Civica_
$28M
Recovered
In year one
-38%
False positives
Versus prior system
94%
Disposition agreement
Investigator audit
2.4x
First-year ROI
Payback in 9 months
The challenge
PUBLIC TRUST FIRST
Civica's investigators were chasing an estimated $80M in annual benefits fraud, but their existing rules screen was flagging 14 percent of legitimate claimants and drawing legislative criticism. A poorly implemented AI pilot had already been pulled by the state ombudsman in 2021, so any new AI for government agencies had to prove it was safer than the status quo.
The bar was clear: defensible accuracy, demographically audited fairness, and full transparency for civil-society oversight groups.
$80M
Estimated annual fraud
14%
False-positive rate
Pulled
2021 vendor tool
The approach
FOUR PHASES
We structured the work as four phases so each release delivered visible savings and rebuilt trust in AI consulting for government agencies across elected officials, civil servants, and community advocates.
Phase 1
Wks 1–10Stakeholder Alignment
Before any model development, we aligned policy, advocacy, and operations stakeholders around clear guardrails for deployment.
Worked with ombudsman, ACLU advisory council, and community panel before any model work
Captured fairness constraints and veto conditions that would govern AI for government agencies
Drafted a public fairness charter and investigator bill of rights
Built a bias dashboard prototype to show what would be monitored
Phase 2
Wks 11–24Model Development
The model phase prioritized both fraud performance and fairness under public-sector constraints and review standards.
Trained constrained gradient-boosted models with monotonicity and equal-odds constraints
Tuned for case-handler productivity, not just fraud catch, to reflect real workloads
Integrated SHAP-based explanations suitable for AI driven public sector solutions under FOIA
Produced disparity reports by demographic segment for independent review
Phase 3
Wks 25–34Investigator Rollout
Rollout focused on investigator usability, explainability, and strict human oversight for every action.
Triage cockpit for 220 investigators that ranked cases by impact and evidence strength
Human-in-the-loop review so no action occurred without staff approval
Public-facing transparency portal that described how the AI ranked cases in plain language
Training program that paired new tools with scenario-based practice sessions
Phase 4
Wks 35–60Public Operations
Post-launch operations centered on transparent reporting, independent re-audits, and safety controls.
Quarterly bias re-audit with external observers and the ombudsman's office
FOIA log and documentation pack generated automatically for every model release
Stakeholder report for legislators and civil-society partners on savings and fairness
Auto-pause guardians that halted scoring if drift or bias thresholds were crossed
"
They knew, before we did, that the politics of public-sector AI matter as much as the math. That is why this stuck. We built consent first, then code. The result is AI for government agencies that our investigators trust and our oversight bodies can defend.
AM
Anita Morales
Deputy Commissioner, Civica State Agency
More like this
RELATED STUDIES