The Great Reinvention logo THE GREAT
REINVENTION

The Paperwork Dividend: Modernizing Federal Government with Intelligent Automation

Proposed legislation: The Government Modernization and Intelligent Automation Act

Modernize Government with AI: Cost-Benefit Analysis

The federal government runs on a strange mixture of cutting-edge missions and antique machinery. The same agencies that operate spacecraft and intelligence satellites also process benefits on decades-old mainframes, route paper forms through manual queues, and reconcile payments using systems that cannot reliably talk to one another. The Internal Revenue Service spent years operating some of the oldest IT systems in the federal government. Social Security, Medicare, Medicaid, and veterans' benefits all depend on layered legacy software that is expensive to maintain and slow to change. The human and fiscal cost of this paper-and-mainframe bureaucracy is enormous — measured in delayed benefits, frustrated taxpayers, billions in improper payments, and armies of staff doing work that software now does better.

The proposition behind modernizing government with artificial intelligence is straightforward: replace manual, paper-based, error-prone processes with intelligent automation for benefits administration, tax processing, records management, and fraud detection. This is not science fiction. Federal agencies are already deploying these tools and already reporting measurable savings. The question this analysis addresses is whether the proposal's target of $100–150 billion in annual savings is realistic, where the savings genuinely come from, and what it would take — honestly — to capture them.

The verdict, developed below, is that the direction is strongly supported by evidence, the early results are real and growing, but the specific $100–150 billion annual figure sits above what can currently be documented and should be presented as a multi-year ambition rather than a near-term certainty.

Where the Savings Come From

Catching Improper Payments and Fraud

The single largest and best-documented savings opportunity is in payment integrity. GAO reported an estimated $162 billion in federal improper payments in fiscal year 2024 (rising to about $186 billion in FY2025), concentrated heavily in a handful of large programs — Medicare, Medicaid, the Earned Income Tax Credit, and SNAP among them. Separately, GAO estimates the government loses between $233 billion and $521 billion annually to fraud. Cumulative improper payments since FY2003 total roughly $2.8 trillion.

AI is already biting into this. The U.S. Treasury reported that AI-enhanced fraud-detection efforts — including its "Do Not Pay" system and related programs — prevented and recovered more than $4 billion in fiscal year 2024, and Treasury later reported that its improper-payment and fraud-prevention work reached $11.7 billion in a subsequent year, a roughly 63 percent increase over the prior year's $7.2 billion. These are official, agency-reported figures, and they demonstrate that machine-learning pattern detection finds improper payments that manual review misses. Crucially, the trajectory is steeply upward as the tools mature and spread across agencies.

Closing the Tax Gap

The IRS estimates a substantial annual "tax gap" — the difference between taxes owed and taxes paid on time — running into the hundreds of billions of dollars per year. GAO has reported that AI could materially help the IRS narrow this gap by improving audit selection, detecting non-compliance, and identifying questionable returns. As of recent reporting, the IRS had on the order of 126 active AI use cases across taxpayer services, operational efficiency, and compliance. Even modest percentage improvements in compliance translate into tens of billions of dollars, because the base is so large. This is revenue recovery rather than spending reduction, but it produces the same fiscal effect: a smaller deficit.

Automating Administrative Labor

Beyond fraud and revenue, intelligent automation reduces the cost of routine administrative work: processing benefit applications, answering taxpayer and claimant questions, adjudicating straightforward claims, digitizing and searching records, and reconciling payments. Tasks that once required large clerical workforces and generated long backlogs can increasingly be handled by automated systems with human oversight reserved for complex or contested cases. The savings accrue both as lower processing cost per transaction and as fewer costly errors and appeals downstream.

Reducing Legacy IT Maintenance

The federal government spends the large majority of its tens of billions in annual IT budget simply maintaining legacy systems rather than improving them — a long-standing GAO finding. Modernization that retires brittle legacy code reduces this maintenance drag over time, though, importantly, it requires up-front investment before it produces net savings.

Projected Figures and the Realistic Range

The proposal claims $100–150 billion per year. Here is the honest accounting.

The opportunity space clearly exceeds that figure: improper payments alone run $160–186 billion annually, fraud losses are estimated at $233–521 billion, and the tax gap adds hundreds of billions more. If AI could eliminate even a third of improper payments and meaningfully narrow fraud and the tax gap, $100–150 billion in combined savings-plus-recovered-revenue would be within reach in principle.

But "in principle" is doing real work in that sentence. The documented AI-driven savings to date are an order of magnitude smaller — Treasury's reported $11.7 billion, plus IRS compliance gains and scattered agency efficiencies. The gap between the opportunity and the realized savings reflects hard constraints: not all improper payments are recoverable (many are documentation errors, not lost dollars), AI cannot prevent every fraud, modernization takes years, and legacy-system replacement is notoriously slow and risky in government.

The fair conclusion: the $100–150 billion annual figure is best understood as a ceiling that becomes plausible over a decade if modernization is pursued aggressively across all major programs, not as a savings rate achievable in the first few years. Today's evidence supports tens of billions in annual savings-and-recovery and a clear upward trajectory. Presenting $100–150 billion as a near-term certainty would overstate the case; presenting it as a credible multi-year target, grounded in the size of the improper-payment and tax-gap problems, is defensible.

Mechanism: How Modernization Would Work

The proposed Government Modernization and Intelligent Automation Act would establish a dedicated modernization framework with three pillars. First, it would create a government-wide payment-integrity platform that extends the Treasury "Do Not Pay" and AI fraud-detection model across all major benefit and payment programs, with data-sharing authorities (subject to privacy safeguards) so that systems can cross-check eligibility and detect anomalies before money goes out the door — prevention being far cheaper than recovery.

Second, it would fund a multi-year legacy-system modernization program with strict accountability for delivery, learning from the failures of past government IT projects by favoring incremental, modular procurement over monolithic "big-bang" replacements. Third, it would scale AI-assisted taxpayer and claimant services and AI-augmented compliance at the IRS and benefit agencies, with mandatory human oversight, transparency, and appeal rights built in.

A capital-investment fund — modeled on the existing Technology Modernization Fund concept — would provide the up-front money that modernization requires, repaid from the savings it generates.

Administrative and Implementation Considerations

This proposal is unusual in that its implementation cost is significant and front-loaded. Realizing AI savings requires investment in software, data infrastructure, cybersecurity, and workforce reskilling before net savings appear. The IRS modernization effort, funded with tens of billions in dedicated money, illustrates both the scale of investment required and the long timelines involved.

Three implementation realities deserve emphasis. First, government IT modernization has a poor historical track record; GAO's "high-risk" list has long featured federal IT management. Success requires disciplined program management, modular procurement, and willingness to halt failing projects early. Second, data quality and interoperability are prerequisites — AI cannot detect anomalies in data it cannot access or trust, so investment in clean, connected data is foundational. Third, workforce transition must be managed humanely; automation will change many federal jobs, and reskilling and attrition-based downsizing are preferable to abrupt displacement.

Finally, AI in government raises legitimate governance concerns — bias, error, due-process protections, and transparency in automated decisions. Any modernization statute must mandate human review of consequential determinations, auditability of AI systems, and clear appeal channels.

International Comparisons and Precedent

Several governments have demonstrated what aggressive digitization can achieve. Estonia's nationwide digital government allows citizens to file taxes and access services in minutes, dramatically lowering administrative cost per transaction and is frequently cited as a model. The United Kingdom's HM Revenue & Customs has invested heavily in digital tax administration. India's Aadhaar-linked direct-benefit-transfer system, while controversial on privacy grounds, has been credited by its government with curbing leakage and duplicate payments at very large scale. These cases show both the promise — large efficiency and integrity gains — and the cautions — privacy, exclusion, and design risks — that accompany government digitization.

Comparison to the Status Quo and Alternatives

The status quo is steady-state legacy maintenance: agencies spend most of their IT budgets keeping old systems alive, improper payments persist at $160 billion-plus per year, and the tax gap goes substantially uncollected. This is not a stable equilibrium so much as a slow, expensive decline, with rising maintenance costs and growing security risk.

The main alternative to AI modernization is incremental staffing — hiring more humans to review payments and process claims. But manual review is precisely what allows $160 billion in improper payments to slip through; humans cannot match machine pattern-detection at scale, and labor costs scale linearly while software scales far better. Another alternative is simply tightening eligibility rules, which can reduce improper payments but at the cost of also denying legitimate claimants — a blunter and less equitable tool than better detection.

Risks, Trade-offs, and Counterarguments

The strongest counterargument is the government's dismal track record on large IT projects. Skeptics can point to a long list of expensive federal modernization efforts that ran over budget, fell behind schedule, or failed outright. This is a fair and important warning. The proposal's response is to insist on modular, incremental procurement with kill-switch accountability — but the risk that modernization money is wasted is real and must be acknowledged.

A second objection concerns privacy and civil liberties. A government-wide system that cross-checks eligibility data across programs is powerful precisely because it links information — and that linkage creates surveillance and breach risks. Robust statutory privacy protections, data minimization, and independent oversight are not optional add-ons but conditions of legitimacy.

A third counterargument is that AI savings are easy to claim and hard to verify. Agency-reported "savings" sometimes count prevented payments that might never have occurred, or recoveries that would have happened anyway. Honest accounting requires independent validation of claimed savings, ideally by GAO or agency Inspectors General.

A fourth trade-off is workforce disruption. Automating administrative work affects real federal employees and the communities that depend on those jobs. Managing this transition fairly is both a moral obligation and a political precondition for success.

Finally, AI errors carry due-process stakes: an algorithm that wrongly flags a legitimate beneficiary as fraudulent can cause genuine harm. Human oversight of consequential decisions is essential.

Conclusion

Modernizing the federal government with intelligent automation is one of the few savings proposals where early, official results already point in the right direction: Treasury's AI fraud tools recovering billions, the IRS deploying scores of AI use cases, and a documented opportunity space — $160 billion-plus in annual improper payments, hundreds of billions in fraud and the tax gap — that dwarfs the proposal's target. The strategic case is strong, and the trajectory is clearly upward.

Candor requires acknowledging that the $100–150 billion annual figure is a multi-year ceiling, not a near-term guarantee: documented savings today are an order of magnitude smaller, modernization is expensive and slow, and not all improper payments are recoverable. Pursued with disciplined procurement, strong privacy and due-process safeguards, independent verification of claimed savings, and humane workforce transition, AI modernization can convert one of government's oldest weaknesses — its paperwork — into one of its largest sources of recurring savings.

Sources

← Back to The Great Reinvention