nothing is checked for it. The report reflects this honestly too: it distinguishes "not checked for this case" from "checked, and nothing was found," rather than treating the two as the same thing.
Every finding cites its rule
Each finding in a report carries a "Why we checked this" toggle showing the literal rule wording that produced it — not a paraphrase or a marketing description of the rule, the actual text used for that check. A broker or compliance officer can see precisely why something was flagged rather than taking it on faith.
A record of exactly what happened
Every case captures which AI model produced it, and the exact rule text as it existed at that moment — so if a rule is later refined, reopening an older case still shows what actually governed it at the time, not today's version. This is available on every report itself, under "Audit details."
Where your data actually goes
Bank statement data is processed using Anthropic's Claude, via their commercial API, under terms that contractually exclude it from training their models and delete it automatically within 30 days. See the full data-handling and security picture →