The Pilot Numbers Look Great. That's the Problem.
Every agency AI pilot produces impressive numbers. Document processing time down 60%. Citizen inquiry resolution up 45%. Workforce training completion rates doubled.
Those numbers are real. They're also misleading.
Pilot economics operate in a controlled environment with vendor attention, discounted pricing, and limited scope. Scaling economics operate in the real world — with government procurement constraints, budget cycles, FISMA requirements, and Inspector General oversight.
The agency that approves a multi-year contract based on pilot ROI is making a financial decision using data from a fundamentally different operating context.
Why Pilot ROI Misleads in Government
In the private sector, a pilot that shows 3x ROI leads to rapid scaling. If the vendor relationship doesn't work out, the company switches. Migration costs are real but manageable.
Government doesn't work that way.
Federal procurement cycles run 12-24 months for competitive acquisitions. Sole-source justifications for contract extensions face IG scrutiny. Budget authority is annual, allocated across fiscal years, and subject to continuing resolution uncertainty.
Once an agency scales an AI deployment, switching vendors isn't a quarter-long project. It's a multi-year procurement action that may require a new ATO, data migration with chain-of-custody documentation, retraining of staff across multiple divisions, and congressional notification if the system touches citizen-facing services.
The switching cost in government isn't a line item. It's a mission risk.
This means the pilot ROI calculation — which implicitly assumes the agency can leave if the math stops working — is fundamentally flawed. In government, scaling an AI deployment is closer to a permanent infrastructure decision than a renewable subscription.
Contract Dependency as Mission Risk
Here's the scenario that keeps government CIOs awake.
The agency deploys an AI platform for workforce training across 15,000 employees. The platform processes training records, generates personalized learning paths, tracks certification compliance, and produces reports for congressional oversight.
Two years in, one of three things happens.
The vendor raises prices. The per-user fee increases 40% at renewal. The agency's budget can't absorb the increase, but the workforce training mission requires the platform. Negotiation leverage is near zero because switching would disrupt training for 15,000 employees during a compliance certification cycle.
The vendor is acquired. The acquiring company shifts the product roadmap toward commercial enterprise clients. Government-specific features — GovCloud deployment, PIV authentication, FISMA-compliant logging — move to a premium tier or a "legacy" product line that receives minimal investment.
The vendor exits government. Margins are lower, compliance requirements are higher, and sales cycles are longer. The vendor decides the commercial market is more attractive. The agency has 12 months to find an alternative, migrate data, achieve a new ATO, and retrain staff.
None of these scenarios are hypothetical. All three have happened repeatedly in government IT over the past decade. AI platforms are not exempt from the same market dynamics.
What Agency CIOs Need to Understand About AI as Infrastructure
The CIO question isn't "which AI tool should we buy?" It's "is AI a tool we use or infrastructure we operate?"
For citizen services, it's infrastructure. AI that processes benefit applications, responds to public inquiries, or assists in adjudication decisions is embedded in the agency's mission delivery. Renting that capability from a vendor who can change terms is operationally equivalent to renting your case management system — technically possible, strategically indefensible.
For workforce training, it's infrastructure. AI that delivers personalized training, tracks certifications, and generates compliance reports for the Office of Personnel Management isn't a nice-to-have tool. It's a core operational system.
For FOIA processing, it's infrastructure. AI that classifies documents, identifies responsive records, and flags exemptions touches the agency's legal obligations to citizens. The system's accuracy and auditability are subject to judicial review.
When AI is infrastructure, the ROI framework needs to account for what happens when someone else controls it. Vendor dependency in infrastructure isn't a procurement risk. It's a mission risk.
The Expanded ROI Framework for Government
The standard AI ROI calculation looks like this: efficiency gained multiplied by labor cost, minus license and implementation cost. This produces a positive number for almost every deployment, which is why every vendor deck includes one.
Government needs three additional variables.
Dependency cost. What does it cost when the vendor raises prices and the agency can't switch within the current fiscal year? What does it cost when the vendor deprecates a feature the agency depends on for compliance reporting? Multiply the probability of each event by its mission impact over a five-year period.
Compliance maintenance cost. Government AI platforms require ongoing ATO maintenance, FISMA assessments, continuous monitoring, and audit support. If the vendor controls the platform's architecture, the agency depends on the vendor's cooperation for every compliance activity. That cooperation is not guaranteed — especially after contract disputes or acquisition.
Exit cost. What does it cost to leave? In government, exit cost includes procurement timeline for a replacement, ATO achievement for the new system, data migration with full chain-of-custody documentation, staff retraining, and potential service disruption during transition.
When agencies run this expanded framework, the math shifts dramatically.
A platform that costs more upfront but includes source code access, self-hosted deployment, and LLM agnosticism often has a lower total cost of ownership over a five-year FYDP than a "cheaper" per-seat subscription.
This is the calculus behind ibl.ai's approach to government deployment — agencies own the platform, deploy it in their own GovCloud or on-premises environment, and maintain operational independence. The pilot ROI matters. The ownership ROI is what survives contact with budget cycles, IG audits, and vendor market dynamics.
Avoiding the Dependency Cycle
Government IT has lived through this cycle before. Agencies adopted enterprise platforms for email, HR, financial management, and case processing. Early results were positive. Then the vendor relationship shifted — through acquisitions, pricing changes, or strategic pivots — and agencies found themselves locked in.
AI is following the same trajectory, but with higher stakes. Email is important. AI that processes citizen data, makes or assists decisions affecting benefits, and touches mission-critical workflows is categorically more sensitive.
The agencies that will extract real, sustainable ROI from AI are the ones that treat it as infrastructure they own and operate — not as a service they rent from a vendor whose incentives will eventually diverge from the mission.
The pilot numbers are encouraging. The question is what you own when the pilot ends.
The Decision That Matters
Agency CIOs face a choice that most vendor conversations obscure.
You can optimize for speed — choose the SaaS tool with the best demo, deploy quickly, and show results to leadership within a quarter. The ROI will look excellent. It will also lock the agency into a dependency that the standard ROI calculation doesn't capture.
Or you can optimize for durability — choose a platform you can own, deploy in your own environment, and operate independently. The implementation takes longer. The upfront investment is higher. The five-year total cost is lower, the compliance posture is stronger, and the agency retains the ability to adapt as missions evolve.
The real ROI of government AI isn't in the pilot. It's in what the agency controls when the vendor relationship inevitably changes.