Learning Objectives

After completing this lecture, participants will be able to distinguish pilot conditions from production realities when scaling government AI, articulating the specific gaps that cause pilots to fail in production. Learners will map the OMB M-24-10 rights-impacting and safety-impacting designations to concrete pre-deployment risk management practices, and connect NIST AI RMF GOVERN, MAP, MEASURE, and MANAGE functions to a phased rollout plan. Participants will build an infrastructure readiness checklist covering FedRAMP authorization boundaries, FISMA moderate or high baselines, redundancy, observability, and disaster recovery. Learners will design change management programs for frontline workers, union representatives, and impacted communities, drawing on lessons from the IRS ID.me rollout and Michigan MIDAS failure. Finally, learners will outline contingency and rollback procedures that allow agencies to pause deployment safely when post-launch monitoring reveals drift, disparate impact, or incident patterns similar to the Dutch childcare benefits scandal. By the end of this 120-minute workshop, each learner should be able to lead a scaling committee review for a rights-impacting AI in their agency.

Key Topics Covered

This lecture examines the full scaling playbook that turns a successful pilot into a dependable production system serving millions of citizens. Topics include the pilot-versus-production gap and why small-scale assumptions collapse under load; OMB M-24-10 minimum practices for rights-impacting and safety-impacting AI including pre-deployment testing, ongoing monitoring, human oversight, and public notice; infrastructure readiness covering FedRAMP moderate or high, FISMA controls, horizontal scaling, failover, backup, and multi-region architectures; change management for operators, caseworkers, AFGE union stakeholders, and the public; staffing models including SRE and MLOps rotations; gradual rollout strategies such as geographic, temporal, canary, and shadow-mode deployments; KPI and SLO definition aligned with GAO AI Accountability Framework; risk management using NIST AI RMF MANAGE function; incident response and rollback; vendor management for FedRAMP-authorized providers; and integration with legacy mainframes common at IRS, SSA, and VA. Each topic is grounded in real federal examples so that managers can translate the framework into their own program.

Why This Matters for Government

Government AI pilots succeed on curated data at small scale. Production must serve every citizen, every day, with legal, equity, and security guarantees. The transition is where most agency programs break. The IRS ID.me deployment verified a few thousand taxpayers in pilot and collapsed when tens of millions of citizens tried to authenticate, forcing Treasury to reverse course in 2022 after civil rights objections from EPIC and members of Congress. The Michigan MIDAS unemployment fraud system performed acceptably in testing but issued more than 40,000 false fraud determinations after scaling, requiring a multi-million dollar settlement. The Dutch SyRI and childcare benefits scandal showed how an apparently successful algorithmic risk model at national scale caused 26,000 families to be wrongly accused of fraud, ultimately bringing down the Rutte government. In every case the gap was not the model. It was the scaling process. OMB Memorandum M-24-10 now requires federal agencies to designate rights-impacting and safety-impacting AI, complete impact assessments, maintain AI use case inventories under EO 14110, and provide public notice before production use. The NIST AI RMF 1.0 MANAGE function requires documented controls for drift monitoring, human oversight, and decommissioning. The GAO AI Accountability Framework requires performance monitoring, data quality tracking, and transparency obligations for any AI that supports agency decisions.

The pilot to production gap is substantial. Pilots are controlled experiments. Data is curated, team is small, scope is narrow, downtime is tolerable, failure is instructive. Production is the opposite. Data arrives messy and adversarial, scale is massive, downtime damages public trust, failure triggers congressional oversight and OIG inquiries, and scope is universal. A pilot that classified 10,000 VA disability claims per month at 94 percent accuracy may drop to 78 percent on the full 1.3 million claim annual workload because edge cases, data drift, and integration errors surface only at full scale. GSA and OMB analyses document that up to 80 percent of federal AI pilots never reach sustained production.

Infrastructure and authorization readiness is foundational. Federal cloud AI requires FedRAMP moderate or high authorization for the platform and any AI APIs, a FISMA system security plan, continuous monitoring, and an authorization to operate signed by the authorizing official. Compute must be load tested at two to three times peak volume. Storage must support Federal Records Act and NARA retention. Networking must be benchmarked for tail latency. Redundancy must include active-active or hot standby failover across at least two FedRAMP regions. Observability must include structured logs, per-decision audit records sufficient for FOIA and Privacy Act requests, real-time metrics, and alerting runbooks. Security must include encryption in transit and at rest, role-based access control tied to agency identity providers, and continuous vulnerability scanning aligned with CISA Binding Operational Directives. Disaster recovery must include tested restore procedures, recovery time objectives, and recovery point objectives.

Change management and workforce considerations drive adoption. Frontline caseworkers at SSA, VA, IRS, and DHS do not simply receive a new tool. Workflows, performance metrics, and bargaining unit protections change. Successful rollouts invest early in stakeholder impact assessment, AFGE and NTEU union consultation where collective bargaining agreements apply, audience-differentiated training, pilot training with iteration, phased rollout in waves, and ninety-day dedicated help desk support. The Department of Labor and EEOC have both issued guidance that AI workplace tools must not produce disparate impact on protected classes.

Gradual rollout is the primary mitigation for scaling risk. Shadow mode runs the new system alongside the legacy process, producing outputs that are logged but not used for decisions. Canary deployment sends a small percentage of real traffic to the new system. Geographic rollout starts with one state or region. Temporal rollout increases share of traffic on a schedule with go or no-go gates. Parallel operation maintains the prior system long enough to fall back. Monitoring during rollout must include model performance, fairness across demographic groups in line with Blueprint for an AI Bill of Rights protections against algorithmic discrimination, incident counts, user satisfaction, and downstream outcomes. A pause-and-fix posture is essential. The Allegheny County Department of Human Services child welfare screening tool exemplifies mature scaling: published validation, community advisors, shadow mode, and explicit carve-outs for the most sensitive decisions.

Case studies and anti-patterns guide practice. The IRS ID.me deployment scaled without adequate accessibility testing. The Houston HISD teacher evaluation model scaled without adequate transparency and was struck down in Houston Federation of Teachers v HISD for due process violations. The COMPAS recidivism tool scaled across jurisdictions without public-sector validation and was found by ProPublica to exhibit racial disparities. Anti-patterns include scaling without load testing, without union consultation, without a rollback plan, without OMB-required public notice, and treating FedRAMP authorization as a box check rather than a continuous control. Agencies that promise leadership they will never pause a rollout are setting up the next Michigan MIDAS.

Overview

Government AI pilots succeed on curated data at small scale. Production must serve every citizen, every day, with legal, equity, and security guarantees. The transition is where most agency programs break. The IRS and ID.me deployment showed how a pilot that verified a few thousand taxpayers collapsed when tens of millions of citizens tried to authenticate, forcing the Treasury to reverse course after civil rights objections from the Electronic Privacy Information Center and members of Congress. The Michigan MIDAS unemployment fraud system performed acceptably in testing but issued more than 40,000 false fraud determinations after scaling, requiring a court-ordered multi-million dollar settlement. The Dutch SyRI and childcare benefits scandal showed how an apparently successful algorithmic risk model at national scale caused 26,000 families to be wrongly accused of fraud, ultimately bringing down the Rutte government. In every case the gap was not the model. It was the scaling process. OMB Memorandum M-24-10 now requires federal agencies to designate rights-impacting and safety-impacting AI, complete impact assessments, maintain AI use case inventories under EO 14110, and notify the public before production use. The NIST AI RMF 1.0 MANAGE function requires documented controls for drift monitoring, human oversight, and decommissioning. The GAO AI Accountability Framework requires performance monitoring, data quality tracking, and transparency obligations for any AI that supports agency decisions. Scaling without these controls is no longer just risky. It is non-compliant.

The Pilot to Production Gap

Pilots are controlled experiments. The data is curated, the team is small, the scope is narrow, downtime is tolerable, and failure is instructive. Production is the opposite. Data arrives messy and adversarial, scale is massive, downtime damages public trust, failure triggers congressional oversight and OIG inquiries, and scope is universal. A pilot that classified 10,000 VA disability claims per month at 94 percent accuracy may drop to 78 percent on the full 1.3 million claim annual workload because edge cases, data drift, and integration errors surface only at full scale. The GSA AI Centers of Excellence have documented that up to 80 percent of federal AI pilots never reach production. The most common reasons are infrastructure insufficient for load, staffing misaligned with operations, governance gaps that surface during ATO review, and change management failures that leave frontline employees unable or unwilling to trust the system. Production readiness therefore requires explicit attention to each of these dimensions before rollout, not after.

Infrastructure and Authorization Readiness

Before a government AI system can move to production it requires a complete authorization boundary. For cloud-hosted systems this means a FedRAMP moderate or high authorization for the underlying platform and any AI APIs, a FISMA system security plan, and an authorization to operate signed by the agency authorizing official. Compute must be load tested at three times peak expected volume. Storage must accommodate historical records required by the Federal Records Act and NARA retention schedules. Networking must be benchmarked for tail latency, not just mean. Redundancy must include active-active or hot standby failover across at least two FedRAMP regions. Observability must include structured logs, per-decision audit records sufficient for FOIA and Privacy Act requests, real-time metrics, alerting runbooks, and dashboards visible to program leadership. Security must include encryption in transit and at rest, role-based access control tied to agency identity providers, continuous vulnerability scanning, and alignment with CISA Binding Operational Directives. Disaster recovery must include tested restore procedures, recovery time objectives, and recovery point objectives set by mission criticality. Without each of these controls in place the authorizing official cannot responsibly sign off.

Change Management and Workforce

Scaling changes jobs. Frontline caseworkers at SSA or VA do not simply receive a new tool. Their workflows, performance metrics, and sometimes their bargaining unit protections change. Successful rollouts invest early in change management. The playbook includes impact assessment mapping every role affected, collaborative design sessions with line staff, AFGE and other union consultation where collective bargaining agreements apply, appropriate training differentiated by operator versus supervisor versus executive audience, pilot training with a small cohort followed by iteration, phased rollout in waves to allow support to scale with demand, dedicated help desks for the first ninety days, and explicit psychological safety for employees who flag problems. The Department of Labor and EEOC have both issued guidance that AI workplace tools must not have disparate impact on protected classes. Training materials must emphasize the human-on-the-loop requirement in M-24-10 so staff understand when to override. Resistance is natural and is often the earliest signal of a real problem in the system.

Risk Management and Gradual Rollout

Scaling introduces risks that were invisible at pilot. A one percent error rate on 10,000 cases is 100 affected citizens. On 10 million cases it is 100,000. For rights-impacting AI this is unacceptable. NIST AI RMF MANAGE function requires that risks be prioritized, mitigated, and monitored continuously. Gradual rollout is the primary mitigation. Shadow mode runs the new system alongside the legacy process without using AI outputs for decisions, producing a direct comparison. Canary deployment routes a small percentage of real traffic to the new system. Geographic rollout begins with one state or region before national expansion. Temporal rollout increases the share of traffic on a weekly schedule with go or no-go gates at each step. Parallel operation maintains the prior system long enough to fall back. Monitoring during rollout must include model performance, fairness across demographic groups in line with Blueprint for an AI Bill of Rights protections against algorithmic discrimination, incident counts, user satisfaction, and downstream outcomes. A pause-and-fix posture is essential. Agencies that promise leadership they will never pause a rollout are setting up the next Michigan MIDAS.

Case Studies and Anti-Patterns

The IRS ID.me deployment scaled without adequate accessibility testing, forcing taxpayers unable to pass facial recognition into long phone queues and triggering a 2022 Treasury reversal. The Houston HISD teacher evaluation model scaled without adequate transparency and was struck down by a federal court in Houston Federation of Teachers v HISD for violating due process. The COMPAS recidivism tool scaled across jurisdictions without public-sector validation and was found by ProPublica to exhibit racial disparities in false positive rates. The Allegheny County child welfare screening tool took a different path, publishing its validation study, engaging community advisors, and using shadow mode before production, offering a model for rights-impacting scaling. Anti-patterns to avoid include scaling without load testing, scaling without union consultation, scaling without a rollback plan, scaling without OMB-required public notice, and treating FedRAMP authorization as a box check rather than a continuous control. The lecture closes with a practice exercise in which learners develop a phased scaling plan for their own agency, including authorization milestones, KPI thresholds, rollback triggers, and communications plans that satisfy the transparency obligations of EO 14110 and the Blueprint for an AI Bill of Rights.

Start Your CLUB Certification

This lecture is part of L3: AI Strategist, 80 hours of comprehensive government AI training aligned with NIST AI RMF, OMB M-24-10, EO 14110, and the GAO AI Accountability Framework. Explore CLUB Certification.

Related Lectures

L3 3.5.1 AI Metrics and KPIs for Government, 90 minutes, workshop and dashboard. L3 3.5.3 Data Infrastructure for Enterprise AI, 90 minutes, lecture and architecture. L3 3.5.4 Cross-Agency AI Coordination, 90 minutes, case studies and framework. L3 3.4.2 Quality Assurance for AI Work Products, 90 minutes, lecture and checklist. L3 3.6.1 Moving from Pilot to Production anti-patterns, 60 minutes, case review.