Gemini Enterprise Placement Decision: A 5-Test Framework

Most teams evaluating Gemini Enterprise are asking the wrong question. "Should we buy it?" is a procurement question. The one that actually determines whether the line item survives its second renewal is different: what job does this hold in our operating model, and who owns that job?

That matters more now than it did a year ago, because the thing being sold changed shape. On April 22, 2026, Google launched the Gemini Enterprise Agent Platform as the evolution of Vertex AI, stating that all Vertex AI services and roadmap evolutions would be delivered exclusively through the Agent Platform rather than as a standalone service. Google now positions Gemini Enterprise as two surfaces: the employee-facing Gemini Enterprise app, and the Agent Platform underneath it.

Two surfaces means two placement decisions. Confuse them and you end up with unused seats, a second agent runtime nobody sanctioned, and governance living in whichever console the last team happened to open.

Here's a five-test framework you can take into a steering meeting.

The Jobs Bundled Under One Name

Before any test, separate the jobs. One name currently covers at least four:

• Workforce assistant, the seat-priced app your team members open to ask questions and draft work.
• Grounded enterprise search, permission-aware retrieval (the assistant only returns what the asker already has rights to see) across connected systems.
• Agent build-and-run surface, where agents get built, deployed, and run.
• Governed distribution plane for models, the layer through which every model your organization uses gets identity, policy, and billing.

These are separable. Model pluralism sits at the platform layer: Agent Platform provides access to more than 200 models through Model Garden, including Gemini 3.1 Pro, Gemma 4, and third-party models such as Anthropic's Claude Opus, Sonnet, and Haiku. So does agent governance. Google's layer here consists of Agent Identity, Agent Registry, and Agent Gateway, meant to give every agent a trackable identity operating within enterprise guardrails. Seats, connectors, and quota economics land at the app layer.

You can reasonably buy one without buying much of the other. Decide that on purpose.

Test 1: What It Replaces

Which AI surfaces are already open on your team members' screens this morning? List them: the copilot bundled with your productivity suite, the in-house retrieval application your data team built, the orchestration framework two squads standardized on. For each, answer plainly. Does Gemini Enterprise retire this, or not?

Be honest when the answer is "not." Consolidation for its own sake is often the more expensive path, especially when your productivity suite and your cloud come from different vendors. Ripping out a working incumbent to standardize the assistant layer mostly buys tidiness.

The displacement test has a hard edge, though, and it's the edition matrix. Google's edition comparison (page last updated 2026-08-27) shows Business excludes:

• using your own full-code agents built outside Gemini Enterprise
• the full data connector ecosystem
• the Agent Marketplace
• Gemini Code Assist Standard
• priority access to the latest Gemini models
• AI developer tools
• the row labeled "enterprise-grade security and compliance," the row your security reviewer will ask about by name

Building and publishing custom no-code agents is marked "(preview)" as of that same update.

If your replacement plan assumes full-code agents or your existing connector set, Business edition won't run that plan. Test it before you sign. Not in month four.

Test 2: What It Sits Beside

Coexistence is a legitimate answer, but only if you draw the map. For every surface you're keeping, write down which one is authoritative for which task, and what the team member is supposed to do when both can answer.

The same discipline applies to capability boundaries. Google's product page footnotes that data sovereignty and compliance capabilities, VPC Service Controls, Customer-Managed Encryption Keys, Access Transparency, data residency, and support for workloads like HIPAA and FedRAMP High, are available in the Standard and Plus editions of the Gemini Enterprise app. Put a regulated workload on the wrong edition and you've bought a control gap.

Test 3: What It Must Not Own

What happens the first time a custom agent quietly inherits a director's file access? Four things should never be ceded to any single AI surface, Google's or anyone else's:

• Identity. Your directory stays the source of truth for who is who.
• Data-access policy. Permission-aware retrieval is a feature. Your governance program still has to exist on top of it. Security analysis published in February 2026 makes the point sharply: because the assistant surfaces whatever a user can already access, existing oversharing and over-broad permissions get amplified at machine speed. Persistent custom agents operating with assigned permissions increase governance complexity when agent ownership and data sources aren't visible.
• System-of-record writes. Reads can be broad. Writes belong behind connectors your platform team owns.
• Model choice. The moment one surface owns which model runs, your exit cost stops being a migration and becomes a rebuild.

That last one has a concrete answer on Google Cloud. In a July 14, 2026 post, Google engineers state that invoking Claude on Google Cloud is operationally identical to invoking any other Google Cloud service. Both use the same IAM policies, the same VPC Service Controls, and the same observability stack via Cloud Logging and Cloud Monitoring, with requests inheriting the project's IAM and VPC configuration. Agent Platform exposes global, regional, and multi-region endpoints for Claude, and the regional option keeps prompts, completions, and intermediate state inside a geographic boundary.

One caveat worth stating plainly rather than glossing over: model pluralism is documented at the Model Garden and Agent Platform layer. Whether every managed surface, including the Gemini Enterprise app assistant itself, offers the same choice isn't something the available documentation establishes. Test it per surface.

Test 4: Who Owns It

Who gets paged when an agent misbehaves at 2am, the team that manages seats, or the team that manages the platform? Google now ships app-level administration separately from platform-level governance. Business edition release notes show admin-side agent management (view, share, enable, disable, suspend, delete custom agents) arriving on August 17, 2026. Agent Identity, Registry, and Gateway sit at the platform layer. Two control planes means two owners, unless someone deliberately unifies them.

Pick that someone before rollout. IT usually owns seats and connectors well. An AI center of excellence usually owns agent standards well. Business units own use cases. What fails is leaving the seam unassigned, which is roughly where the industry sits. Writer's April 2026 enterprise AI adoption survey of 1,200 executives and 1,200 team members reports 36% of companies have no formal plan for supervising AI agents. It also finds 35% admit they couldn't immediately pull the plug on a rogue agent. And 79% say AI applications are being created in silos.

Test 5: How It's Paid For

The bill is a seat line plus consumption. Both behave in ways spreadsheets tend to miss.

Feature quotas pool per edition within a project and location and adjust with seats. Pooling means the daily allowance is shared across all licenses in that project, not handed out per person. Google's documented example: 100 Standard licenses share 16,000 assistant queries per day, 160 per license. 50 Plus licenses share a separate 10,000 per day, with storage and indexing pooling across editions in the project. A handful of heavy users can draw down a shared pool fast. Which project a seat lives in shapes both service levels and cost. The Pay-as-you-go subscription removes feature quota limits entirely; you pay for what you use.

On seat prices, be careful in your business case. Third-party partner and analyst pages report roughly $21 per seat per month for Business, around $30 for Standard on annual terms (about $35 month-to-month), and $50, 60 for Plus. Consumption beyond quota is additional. Google doesn't publish a Plus list price. Treat those figures as reported, and confirm current rates with your Google contract owner before committing.

One funding lever is worth checking: software purchased through Google Cloud Marketplace counts toward committed Google Cloud spend. Google's 2025 channel update stated qualifying software purchases through authorized channel partners count 100% toward committed spend, capped at 25% of the total commitment. That channel guidance dates to Google's 2025 update and may have moved. Verify before you build a case on it.

And the packaging itself is a moving target. Renamed, restructured in April 2026, edition rows still shifting month to month, Gemini 3.7 Flash added to the Business model selector on August 13, 2026, and Gemini 3.8 Flash shipping September 2, three Flash releases in seven weeks. A 12-month commitment gets priced against something that won't look identical at renewal. Google Cloud is also vertical-packaging the product: Gemini Enterprise for Legal was announced August 25, 2026 as a purpose-built industry solution embedding governance and permissions inheritance.

Momentum is real, for what it's worth. Alphabet's Q1 2026 results reported 40% quarter-over-quarter growth in Gemini Enterprise paid monthly active users, with Bosch, Citi Wealth, Merck, and Mars named as customers. That's a relative growth figure and the freshest first-party number available. Read it as directional interest, and build the case on your own usage.

A 90-Day Placement Plan

Weeks 1, 4, prove placement. Run tests 1 and 2 against a real inventory of AI surfaces. Pick one workflow and one agent. Confirm the edition you're scoping actually includes the capabilities that workflow needs.

Weeks 5, 8, govern before you scale. Register every agent with an owner and named data sources. Set the identity, policy, and write boundaries from test 3 in writing. Instrument consumption per model and per project. Long-running agents are a new class of runaway spend, and a monthly invoice is a slow way to find one. This is the kind of visibility WALT Labs' Cloud Companion is built for: consumption monitoring per model and per project.

Weeks 9, 12, decide seats and defer the rest. Allocate seats by project against the quota pool; headcount is the wrong unit here. Defer anything still marked preview until it ships.

Getting It Implemented

A placement framework only pays off when someone builds the thing it describes. WALT Labs stands up Claude on the Gemini Enterprise Agent Platform inside your existing Google Cloud estate, using the IAM, VPC Service Controls, and billing you already own, so model choice stays a configuration decision rather than a contract you have to escape. Heading into a renewal with this question unresolved? That's the conversation to have.

Which layer are you actually buying?