Government / Public Sector
Federal agency automates identity verification and citizen payouts
CLIENT
A federal agency administering citizen entitlement payouts
Cadence
Daily run
Client Overview
Refunding unclaimed moneys means verifying that the claimant, the account owner, and the receiving bank account are all genuinely connected, checked against identity documents, bank statements, and sometimes a name-change letter, before a government payment goes out. That verification ran manually across three separate systems, with a large team working through a high daily claim volume, and roughly one in ten claims already needing an exception path.
How the Workflow Runs
The refund process now runs as a single scheduled pipeline across the claims portal, ledger, and financial ERP. Every claim follows the same four phases, and anything that does not cleanly pass is routed to a person rather than guessed at.
FIG. 01 · Daily citizen-payout run across three systems
01
Collect & read
BOT + Agent
Claims and supporting documents are gathered automatically. A document-intelligence layer classifies each attachment and extracts identity and bank details, sending anything it cannot classify to human review.
02
Match on explicit rules
Agent
Names, IC numbers and bank details are cross-checked against deterministic match rules, with no black-box scoring. Claims under a reserved payment code always escalate.
03
Human checkpoint
Human
Case officers clear every claim that failed matching, carries the reserved code, or had a low-confidence extraction, approving, rejecting, or sending it back for more documents. Nothing is auto-decided on their behalf.
04
Post, certify & prove
Bot + Agent
Verified claims post to the ledger under maker-checker re-confirmation, pass a budget-sufficiency check that defers rather than fails, and are certified twice with a one-time code before disbursement.
How Cognitive Consulting helped
- Public money changes the arithmetic. We ranked the manual checks by what happens when one goes wrong, not by how long each one took.
- The client wanted the document layer to call ambiguous identity documents. We argued the other way: the confidence threshold is a business setting, not a model artefact.
- Every match rule is written out in full and readable by an auditor: no scoring, no weights. A budget shortfall defers the payment rather than failing the claim.
OUTCOMES
- Low-confidence extractions auto-escalate to a case officer, never guessed at
- Every match and budget check runs on explicit deterministic rules, no black-box scoring
- Payments certified twice with a one-time code before disbursement, every step logged
RELATED CASE STUDIES