The agencies that moved first on cloud AI are not waiting for approval anymore. The agencies still evaluating are running out of runway.

By Shashi Bellamkonda · April 30, 2026

Somewhere in Sacramento, a Covered California applicant submitted proof of income and received a response in seconds. A year ago, that same interaction would have triggered a wait of up to three weeks, a follow-up call, and in many cases a second document submission. The change was not a new policy. It was a Document Artificial Intelligence system from Google Cloud, implemented with Deloitte, running inside CalHEERS, the state's health insurance eligibility and enrollment platform. That deployment was announced at Google Cloud Next '26, and it is the most operationally credible government AI result Google has produced inside a regulated workflow to date.

Karen Dahut, Chief Executive Officer of Google Public Sector, described Next '26 as the arrival of the agentic era for government, arguing that a fully integrated AI stack spanning infrastructure, models, data management, and agents is what makes transformation at mission scale achievable now. CalHEERS is the evidence she brought to that argument.

A regulated workflow with fraud at stake is the hardest test to pass

The CalHEERS system handles eligibility verification for millions of Californians, including residents with limited or no credit history. Document AI now automates verification across 25 document types, from income records to residency proof, routes ambiguous cases to human specialists, and runs on a FedRAMP-certified foundation that satisfies the compliance requirement that blocked cloud AI adoption in state programs for years.

The system is trained to detect fraud markers that manual review under volume pressure tends to miss: missing signatures on scanned documents, mismatched dates, subtle formatting inconsistencies. It does not auto-reject. It flags for specialist review, which satisfies due process requirements and protects equitable access for applicants whose documentation is legitimate but irregular.

The 40% reduction in manual tasks is a vendor-reported figure. The compression from a three-week wait to real-time on-screen feedback is harder to dispute, because it represents an architectural change with a before and after that staff and applicants both experience directly.

Fraud, waste, and abuse reduction inside a health eligibility system maps directly to a line item a state legislature will recognize and fund. That is what makes CalHEERS a replicable procurement story, not just a case study.

Austin shows the platform moving into a different problem class entirely

The City of Austin use case sits in a different part of the portfolio. Extreme heat causes more fatalities in Austin than all other extreme weather events combined. The city is working toward 50% citywide canopy cover by 2050, but shade is not distributed equally. Lower-income communities east of Interstate 35 carry a disproportionate share of the heat burden and have significantly less tree cover than wealthier neighborhoods to the west.

Austin is using high-resolution aerial imagery and tree canopy data from Google's Environmental Insights Explorer, combined with Google Earth Engine's pre-processed datasets, to identify urban heat islands and direct planting resources to the neighborhoods that need them most. Three out of every five trees distributed go to high-priority, low-canopy areas. The city uses the Tree Equity Score, a national metric developed by the nonprofit American Forests and informed by Google's canopy analysis, alongside its own internal prioritization process.

Large-scale tree canopy modeling has historically been cost-prohibitive at the municipal budget level. Google Earth Engine makes it accessible. The first use case is rarely the last one, and Austin illustrates the pattern: a city solves one resource-intensive data problem with a cloud platform, and the infrastructure is then available for the next problem without starting the procurement process over.

Agent Designer is how the footprint grows without a new procurement cycle

Google Public Sector announced that Gemini 3.1 Pro, its most capable reasoning model, is now available through Gemini for Government. Government buyers have watched commercial AI capabilities run ahead of their own access for three years, constrained by compliance certification timelines. Closing that gap inside a sovereign deployment environment removes a standing objection that has stalled procurement conversations in regulated agencies.

The more consequential announcement is Agent Designer, a feature inside the Gemini Enterprise App that lets agency employees, not just developers, build trigger-based and schedule-based agents. Department of Defense civilian and military personnel are already using it to build agents for unclassified tasks on GenAI.mil. The tool includes inspection and approval workflows before agents go into production.

This is where the entrenchment dynamic becomes visible. Once a platform is certified, contracted, and running a core workflow like eligibility verification, expansion happens at the employee level rather than through procurement. An employee building an agent for a scheduling task inside an existing Gemini for Government environment does not trigger a new vendor evaluation. The platform is already there. The agent is an extension, not a decision.

That is how cloud AI becomes embedded in government operations. Accumulated workflow dependencies, each individually small, each making the next one easier to build.

Infrastructure is the long-duration bet agencies are implicitly making

Google announced eighth-generation tensor processing units at Next '26, with TPU 8t for training and TPU 8i for near-zero latency inference, alongside Virgo Networking, a megascale data center fabric for high-performance workloads. All seventeen U.S. Department of Energy national laboratories are using an AI co-scientist built on tensor processing units for the Genesis Mission. Infrastructure at that scale shapes the cost basis for inference across multi-year contract durations. An agency selecting a cloud AI partner today is also selecting whose infrastructure roadmap it depends on five years from now.

About Shashi Bellamkonda

Loading bio…

LinkedIn Profile

Get new posts by email:

Subscribe

Powered by