How our risk scores are calculated

Two scores per document

Each uploaded document is assessed on two independent dimensions, each on a 0–100 scale where a higher number means higher concern:

• Document validity — how genuine, complete and internally consistent the document appears.
• Client risk — how much the document's content raises the client's money- laundering / financing-of-terrorism risk.

From score to risk band

Each 0–100 score maps to a band that drives the workflow:

• Low — score below 34
• Medium — score from 34 to 66
• High — score of 67 or above

Higher bands pull in additional evidence requirements (enhanced due diligence) for that client.

What goes into a score

The assessment is grounded only in verifiable inputs, never guesswork about the person:

• The fields extracted intelligently from the document by AI (names, dates, identifiers, addresses, amounts), each with its own extraction confidence.
• The document's type and the client's context — individual or business, jurisdiction, and current risk band.

An Azure AI Foundry model reviews these inputs against a fixed, versioned KYC/KYB rubric and returns the two scores, a written rationale, and any risk flags it raised.

How a case's overall risk rating is assessed

Separately from the document scores above, each case gets one overall risk rating. An analyst rates six factors — each Low, Medium or High — and the platform's rules turn those ratings into a score from 0 to 100 and a band, using the same cut-offs as above. The rules are the same whether the ratings are typed in by the analyst or pre-filled from the case's own screening and document results, and the analyst is always accountable for them.

A Low rating counts as 15 points, Medium as 50 and High as 85; the score is the weighted average. Three rules stop an average from hiding something serious:

• a confirmed sanctions, PEP or serious adverse-media match makes the case High;
• a case flagged for enhanced due diligence is High;
• a case with any High-rated factor is never rated Low.

Confidence and human review

The system routes to a human rather than scoring confidently whenever it is unsure:

• If the model's own confidence is below 0.70, or
• if any extracted field's confidence is below 0.60 (for example a poor scan),

the document is flagged for manual review instead of being taken at face value.

Reproducible and auditable

For every score we store the exact extraction and review model identifiers, the prompt version, and a snapshot of the extracted fields the score was based on. Any past assessment can therefore be reproduced and explained against the precise logic and data used at the time.

Human decision, always

The score never decides anything on its own. An analyst reviews the AI assessment alongside the source document, then accepts or overrides it with a recorded reason. Every step is captured in an append-only audit trail, aligned to AMLA, FATF and FinCEN expectations.