Google is teaching Gemini government procedure. The design question is where a draft becomes a decision

A statute is not a prompt. It has definitions in one section that change the meaning of a word in another, exceptions that override the general rule, effective dates, and agency guidance that interprets it differently from the plain text. Government work is applying all of that, case by case, and leaving a record that explains why.

That is the problem Google set out to package on October 20 with Gemini for government services, announced by Google Cloud CEO Thomas Kurian and available in preview.

How Google structured it

The offer has five parts:

• Skills for government work: reusable instructions and context written by domain experts, each applying the statutes, regulations and agency guidance that govern a task.
• Connectors to agency systems: case management, ERP, HR, collaboration tools and federal data sources such as the Federal Register, regulations.gov and the eCFR, queried in place without migrating data.
• Partner agents from public sector software providers, such as SZNS Solutions’ FOIA discovery and redaction agents.
• A governed control plane that applies policy across agents, with VPC Service Controls, customer-managed encryption keys and Assured Workloads, including a FedRAMP High data boundary.
• Grounded outputs with citations to source documents.

Access follows the role-based and record-level permissions agencies already have. Google says agency data and outputs are not used to train its foundation models without permission. The State of Missouri and the U.S. Department of Health and Human Services are named as working with Google to validate the capabilities.

Draft, proposal or decision

The most useful way to read the announcement is to sort its example workflows by what the agent’s output actually is. Google’s own descriptions already draw the line in most cases.

Research drafts. Policy and regulatory research returns answers from the CFR, the Federal Register, state administrative codes and internal manuals, with page and paragraph citations, including which rule governs when federal and state rules both apply. That last part is legal interpretation. A citation lets a reviewer check the source; it does not make the conclusion correct. The output is a draft for an analyst or counsel.

Proposals that need confirmation. FOIA discovery gathers responsive records into a review set for the records officer and counsel. Redaction proposes redactions labeled by statutory exemption, and exports redacted copies and an exemption log only after a reviewer confirms. Here the agent prepares and a person commits, which is the right design. The risk is volume: when an agent proposes hundreds of redactions, confirmation can turn into a click. We argued that a human in the loop is only a control if the human can challenge the AI; grouping findings for review in passes, as Google describes, helps, but reviewer capacity has to be planned.

Inputs to decisions. Vendor bid evaluation checks bids against RFP requirements, the FAR and rubrics, and gives contracting officers “a documented basis for each award decision.” Grant audits flag unallowable costs against 2 CFR 200 with line-item citations. These outputs feed decisions that carry legal accountability. The officer decides; the agent’s analysis becomes part of the file and should be retained as such.

Decisions the agent should not make. Google does not claim the system makes final eligibility, enforcement or award decisions, and enterprises building on it should not either. That boundary belongs in the design, written down per workflow, rather than in a disclaimer.

What to settle before a pilot

Who owns each skill. A skill that encodes how a statute applies is a procedure. When the regulation changes, someone must update it, test it and record the version in effect for each case. Treat skills like controlled guidance documents, with an owner, an effective date and a review cycle.

Which version answered. If a citizen or an auditor challenges a decision months later, the agency needs to know which skill version, which sources and which model produced the analysis behind it.

Where authorization lives. Connectors that query systems in place, under existing permissions, keep the authority where it already is, in the system of record. That is the strongest part of the design, and the reason the architecture can work: the agent reasons, the system that owns the transaction authorizes.

Private sector teams in regulated industries should watch this preview closely. Replace “statute” with “SOP” or “credit policy” and the architecture is the same, and so are the questions.