A V-CIP session that ends in a binary pass or fail misses the most important category of outcome: the edge case that deserves a second look but not a rejection. A face match score of 78 against an Aadhaar photograph from 2009 is not a failed session — it is an aged photograph requiring a different calibration. A geolocation that shows Assam while the borrower says they are in Shillong is not a spoofing attempt — it is a GPS boundary accuracy issue. The Video KYC Moderator AI knows the difference. The refer decision is where that knowledge is most consequential.
Why the refer decision is the hardest — and the most important
A clean pass is straightforward: all checks passed above threshold, the agent confirms, the session is complete. A clear fail is also straightforward: fraudulent documents, identity mismatch, active deception. The edge case — where the evidence is genuinely ambiguous, where the right answer requires context and judgment — is where a poorly designed V-CIP system causes the most damage.
A system that defaults to rejection on any ambiguity rejects creditworthy borrowers for circumstances beyond their control: an old photograph on an Aadhaar they have not updated, a poor connection that degraded video quality at the document capture moment, a name in their CKYC record that predates a marriage. A system that defaults to pass on any ambiguity creates compliance exposure by certifying sessions where the required checks have not been conclusively met.
The Video KYC Moderator AI's refer decision is not a failure to decide — it is a structured escalation with a specific reason, a specific evidence brief, and a specific resolution pathway. The human agent receiving a refer session knows exactly what the AI found, why it could not auto-resolve it, and what they need to do to reach a determination.
The three-tier decision framework
All 14 compliance checks met — agent visual confirmation received
Liveness above 90, face match above 85, all documents verified, geolocation confirmed, no flags raised. Agent has reviewed the AI compliance summary and confirmed visual identity match. Session record sealed and archived. Application proceeds to underwriting. Accounts for 84.1% of all V-CIP sessions.
AI has identified a specific ambiguity that requires human judgment to resolve
The session has progressed but a specific check produced a result in the ambiguous zone — not a clear fail, not a clear pass. The AI generates a structured refer brief: the specific check that triggered the refer, the evidence for and against, the likely benign explanation, and the agent's decision options. The agent reviews the brief and makes a determination — pass with note, request re-session, or fail. Accounts for 12.7% of sessions.
Evidence of identity fraud, document forgery, or deliberate misrepresentation
Liveness failure combined with high face match (deepfake signature), document tampering evidence, identity mismatch between live face and presented documents, or deliberate location spoofing. Application rejected. Fraud team notified. Device fingerprint and identity elements flagged for consortium sharing. Accounts for 3.2% of sessions.
The 10 most common edge cases — and how the AI handles each
Face match 74% — Aadhaar photo is 11 years old
The borrower is 34 years old. Their Aadhaar was issued in 2014. The 128-point face match against the decade-old photograph returns 74% — below the standard 85% threshold. The AI applies an age-adjusted threshold: for Aadhaar photographs older than 7 years, the threshold is reduced to 70% and the match is supplemented by a secondary check against the CKYC photograph (which is more recent). With age adjustment applied, the match at 74% is above the adjusted threshold of 70%.
→ Auto-pass with age-adjustment note logged · No refer requiredDocument OCR failed — bandwidth degradation during Aadhaar capture
The borrower's video quality dropped to below-threshold resolution at the moment they held up their Aadhaar card — a network packet loss event, not a deliberate act. OCR confidence is 54%, insufficient for reliable text extraction. The AI does not fail the session: it flags the specific failure, prompts the borrower to reposition and recapture, and retries up to three times. If retries succeed, the session continues. If all three retries fail, the AI generates a refer brief recommending a session retry at a better connection time rather than a failure decision.
→ Refer: session retry recommended · Not a failure — technical constraintAadhaar "Sunita Devi Sharma" vs PAN "Sunita Sharma" — middle name absent
The name on the Aadhaar includes a middle name that the PAN record does not include. The AI's name reconciliation module computes a fuzzy match score of 88% (middle name omission is a known and common Indian name rendering variation). At 88%, this is above the name match threshold of 80% for this category of variation. The discrepancy is logged in the session record with the reconciliation rationale — so the agent and any future reviewer can see why it was accepted without the agent needing to independently assess.
→ Auto-pass with reconciliation note · Discrepancy documentedGPS shows location 800m from Bangladesh border — borrower states they are in Dhubri, Assam
The GPS coordinate falls within India — Dhubri district, Assam — but the proximity to the international border triggers the AI's border-zone flag. The IP geolocation confirms an Indian ISP. The borrower's stated location (Dhubri) is consistent with both the GPS and the ISP. The flag is a geographic proximity alert, not a location spoofing indicator. The AI generates a refer brief documenting the border-zone flag, the confirming IP evidence, and the likely benign explanation — a resident of a legitimate border district conducting a normal V-CIP session.
→ Refer: border-zone flag with confirming evidence · Agent decision requiredFace match 81% — borrower now wears spectacles not present in 2018 Aadhaar photo
The live face match is affected by spectacle frame occlusion of the eye-zone landmarks — a significant contributor to the 128-point geometry score. The AI detects spectacle-frame occlusion and applies a spectacle-adjusted threshold, downweighting the occluded eye-zone landmarks and upweighting the unoccluded facial geometry zones (jaw, nose, cheek contour). Adjusted score: 87% — above the 85% threshold. Match confirmed.
→ Auto-pass with spectacle-adjustment applied · Adjustment rationale loggedBorrower failed the first liveness challenge — misunderstood the instruction
The active liveness challenge (randomised gesture sequence) was failed on the first attempt — the borrower performed the gestures in the wrong order, likely due to misunderstanding the prompt. Passive liveness score is 94% (strong — live person confirmed). The AI issues a second challenge attempt automatically, with a simplified instruction. The borrower passes the second attempt. The first failure is logged with the passive liveness score context — the session is not flagged as a liveness failure because the passive analysis provides strong corroborating evidence of liveness throughout.
→ Auto-pass after second attempt · First failure and passive score loggedFace match 92% but liveness score 38% — probable deepfake attack
The face matches the Aadhaar photograph with high confidence — but the liveness analysis identifies the micro-texture and frame-level inconsistencies characteristic of AI-generated video rather than a live camera feed. This is the canonical deepfake signature: the face is recognisably correct (because a real photograph of the victim exists) but the video is generated. The session is immediately failed. Fraud team alert generated. Device fingerprint logged for consortium sharing.
→ Fail: deepfake signature · Fraud alert + device flaggingFace match 68% — Aadhaar photograph partially obscured by worn laminate
The Aadhaar card presented has a worn laminate that creates a milky occlusion over part of the photograph — a physical degradation issue, not a fraudulent alteration. The AI detects laminate occlusion pattern (distinguished from deliberate tampering by the texture and extent of the obscuration) and generates a refer brief: the face match is below threshold due to physical document condition, the liveness score is strong at 94%, and the PAN cross-check is exact. Recommended resolution: accept with note, or request DigiLocker Aadhaar as an alternative clear photograph source.
→ Refer: laminate occlusion · DigiLocker alternative recommendedPAN OCR extracts wrong PAN — borrower showing PAN of a family member
The PAN extracted by OCR does not match the PAN entered in the application. The borrower, when the mismatch is flagged in real time, explains they picked up their spouse's PAN card by mistake. They retrieve their own PAN and re-present it — the second card OCR matches the application exactly. The first PAN mismatch, the borrower's explanation, and the re-presentation are all logged in the session record. The agent reviews the sequence and confirms the explanation is consistent with the visual evidence before issuing a pass.
→ Refer: PAN mismatch on first attempt · Agent confirms honest error and re-presentationBorrower refuses to state Aadhaar last 4 digits verbally — citing privacy concerns
The RBI guidelines require the customer to verbally confirm their Aadhaar last 4 digits on the video session — this is a mandatory requirement, not an optional step. A borrower who refuses this step — regardless of their stated reason — cannot have their V-CIP session passed, because a specific regulatory requirement has not been met. The AI generates a fail determination with a clear regulatory basis (MD 16(c)(i)) and a borrower communication explaining the requirement and offering to reschedule if they wish to reconsider.
→ Fail: mandatory RBI requirement not met · Clear basis logged · Session reschedule offeredThe refer brief: what the agent receives for every escalated session
When the Video KYC Moderator AI generates a refer, the human agent does not receive a raw video and a flag. They receive a structured brief that takes them directly to the decision: the specific check that triggered the refer, the evidence supporting and weighing against the benign interpretation, the AI's assessment of the most likely explanation, the borrower's response if they were prompted about the anomaly during the session, and the agent's three decision options with the regulatory basis for each.
A well-constructed refer brief reduces the agent's review time for a refer session from 7–10 minutes (watching the full recording to find the relevant moment) to 90–120 seconds (reading the brief, reviewing the flagged segment, making a determination). This efficiency is not just operational — it is quality-improving. An agent reviewing a concise, evidence-organised brief makes more consistent and better-reasoned decisions than an agent reviewing an unstructured raw recording under time pressure.
The quality of a V-CIP operation is measured in its refer rate — not its pass rate
An operation that passes 98% of sessions is not a high-quality operation — it is likely a permissive one. An operation that fails 20% of sessions is not a rigorous one — it is likely an inconsistent one that is penalising borrowers for circumstances the system cannot interpret correctly. The 12.7% refer rate — the sessions that genuinely require a human judgment call — is where the quality of the AI's edge case framework is most visible. A structured refer with a clear brief, a specific evidence summary, and a defined decision pathway produces better human decisions than an unstructured escalation. That is the Video KYC Moderator AI's most consequential contribution: not the 84% it automates, but the 13% it elevates.