Learning Objectives

By the end of this lecture you will (1) define the six families of metrics and KPIs relevant to federal AI: model performance, fairness and equity, operational, cost and efficiency, trust and transparency, and mission-outcome; (2) align AI metrics with the Government Performance and Results Modernization Act of 2010 (GPRA Modernization Act) and the Foundations for Evidence-Based Policymaking Act of 2018, which jointly structure agency performance reporting; (3) apply NIST AI RMF MEASURE function metrics for trustworthiness characteristics including validity and reliability, safety, security and resilience, accountability and transparency, explainability and interpretability, privacy, and managed bias; (4) meet OMB M-24-10 expectations for performance measurement including the annual AI Use Case Inventory, which must include intended use, deployment status, and performance indicators where appropriate; (5) interpret the GAO AI Accountability Framework (June 2021) across its four principles (Governance, Data, Performance, and Monitoring) and implement the associated key practices; (6) distinguish vanity metrics (pilot count, press mentions) from mission-impact metrics (service-delivery outcomes, equity effects, cost savings); (7) implement stratified measurement so disparate impact is visible, not hidden in aggregate; (8) design measurement infrastructure including dashboards, automated alerting, incident tracking, and executive reporting; and (9) communicate AI performance credibly to OMB, GAO, Congress, and the public through structured performance reports, testimony, and the inventory.

Why Measurement Matters

Measurement is the hinge between promise and accountability. A federal AI program that cannot demonstrate measurable outcomes has no defense when OMB questions its budget, when GAO audits its performance, when a committee of Congress demands evidence, or when a disparate-impact complaint reaches HHS OCR, EEOC, or DOJ Civil Rights. The GPRA Modernization Act of 2010 requires agencies to publish performance plans with measurable indicators; the Foundations for Evidence-Based Policymaking Act of 2018 strengthened this through Learning Agendas and Chief Data Officer requirements. For AI specifically, OMB M-24-10 requires agencies to include performance information in the annual AI Use Case Inventory where appropriate. The GAO AI Accountability Framework of June 2021 organizes practices around Governance, Data, Performance, and Monitoring. Failure modes are well documented: the Epic Sepsis Model operated for years at most sites without local measurement, allowing underperformance to persist; Optum's chronic-care algorithm was not measured for equity until academic researchers intervened; Michigan MIDAS lacked outcome measurement of its 93 percent false-positive rate. Each case illustrates what happens when measurement is absent or vanity-focused. Conversely, VA REACH VET's measured outcomes through peer-reviewed publication create a template: a federal AI program that publishes its evaluation, updates it over time, and engages academic partners builds lasting credibility.

Model Performance Metrics

Model performance metrics vary by task. For binary classification: accuracy, precision, recall, F1, area under ROC (AUROC), area under precision-recall curve (AUPRC), balanced accuracy for imbalanced classes. For multi-class: per-class precision/recall and macro/micro F1. For ranking: nDCG, MAP, MRR. For regression: MAE, RMSE, MAPE, R-squared. For generative: BLEU, ROUGE, BERTScore, GPT-eval, human preference; faithfulness and grounding metrics for RAG; task-specific domain evaluations (MedQA, SQuAD, GLUE, MMLU, HumanEval). Calibration should always accompany accuracy: Brier score, expected calibration error, reliability diagrams. For safety-critical systems, operating thresholds matter more than optimal F1; publish the operating point and the downstream workflow. Federal programs should pre-register their evaluation plan including held-out test sets, stratified analysis strata, and acceptance criteria. Avoid benchmarks contamination: do not let vendors train on your test set. Use published academic benchmarks where appropriate but recognize their limitations; real-world deployment performance can differ substantially, as the Epic Sepsis case demonstrated. NIST ARIA evaluations, HELM, BIG-bench, and domain-specific benchmarks complement but do not replace local evaluation.

Fairness and Equity Metrics

Fairness and equity metrics are mandatory for rights-impacting AI under OMB M-24-10, Section 1557 of the ACA for federally-funded health programs, Title VI/VII/IX and ADA for other protected classes. Core metrics include demographic parity (equal positive rates across groups), equalized odds (equal true positive and false positive rates across groups), calibration by group, and group-conditional precision/recall. Famous results from Kleinberg et al. and Chouldechova show that demographic parity, equalized odds, and calibration cannot all be satisfied simultaneously when base rates differ by group; practitioners must choose which metric to prioritize and document the choice. Intersectional analysis is important: a model might be fair across race and across sex separately but unfair for Black women specifically. Subgroup sample sizes must be reported; conclusions on small subgroups are unreliable. Equity measurement also requires upstream data: demographic labels must exist and be accurate, which is its own challenge under federal confidentiality rules. Proxy methods (BISG for race estimation from name and geography) are imperfect substitutes. Standard references include the Aequitas Toolkit, Fairlearn, Google's fairness library, and academic work including Barocas-Hardt-Narayanan. Federal agencies should publish equity metrics alongside aggregate performance; OCR's 2024 Section 1557 Final Rule reinforces this expectation.

Operational Metrics

Operational metrics address system behavior in production. Availability (uptime percentage, SLA attainment), latency (median, p95, p99), throughput (queries per second), error rates (timeouts, 5xx, model errors), drift indicators (input distribution drift, output distribution drift, label drift proxies), retraining frequency, model version rollback count, incident count by severity, mean time to detect (MTTD), mean time to remediate (MTTR), and change failure rate. For generative systems: hallucination rate as measured by sampling and review, refusal rate, jailbreak success rate from red-team metrics, prompt injection detection rate. For agentic systems: task success rate, tool call correctness, goal alignment score, human override rate. Data quality metrics include completeness, consistency, timeliness, freshness, and validity. These operational metrics feed into the NIST AI RMF MEASURE function and support continuous monitoring under M-24-10 minimum practices. Dashboard design should surface key metrics prominently without information overload; use thresholds and alerting rather than visual interpretation alone.

Cost and Efficiency Metrics

Cost metrics address federal stewardship of taxpayer resources. Total cost of ownership should include compute costs (GPU/TPU hours, inference costs per request), data costs (collection, labeling, storage, lineage), platform costs (MLOps, monitoring, governance tooling), workforce costs (federal FTE, contractor hours), and governance overhead. Unit economics can be expressed as cost per decision, cost per case processed, cost per hour of analyst time saved. Return on investment should compare against the baseline (status quo) rather than a hypothetical optimum. Avoid counting hypothetical savings; only realized savings count for federal reporting. The Office of Management and Budget Circular A-11 provides general federal performance-budget integration; OMB Circular A-94 provides benefit-cost analysis methodology. The FinOps Foundation provides broader cloud cost discipline applicable to AI. Vendor cost opacity can be significant; demand itemized billing and understand pricing levers (input tokens, output tokens, context length, model tier). Plan for cost growth as AI usage scales; what is cheap at pilot scale can be prohibitive at agency scale.

Trust and Transparency Metrics

Trust metrics measure whether the AI's users, subjects, and oversight stakeholders trust it. User-side metrics: adoption rate, time to first successful use, user-reported satisfaction (CSAT, NPS adapted), override rate (how often humans override AI). Subject-side metrics: complaint rate, appeal rate, appeal resolution outcome distribution. Oversight-side metrics: GAO and IG finding closure rate, completeness of AI Use Case Inventory, model card currency, data statement currency. Transparency metrics: proportion of use cases with published model card, intended-use statement, fairness report, and red-team summary. Content provenance attachment rate for official communications. Reporting channel accessibility (how quickly do affected parties receive response). Federal agencies can be tempted to conceal trust problems; this invariably backfires when OMB, GAO, or the press discovers them. Mature agencies publish problem statistics and improvement trends; opacity is a GAO audit flag, not a feature.

Mission-Outcome Metrics

Mission-outcome metrics are what agencies ultimately exist to deliver. For VA, reduction in veteran suicide attempts among REACH VET cohort. For IRS, taxpayer service improvements, fraud reduction, and dispute resolution times. For CMS, reduction in improper payments and improvement in beneficiary outcomes. For DoD, mission effectiveness in specific capability areas, with detail classified. For SSA, claim accuracy and processing time. For Social Security disability determinations, concordance with medical reviewers. For USCIS, processing time and approval consistency. For ICE and CBP, lawful and equitable operations. Mission outcomes should be tied to specific agency mission elements and not invented post hoc to justify a deployed AI. Agency strategic plans under GPRA Modernization set mission objectives; AI programs should demonstrably contribute to those objectives. Counterfactual analysis is essential: would the outcome have improved anyway without the AI? Randomized pilots, stepped-wedge rollouts, and quasi-experimental designs help answer this; OMB A-11 Section 270 addresses evaluation. Avoid post-hoc story-telling; pre-register evaluation designs and share results.

GAO AI Accountability Framework

The GAO AI Accountability Framework, published June 2021 as GAO-21-519SP, organizes federal AI accountability around four principles. Governance: establish governance structures to manage risks, including AI governance board, clear roles, and escalation paths. Data: ensure quality, reliability, and representativeness of data used to develop AI. Performance: validate and monitor performance to achieve goals while avoiding unintended effects. Monitoring: continuously monitor and assess AI throughout its life cycle. Each principle has specific key practices; GAO uses the framework during audits. For measurement, the Performance principle and Monitoring principle are central: they specify the kinds of metrics auditors expect to see, including stratified performance, fairness evaluation, reliability testing, drift monitoring, incident response, and documentation of findings and remediation. Agencies that have internalized the framework tend to score better on GAO audits and respond more effectively to Congressional inquiries. Federal program managers should treat the GAO framework as a measurement design reference, not merely as an audit checklist.

Public Reporting and the AI Use Case Inventory

OMB M-24-10 requires agencies to publish an annual AI Use Case Inventory. The 2024 inventory format called for each use case to include intended use, development stage, expected benefits, whether rights- or safety-impacting, and where appropriate, performance indicators. The inventory is published on each CFO Act agency's open data page. As of 2024 inventories the range of quality is wide: some agencies publish detailed, quantitative information; others publish minimum required fields. OMB has iterated the inventory format and will continue to do so. Public reporting raises tensions between transparency and operational security: a fraud-detection algorithm's precise thresholds might aid evasion if published. The standard resolution is to publish intended use, general performance characteristics, and governance arrangements while withholding operational specifics. FOIA handling follows standard federal practice. Agencies should also report to Congress through statutory performance reports, responses to oversight letters, and testimony; the tone should match the evidence. Over-promising AI capability is a recipe for later-stage credibility loss.

Measurement Anti-Patterns

(1) Aggregate Only: reporting aggregate accuracy while hiding stratified performance. (2) Vanity Metrics: counting pilots, press releases, slide decks. (3) Benchmark Tunneling: optimizing for public benchmarks that don't match production. (4) Pilot Inflation: endless pilots inflate activity statistics without mission impact. (5) Missing Counterfactual: claiming causation from correlation without baseline comparison. (6) Measurement-Free Go-Live: deploying without pre-specified acceptance criteria. (7) Ghost KPIs: KPIs that are never updated after the initial plan document. (8) Post-hoc Framing: inventing metrics after deployment to justify outcomes. (9) Opaque Vendor Metrics: accepting vendor-claimed performance without independent verification. (10) Under-Reporting Incidents: classifying problems as noise to keep dashboards green. (11) Dashboard Theater: beautiful dashboards disconnected from actual decisions. (12) Measurement Without Action: metrics exist but nobody acts on them. GAO and IGs look for these patterns; so should internal program managers. A mature AI program audits itself against this list quarterly.

Summary and Next Steps

Measurement is the hinge of accountability. Six metric families (model performance, fairness and equity, operational, cost and efficiency, trust and transparency, mission-outcome) cover the AI measurement space. GPRA Modernization Act, Foundations Act, OMB M-24-10, NIST AI RMF MEASURE function, and GAO AI Accountability Framework structure federal practice. The annual AI Use Case Inventory and performance reports are the primary public artifacts. Avoid the twelve anti-patterns including aggregate-only reporting, vanity metrics, benchmark tunneling, and measurement-free go-live. The next lecture, Moving from Pilot to Production, applies measurement discipline specifically to the production transition, where most federal AI programs fail.