Vendor comparison

Gemini vs Vertex AI: compliance comparison for Google buyers

Independent compliance comparison from Janus Compliance. Reviewed by Michael K. Onyekwere, CIPP/E. Last reviewed 2026-10-15. Not legal advice.

Profiles: Google Gemini (Workspace) · Google Vertex AI

Share this Google Gemini (Workspace) vs Google Vertex AI profile:Share on X Bluesky

TL;DR. Gemini in Workspace (formerly Gemini for Google Workspace) and Gemini Enterprise Agent Platform (formerly Vertex AI) both run on Google's Gemini model family, but their compliance pictures differ. Gemini in Workspace is an office-productivity feature that works inside an existing Google Workspace tenant and contract. Agent Platform is a Google Cloud developer platform with endpoint-based data residency, customer-managed encryption keys, a BAA covering the listed services, and a Model Garden of Google, partner and open models. Partner models such as Anthropic's Claude are called through Google's API, Google states that prompts and responses are not shared with third parties, and each partner or open model comes with its own separate terms. Gemini in Workspace fits office productivity; Agent Platform fits product builds and regulated workflows.

The two surfaces

The shared underlying technology is Google's Gemini model family. The product wrapper, contract, and compliance surface around it differs.

Both surfaces are covered in the Google Gemini profile, which also covers the Gemini Developer API and consumer Gemini. The comparison below sets the two enterprise surfaces side by side.

Gemini in Workspace

The Gemini app is a core service in most Google Workspace editions (Business Starter upwards, the Enterprise and Frontline editions, Essentials, Enterprise Essentials, Enterprise Essentials Plus and Nonprofits); Essentials Starter gets it only as an additional Google service, and the Gemini features inside Gmail, Docs and the other apps vary by edition
Operates inside the buyer's existing Workspace tenant
Customer data lives in the buyer's Workspace storage
Contract is the Google Cloud Terms of Service (which absorbed the Google Workspace Terms of Service in a version dated 1 June 2026, effective at renewal) or an offline Google Cloud Master Agreement, plus the Cloud Data Processing Addendum and the Google Workspace Service Specific Terms

Gemini Enterprise Agent Platform (formerly Vertex AI)

A Google Cloud product the buyer subscribes to separately
Operates inside the buyer's Google Cloud project, with project-scoped IAM and VPC controls
Customer data lives in Google Cloud Storage, BigQuery, or wherever the buyer pipes it
Contract is the Google Cloud Terms of Service (online) or a Google Cloud Master Agreement (offline), plus the Cloud Data Processing Addendum and the Google Cloud Service Specific Terms. Google states that "Only customers whose use of Google Cloud is governed by the Google Cloud Platform Terms of Service are subject to prompt logging for abuse monitoring", and those customers can request an exception

How the two compare

1. Tenant model and where customer data lives

| | Gemini in Workspace | Agent Platform | |---|---|---| | Data location | Buyer's Workspace tenant (mail, drive, chat, meet, docs) | Buyer's Google Cloud project; data at rest stays in the location the buyer selects | | Region selection | Workspace Data Regions (United States or Europe), covering Gemini prompts and responses but not Gemini Notebook. Google's Data Regions help page (updated 13 October 2026) gives storage at rest on most Business, Enterprise and Frontline editions (not Business Starter) and in-region processing on Enterprise Plus, Enterprise Essentials Plus and Frontline Plus or with the Enterprise Data Regions or Assured Controls add-ons. Since its version of 12 October 2026, the Workspace Service Specific Terms define the in-scope editions by reference to that page | ML processing stays in a region only when the request calls a regional or multi-region endpoint for a model listed for that location. The global endpoint gives no residency guarantee, Google's data residency terms exclude grounding features and RAG Engine, and the EU endpoint excludes the UK and Switzerland | | Customer-controlled encryption keys | Client-side encryption, with keys held in an external key service the customer chooses, on Frontline Plus, Enterprise Plus, Education Standard and Education Plus | Customer-managed encryption keys (CMEK) on most features; not supported for Preview models, grounding with Google Search or Google Maps, durable caching, Veo, the Gemini Live API native audio model, Gemini Embedding 2, or batch inference on several models | | VPC Service Controls | Not applicable: VPC Service Controls is a Google Cloud control. Workspace relies on its own controls, such as data loss prevention, information rights management, client-side encryption and context-aware access | Yes, a VPC Service Controls perimeter applies to Agent Platform (not to grounding with Google Search or Google Maps, or to Veo) |

Agent Platform provides the controls regulated buyers usually ask for (CMEK, VPC Service Controls and data residency), and its residency commitment depends on the endpoint each request calls. Gemini in Workspace relies on the controls the Workspace tenant already has, with Data Regions on most editions and client-side encryption and Assured Controls on higher editions.

2. Training default and the Model Garden question

| | Gemini in Workspace | Agent Platform | |---|---|---| | Use of customer content to train Google base models | No (Google Workspace Service Specific Terms, section 12.11) | No (Google Cloud Service Specific Terms, section 18) | | Fine-tuning | None documented for Workspace | Available, supervised fine-tuning, customer-scoped | | Third-party models available | N/A | Yes, Model Garden offers partner models such as Anthropic's Claude and Mistral as managed APIs, and open models such as Meta's Llama (in Preview) | | Third-party model use of customer content | N/A | Requests go to Google's endpoints, and Google states that prompts and responses are not shared with third parties, including when partner models are used | | Buyer responsibility for third-party model choice | N/A | Yes. Enabling a partner model needs a user with the Consumer Procurement Entitlement Manager role, and the model is a "Separate Offering" with its own terms |

A buyer who picks Claude or Mistral through Agent Platform calls the model through Google's API, and Google states that "Customer prompts and model responses are not shared with third-parties when using the Gemini Enterprise API, including partner models." Google's Cloud subprocessor list (last modified 20 August 2026) names no model provider. Part of the contract comes from the model publisher. Google's documentation says Anthropic models "are not a Google product" and that their availability is subject to the terms for "Separate Offerings" in the Service Specific Terms and to separate terms in the model card, and section 19 of those terms disclaims Google's liability and indemnity for Separate Offerings. Llama models appear in Google's list of open models, marked Preview.

3. HIPAA BAA position

| | Gemini in Workspace | Agent Platform | |---|---|---| | BAA available | Yes, once a Workspace super administrator accepts the BAA in the Admin console | Yes, the Google Cloud BAA covers Gemini Enterprise Agent Platform and Generative AI on Agent Platform | | Tier required | Any Workspace edition except the legacy free edition | Google Cloud BAA accepted by the customer (it covers Google Cloud's entire infrastructure and the listed services); PHI only in listed services, with unsupported services disabled | | Consumer / free tier | No BAA | No BAA | | Eligible services | Gemini in Google Workspace, the Gemini app (excluding Gemini in Google Chrome) and the Gemini Mac app; not Gemini Notebook | Gemini Enterprise Agent Platform and Generative AI on Agent Platform; no pre-GA (for example Preview) offerings with PHI; confirm partner-model coverage with Google in writing |

Google's BAA list names Generative AI on Agent Platform, not individual models. Google states that the certifications for generative AI on Agent Platform continue to apply when partner models are used as a managed API, and its HIPAA guidance says not to use pre-GA offerings, which include anything marked Preview, with PHI. A buyer that plans to send PHI to a partner model should confirm in writing with Google that the specific model is covered. Both the Workspace and the Cloud Service Specific Terms also bar use of Google's generative AI services for clinical purposes.

4. EU/UK transfer mechanism

| | Gemini in Workspace | Agent Platform | |---|---|---| | SCC version | EU-U.S. Data Privacy Framework first; 2021 SCCs as the fallback (see below) | EU-U.S. Data Privacy Framework first; 2021 SCCs as the fallback (see below) | | UK Addendum | UK Extension to the Data Privacy Framework from 16 September 2024; in the fallback, the ICO International Data Transfer Addendum is only in Google's controller-to-processor SCCs (see below) | UK Extension to the Data Privacy Framework from 16 September 2024; in the fallback, the ICO International Data Transfer Addendum is only in Google's controller-to-processor SCCs (see below) | | Region commitment | Workspace Data Regions (see section 1) | Set by the endpoint each request calls (see section 1); the EU endpoint excludes the UK and Switzerland | | Sovereign cloud option | Assured Controls and Assured Controls Plus, paid add-ons to Frontline Plus or Enterprise Plus | Assured Workloads for regulated and public-sector buyers | | What the transfer assessment records | The standard Workspace transfer position | For each model the application calls, the endpoint used and that model's listed processing locations, because partner and open models have different regional support from Google's own models |

The transfer position is the same for both surfaces. Google states that "As of September 1, 2023, Google has adopted an Alternative Transfer Solution for transfers of EU/EEA personal data to, and onwards from, the U.S." That solution is the EU-U.S. Data Privacy Framework, and Google adopted the UK Extension and the Swiss-U.S. framework on the same basis from 16 September 2024. When the participant registry was checked on 15 October 2026, Google LLC's EU-U.S., UK Extension and Swiss-U.S. certifications were all Active, to 13 September 2027. If Google stopped relying on the framework, the Cloud Data Processing Addendum would fall back to the 2021 SCCs: Module 3 (processor to processor, Google exporter) between the Google contracting entity (Google Cloud EMEA Limited for most EMEA customers) and Google LLC, or Module 4 (processor to controller) between Google and the customer where the Google entity is in an adequate country but the customer's billing address is not, and Module 2 or 3 between Google and the customer where the Google entity is not in an adequate country. The ICO's International Data Transfer Addendum for UK transfers is only in Google's controller-to-processor SCCs; the Module 3 (Google exporter) SCCs carry supplementary terms for Swiss transfers only.

5. AI Act deployer-side read

The Gemini models are general-purpose AI models under the EU AI Act (Article 3(63)), and Google is their provider. Google's obligations as a model provider under Chapter V, including Article 53, have applied since 2 August 2025, and models already on the market before that date have until 2 August 2027 (Article 111(3)). Gemini in Workspace and Agent Platform are services built on those models: Google provides Gemini in Workspace as an AI system, while Agent Platform gives buyers access to models to build their own systems. A buyer that only uses an AI system is a deployer; a buyer that develops an AI system and puts it into service under its own name is its provider (Article 3(3)). Every date in this section is a fixed date in the Act as amended by Regulation (EU) 2026/1744, which set the high-risk dates at 2 December 2027 (Annex III) and 2 August 2028 (Annex I). The obligations then follow the use case:

Gemini in Workspace for office productivity (drafting emails, summarising meetings, document search) is not high-risk unless it is used for an Annex III purpose. The deployer transparency duties in Article 50(4), for deep fakes and for AI-generated text published to inform the public on matters of public interest, apply from 2 August 2026.
A buyer that builds an application on Agent Platform and puts it into service under its own name is the provider of that AI system. A customer-facing chatbot carries the Article 50(1) duty to tell people they are interacting with an AI system, from 2 August 2026. Where the application generates synthetic text, audio, images or video, Article 50(2) also requires its output to be marked in a machine-readable format, with systems already on the market before 2 August 2026 given until 2 December 2026 (Article 111(4)). A recruitment screening tool is high-risk under Article 6(2) and Annex III, point 4, and the high-risk obligations apply from 2 December 2027. A clinical decision-support feature that is a medical device requiring third-party conformity assessment is high-risk under Article 6(1) and Annex I, with obligations from 2 August 2028; Google's Cloud Service Specific Terms also bar use of its generative AI services for clinical purposes.
Where the application calls a partner or open model, such as Anthropic's Claude, the model's publisher is the general-purpose AI model provider, and Article 53(1)(b) requires it to make information and documentation available to providers that integrate the model (Article 53(2) lifts this for free and open-source models without systemic risk). Google points buyers to the model card and the model publisher for details about partner models, so the buyer's assessment uses the documentation Google publishes plus the model publisher's material.

See the EU AI Act hub for the deployer assessment workflow.

6. Model providers and integrations

| | Gemini in Workspace | Agent Platform | |---|---|---| | Foundation model provider | Google (first-party) | Google, plus optional partner and open models from Model Garden (for example Anthropic, Mistral and Meta), each under its own separate terms | | Cloud subprocessor | Google Cloud (first-party) | Google Cloud (first-party) | | Third-party connector ecosystem | Workspace add-ons (controlled at tenant) | Agent integrations on Agent Platform; Vertex AI Extensions was deprecated on 26 May 2026 and will be shut down after 26 November 2026 | | Transfer assessment records | The standard Workspace transfer position (see section 4) | Each model the buyer's app actually calls, with its endpoint and listed processing locations |

Picking between them

Pick Gemini in Workspace when:

The use case is office productivity inside an existing Google Workspace tenant
The buyer wants tenant-bound defaults without additional Google Cloud subscription, project structure, or developer overhead
The compliance frame is minimal-risk or transparency-tier under the AI Act

Pick Agent Platform when:

The buyer is building a product, an internal tool, or a regulated-industry workflow on top of Gemini or another model
The deployment needs endpoint-based data residency, CMEK, VPC Service Controls, or Assured Workloads
The workload handles PHI in a custom application, and the buyer can keep that PHI inside the Google Cloud services listed in Google's BAA and disable unsupported services where PHI is handled
The buyer wants to call partner and open models (Claude, Mistral, Llama) through one Google Cloud account and API, accepting that each model's separate terms still apply

Run them in parallel when:

Gemini in Workspace covers the office-productivity layer for staff
Agent Platform covers the production AI surface the buyer ships to customers or uses in regulated workflows
The compliance owner manages two distinct deployer assessments because the use cases differ

What Model Garden adds to the buyer's records

Calling a partner model through Agent Platform keeps the data flow with Google: Google states that prompts and responses are not shared with third parties when partner models are used through its API, and its Cloud subprocessor list names no model provider. The buyer's records still need to track each model choice. For each model the application calls, the Transfer Impact Assessment should record the endpoint used and that model's listed processing locations, because Google says that for partner models "the regionalization of data processing may vary". The AI Act assessment should reference the model publisher's documentation, and where PHI is involved the buyer should confirm with Google in writing that the model is covered by the BAA.

Each partner or open model is a "Separate Offering" under the Google Cloud Service Specific Terms, subject to its own terms (for Claude, the terms on its model card), and Google disclaims liability and indemnity for Separate Offerings. The buyer therefore accepts the model publisher's terms when it enables a model, even though it calls Google's endpoint and pays Google.