The number that matters in the USDA Inspector General's report is not 82. It is 73.
Eighty-two federal artificial intelligence use cases were catalogued. Seventy-three of them were running in production without an Authority to Operate โ no signed risk acceptance, no accreditation boundary, no name attached to the file. Nine had paperwork. The rest were ghosts wearing government badges, executing procurement logic and benefits determinations inside systems no compliance officer had ever entered into a registry.
On October 19, 2026, Google Public Sector and Accenture Federal Services announced Gemini Launchpads, a program to industrialize federal AI agent deployment, backed by a thousand forward-deployed engineers and a physical prototyping lab called The Forge in Washington. Karen Dahut signed from Google's side; Ron Ash from AFS, a fifteen-thousand-person subsidiary nested inside an eight-hundred-thousand-person Accenture.
It is an impressive announcement. And it cannot see the seventy-three. No tool can. That is the entire story, and almost nobody is telling it.
Launchpads is best understood not as a product but as a delivery vehicle. There is no price list, no subscription tier, no self-service console anywhere in the material. What exists is a template library of pre-built agents for mission-support functions โ human resources, finance, supply chain โ wrapped in a governance scaffold mapped to NIST AI RMF, FIPS 140-3, and NIST SP 800-53 Rev.5, sitting inside Google Cloud's FedRAMP High boundary.
The scaffolding is the interesting part, not the agents. FedRAMP authorization for a cloud service provider runs twelve to thirty-six months and, by the figures in circulation, more than three and a half million dollars. The Consolidated Rules of 2026 were published on June 24, 2026, and take effect January 1, 2027. Six months. A process that structurally demands one to three years was handed half a year to comply. That contradiction is the actual news buried beneath the press release, and it guarantees that most existing deployments will pursue remediation rather than native compliance.
The technical heart of the program โ the one section worth reading twice โ is its insistence that an AI agent be governed as a Non-Person Entity. Four requirements: enterprise-grade identity management, time-bound scoped credentials, independent revocability, and operation-level audit logging recording which agent touched which data, performed which action, and which human approved it.
This is correct, and it is underrated. Most enterprise agent deployments in 2026 still run on a shared service account with standing privileges and a log that says nothing. I have spent enough hours inside that wreckage to recognize its silhouette. In 2017, auditing a DAO successor project in Zurich, I found a reentrancy path worth 500 ETH and watched the frontend team reject my report as too academic. The lesson was never about Solidity. It was that a system's identity model determines its failure mode long before any specific bug does. Identity is a protocol; soul is the private key.
Here is the ceiling, though. Google's FedRAMP High authorization covers its infrastructure โ the IaaS and PaaS layers. It does not cover your application. Each agency's Authorizing Official signs each specific deployment, and that signature is a human organizational act, not a configuration step. Launchpads can compress engineering effort. It cannot compress the sequence of internal risk assessment, security review, and AO signature, because that sequence is not a technical process at all. It is a chain of accountability with a person's name at the end of it.
The program is ATO-ready. It is not ATO-granted, and the distance between those two phrases is measured in quarters, not sprints.
Then there is the question nobody asked at the launch: where does the audit chain live? If an agent's operation-level telemetry sits on Google's side of the boundary, then migrating away later means abandoning the compliance history itself. That is a far more severe dependency than model lock-in, because model lock-in costs you money while audit lock-in costs you your accreditation. The audit is not a check; it is a confession. And a confession you cannot carry with you is a confession you do not own.
The economics have a silent second ceiling. Agent workflows are not single-turn queries. A planning, tool-calling, reflection loop can consume five to twenty times the tokens of a comparable question-and-answer interaction. At frontier pricing, a single multi-step task can cost dollars. Run that against a back office processing hundreds of thousands of tickets monthly and the automation bill can exceed the payroll it replaced. No federal contract I have seen prices this properly, and Launchpads does not appear to either.
And the competitive framing is incomplete. The forward-deployed engineer model, the compliance wrapper, the mission-level agent โ Palantir has been selling that configuration to the federal government for nearly twenty years, with AIP already holding signed authorities inside defense and law enforcement environments. In the code, I found the ghost of the architect โ and the architect here is not in Mountain View. Gemini Launchpads is a follower entering an incumbent's territory, armed with cheaper inference and a channel partner, which is a real advantage and also an admission.
Which brings me to the contrarian reading. The source material frames the federal AI problem as an infrastructure bottleneck: not enough tooling, not enough pre-built components, not enough compliant scaffolding. That framing is vendor-friendly in a way that should make any auditor uneasy, because it redefines an incentive problem as a procurement problem. The seventy-three agents are not missing tools. They are missing consequences. An agency that deploys without authorization faces no measurable penalty for doing so, and a program that ships templates does not change that arithmetic. When the pool empties, only the intent remains โ and the intent of a compliance tool sold by a cloud provider is not always the intent of the compliance officer who signs.
There is an irony worth naming. Non-human identity governance is a problem the crypto industry confronted first and solved badly. Soulbound tokens were proposed three years ago as the answer to persistent, non-transferable on-chain identity, and they went nowhere, because nobody wants their credit record welded to a wallet forever. The federal government is about to discover the same truth at a scale of millions of employees: durable machine identity is easy to issue and nearly impossible to revoke cleanly. Web3 did not fail at this because the technology was immature. It failed because the incentives were wrong.
The question I keep returning to is not whether Launchpads ships. It is what happens the first time an autonomous agent makes a determination that a judge or an Inspector General overturns โ a wrongly denied benefit, a badly scoped procurement โ and the audit trail shows a human approval that was, in practice, a rubber stamp. Who signs then? The agency, the integrator, or the model provider? No contract clause I have read answers that, and no Federal Acquisition Regulation clause does either. Until one does, a thousand forward-deployed engineers are accelerating a vehicle toward a wall nobody has drawn yet.