Google puts its Gemini agent to work on FOIA requests and grant audits

Google Cloud has put Gemini for government services into preview. It points the company’s Gemini agent at public-sector work such as FOIA record searches, redaction and bid evaluation, with a FedRAMP High deployment option, according to CEO Thomas Kurian.

Seven skills, nine connectors and two partner agents

Google says the product has five parts: skills, connectors, partner agents, a governed control plane and a partner ecosystem. The skills are reusable instructions written by domain experts that apply the statute, regulation and agency guidance behind a task. Google lists these tasks:

• Policy and regulatory research, with page and paragraph citations from the CFR, the Federal Register and state codes
• FOIA and state Sunshine Law record discovery across SharePoint, Drive and email archives
• Batch redaction, with each proposal labeled by statutory exemption and an exemption log produced after a reviewer confirms
• Vendor bid evaluation against RFP requirements and FAR guidelines
• Grant and program expense audits against 2 CFR 200 Uniform Guidance
• Document intake checks on items such as pay stubs, W-2s and leases
• Plain-language public notices built from policy memos

The connectors cover the Federal Register, regulations.gov, eCFR, Oracle E-Business Suite, Oracle PeopleSoft, Salesforce, ServiceNow, Google Workspace and Microsoft 365 with SharePoint. Google says Gemini queries each system in place, so nothing is migrated, and access follows the permissions the agency already has.

Two pre-built agents from SZNS Solutions, one for FOIA and public records discovery and one for document redaction, are the partner agents the post describes. Google says a person approves each consequential step. Accenture Federal Services, Deloitte, SAIC, Slalom and nine other firms are named as integration partners.

On governance, Google says security teams can enforce VPC Service Controls, CMEK and Assured Workloads across all agents. Outputs are grounded in source documents with citations. Customer data is not used to train Google’s foundation models without permission. The Assured Workloads FedRAMP High Data Boundary option enforces U.S. data residency and personnel controls.

Who can try it, and what remains unstated

The product is in preview as of October 20, 2026, and Google invites government leaders to explore it. The post gives no price, no general availability date and no list of regions beyond the U.S. data boundary. Google says it is building the product with a limited set of federal and state customers, and it names Missouri’s state government and the federal health department as two of them, but it reports no results from either.

The only customer voice is Tim Marczewski of the Missouri Office of Administration, who said the state is “delivering faster, better services to Missourians” by automating routine work. Claims such as shortening FOIA discovery that “often takes weeks” are Google’s own and are not independently measured in the post. Nor does the post say how accurate the redaction or bid-scoring skills are. Because the agent proposes and a human confirms, error rates matter for agencies deciding how much review time they can actually save.

For readers who followed Google’s earlier government push, this is a different layer. It adds task-specific skills and agents on top of general access to the model.

A pricing question the preview leaves open

Cost is the main gap. In August 2025, Google and the U.S. General Services Administration agreed a OneGov deal that, according to TechRepublic, set a price of $0.47 per agency for Gemini for Government tools. A reseller page cited in the same search lists that promotional pricing as running through September 30, 2026, so the terms may have changed. This new product’s pricing is not covered by anything published so far.

Competitors have used price to win federal seats. Microsoft’s blog said its OneGov agreement included Microsoft 365 Copilot at no cost for up to 12 months for existing G5 users. A licensing blog reports Copilot is also available in GCC High. Against free or near-free entry offers, Google’s pitch is different: it says that because it runs the whole stack, it can route high-volume parsing to cheaper models and keep high-reasoning models for complex policy analysis, so spending does not grow as fast as workload.

That claim cannot be tested yet. Agencies will want to know how usage of the specialized skills is metered, what the partner agents cost on top, and whether the discounted federal terms carry over. The Microsoft figures come from the company’s blog, and the reseller figures are marketing claims that should be checked with GSA.