{"stable_id":"957e48fc908e26a9","slug":"dwp-uc-advances-fraud","company":"UK Department for Work and Pensions","workflow_name":"ML risk referral with human fraud decision","function_code":"insurance","pattern_codes":["threshold_as_control","continuous_decisioning"],"changed_assumption":"Fraud review can focus on model-high-risk claims if adverse decisions remain human.","evidence_strength":"verified","publication_tier":"showcase","freshness":"current","reviewed_at":"2026-08-23","updated_at":"2026-08-22","source_quality_summary":"1 independent, 1 primary; publication outcomes are verified.","caveat_summary":"Department/NAO estimates; subgroup disparities require monitoring.","freshness_caveat":null,"workflow_summary":{"before":[{"actor":"Universal Credit advance process","action":"Receives a new claimant request and applies existing eligibility and fraud controls.","actor_type":"control"},{"actor":"Human decision maker","action":"Reviews evidence and alone determines whether an advance is fraudulent.","actor_type":"human"}],"hinge":"Fraud review can focus on model-high-risk claims if adverse decisions remain human.","after":[{"actor":"Advances machine-learning model","action":"Scores new advance requests and refers predicted high-risk cases.","actor_type":"ai"},{"actor":"Human decision maker and assurance sample","action":"Reviews referred cases; equal high/low-risk random samples monitor hit rate and false negatives.","actor_type":"human"}],"decision_mode":"shared","decision_marker":"Human sets the threshold; AI decides each instance"},"before":[{"order":1,"actor":"Universal Credit advance process","action":"Receives a new claimant request and applies existing eligibility and fraud controls.","handoff_to":"Human decision maker","control":"The fairness report does not document a dated pre-model operating queue; random sampling is the evaluation benchmark, not claimed as the full former workflow."},{"order":2,"actor":"Human decision maker","action":"Reviews evidence and alone determines whether an advance is fraudulent.","handoff_to":"Payment or fraud action","control":"Human adverse-action authority."}],"after":[{"order":1,"actor":"Advances machine-learning model","action":"Scores new advance requests and refers predicted high-risk cases.","handoff_to":"Prioritized human review queue","control":"Model is a prevention control; it does not make the fraud decision."},{"order":2,"actor":"Human decision maker and assurance sample","action":"Reviews referred cases; equal high/low-risk random samples monitor hit rate and false negatives.","handoff_to":"Payment, fraud action, or model monitoring","control":"Legitimate referrals are prioritized for timeliness; DWP conducts fairness assessment."}],"decision_rights":"A human confirms fraud; DWP owns thresholds and fairness review.","exception_path":"Legitimate referred cases are prioritized; random sampling checks misses.","removed_work":["Lower-yield random referral as primary control"],"outcomes":[{"metric":"Fraud identification","baseline":"Random sample","result":"2.9 times as effective; estimated £4.4m saved","period":"2024–25; savings since 2022","scale":"1.4 million advances worth £0.8bn in 2024–25","attribution_caveat":"Department/NAO estimates; subgroup disparities require monitoring.","evidence_label":"verified"}],"executive_lesson":"A defensible model narrows review; it does not own denial.","anti_pattern":"Do not automate adverse action from risk alone.","questions_for_leaders":["Where is the operating threshold set and who can override it?","What measured result would trigger rollback or retraining?","Which residual decisions must remain human-owned?"],"collections":["regulated-autonomy"],"bundle_version":"1.0.0","bundle_fingerprint":"sha256:c23c6cc2b88153d008ea8fda928f632ce0011fc2d4c5036672a16e5d895bab93","canonical_url":"https://brianletort.ai/transformations/dwp-uc-advances-fraud"}