Federal AI is scaling faster than its ability to learn

The federal AI conversation has quietly changed. For two years the question was whether agencies could get access to capable models and the compute to run them. That question is largely settled. Centralized purchasing vehicles, governmentwide agreements, and hyperscaler contracts have pushed AI into the mainstream of federal operations, and recent oversight reporting shows agency AI use roughly doubling in a single year. Access is no longer the constraint.

The new constraint is memory. As agencies increasingly consume AI as a service—paying for capability on an ongoing basis rather than buying a system once—the knowledge about what actually works, where it fails, and what it costs is being generated faster than any agency is capturing it. The Government Accountability Office has been direct about this: agencies are not systematically collecting lessons learned, and it has recommended that major departments update their policies to require it. That is a plain-language warning that the government is spending at scale while learning by accident.

Buying AI is easy now. Understanding it is not.

The current wave of consolidation—governmentwide purchasing initiatives, multibillion-dollar cloud awards, and large platform modernizations—was designed to reduce friction, and it has. But friction was never the whole problem. The harder problem is that AI deployed inside a mission is not a static asset. Its behavior drifts, its costs move with usage, and its value depends on context that a contract line item cannot describe. When an agency treats an AI service like a commodity purchase, it optimizes for the transaction and neglects the operating knowledge that makes the transaction worthwhile.

This is where the "as a service" model cuts both ways. Consuming AI as a service lets an agency move quickly and avoid owning depreciating infrastructure. It also means the most important information—how the model performs against real mission data, which guardrails held, what the true cost per outcome was—lives in the seams between the vendor, the program office, and the end users. If no one owns the job of pulling that information back into the institution, it evaporates when the period of performance ends.

Access to AI is now abundant. Institutional judgment about AI is still scarce—and scarcity is where risk concentrates.

The FY2026 turnover is a forcing function

Several prominent AI and cloud arrangements are set to turn over around the end of the fiscal year. Turnover is normal and often healthy; it introduces competition and resets pricing. But a contract boundary is also the moment when institutional memory is most fragile. Teams disperse, vendor staff rotate off, and the tacit knowledge accumulated over a period of performance walks out the door unless it has been deliberately written down. An agency that has not been capturing lessons throughout the engagement discovers, at exactly the wrong time, that it cannot tell a successor team what it already learned.

The agencies that will navigate this turnover well are not the ones with the most sophisticated models. They are the ones that treated every deployment as a source of reusable knowledge—recording assumptions, decisions, failure modes, and cost realities as they went. For them a contract transition is a handoff. For everyone else it is amnesia.

Build the operating loop, not just the pipeline

The instinct in a moment like this is to invest in more capability: another platform, another model, another vehicle. The higher-return move is to invest in the loop that converts deployed AI into institutional learning. That loop is unglamorous and entirely within an agency’s control. It does not require new authorities or a larger budget—it requires discipline and ownership.

• Name an accountable owner for lessons learned on every AI engagement, with the explicit job of pulling knowledge back into the agency before the period of performance ends.
• Instrument outcomes, not just uptime—measure cost per useful result and model performance against real mission data, and treat those numbers as program artifacts.
• Write down decisions and their rationale as the work happens, so a successor team inherits judgment rather than a blank slate.
• Make failure modes reportable and shareable across programs, so the same expensive mistake is not independently rediscovered by three agencies.
• Structure contracts and transitions so knowledge transfer is a deliverable, not a courtesy.

None of this slows delivery down. In practice it does the opposite: a team that can see how its AI is actually performing makes faster, safer decisions and wastes less effort relitigating settled questions. The operating loop is what turns a series of disconnected buys into a compounding capability.

Where GovArc sits on this

Our view is consistent with how we approach every modernization: the mission is modernized when the operating model changes, not when a new tool arrives. AI raises the stakes because it scales both value and error quickly, and because so much of the important information now lives outside the agency by default. Federal leaders do not need to slow their AI adoption. They need to pair it with a deliberate learning loop so the government owns the judgment its own spending is generating.

The agencies that get this right over the next year will not necessarily be the ones that bought the most AI. They will be the ones that remembered what it taught them.