A small business that works in Gmail, Drive and Calendar now gets one more offer from Google: an agent platform. Gemini Enterprise lets an employee describe an agent in plain language, connect it to Gmail, Drive or HubSpot, and have it run every morning. Hence a fair question: why pay for a custom AI agent if Google already does it?
The detail that settles most of it sits in Google's help pages: a no-code agent in Gemini Enterprise acts on behalf of the person who built it, with their access and their credentials. For a personal assistant, that is exactly right. For work that belongs to the company rather than to one employee, it is a limit.
The short answer, for the busy
Gemini Enterprise is Google's agent platform, sold separately from Google Workspace and from the Gemini app. It searches across your applications (Drive, Gmail, Microsoft 365, HubSpot, Jira and others), ships ready-made agents such as Deep Research, and lets an employee build a no-code agent that can run on a schedule. The Business edition, aimed at small businesses, is bought online with a 30-day trial. These agents work on behalf of one employee: they use that person's access, their scheduled runs stop unless the person's credentials are refreshed every 14 days, and a scheduled action that involves other people waits for that person's approval. The Business edition does not let you choose where your data is processed. For searching and drafting, the tool is good. Custom work becomes necessary when the process has to write into software missing from the connector list, run with nobody behind it, and keep a log that belongs to the company.
Three Geminis, three products
Google uses one name for three things, and its own help page takes care to separate them.
The Gemini app is the everyday assistant. For Google Workspace customers it comes at no additional cost and draws on Gmail, Drive and Chat.
Gemini in Workspace is the assistant built into Gmail, Docs and Sheets, helping people write, summarize and organize inside each app. The subscriptions themselves, Workspace included, are compared in our article on Copilot, ChatGPT and Claude.
Gemini Enterprise is a separate product, which Google describes as a standalone agent platform that connects to many internal and third-party systems to automate multi-step workflows. It comes in editions: Business for small and mid-sized businesses and small teams, with no complex IT setup; Standard and Plus for large organizations; Frontline for frontline staff. Standard, Plus and Frontline go through Google's sales team. Business is bought online, with a 30-day trial, and can also be added to a Workspace subscription.
The developer tool, Gemini Enterprise Agent Platform (the evolution of Vertex AI, billed on consumption), is purchased separately. And before paying for one more license, note that Workspace Studio, Google's tool for designing flows and agents inside Workspace, is included in its Business and Enterprise plans.
Agent Designer, the no-code tool
In Agent Designer, the employee writes in plain language what the agent should do, picks its sources, then refines it on a visual canvas, with sub-agents for multi-step tasks. They can then share it with the team and give it a schedule (hourly, daily, weekly) along with the prompt to run.
Connectors can write, too: in the Business edition, an agent can create a calendar event, send an email through Gmail, update a Salesforce record or open a Zendesk ticket. Google also supplies its own agents, such as Deep Research, which produces a sourced report, and Gemini Notebook, which turns a set of documents into a knowledge base.
The use it suits: every Monday, an agent reads the week's support tickets, compares them with the documentation in Drive and sends a summary to the manager. Nobody else is affected, and a person reads the result before anything is decided. Custom development would be hard to justify for that. If the real issue is the quality of answers drawn from your documents, the work starts in your files: see our article on assistants that answer from your documents.
An agent that works on behalf of one employee
Four rules from Google's help pages are worth reading before you hand a process to the platform.
Scheduled runs use the user's credentials, and those credentials expire every 14 days. If nobody clicks "Refresh the existing schedules," every scheduled agent that person owns stops running. Three weeks of vacation will do it.
On a schedule, the agent acts alone on personal tasks, such as creating an event in the user's own calendar or emailing the user. Any action involving other people pauses until the user reviews and approves it. A sensible safeguard, but the agent meant to follow up with your customers is then waiting for someone.
Sharing an agent shares its sources. Google spells this out as a caution: colleagues who receive the agent can ask it about the content of its connectors and files, even without direct access to them.
Finally, when an employee leaves the organization, their agents stay in place until an administrator deletes them.
These rules make sense for a personal assistant. They also draw the line: a process that has to run on the 1st of the month, whether or not the employee is around, needs a service identity with its own permissions and its own log. Microsoft draws a similar line, through licensing instead, as we described in the Copilot Studio article.
Connectors: reading is easy, writing depends on the list
The Business edition's connector list is long: Microsoft 365, Slack, Notion, Asana, Jira, Dropbox, Zendesk and many more. For searching and reading, a small business will almost always find what it needs.
Writing is narrower. Each connector has its own list of actions, and that list decides what the agent can do. On the accounting side, Google's list includes Zoho Books, added on October 15, 2026 in public preview, with actions to create invoices, expenses and payments. QuickBooks, Sage and Acomba are not on it.
When your software isn't listed, Google offers a custom MCP server connection, also in public preview. The requirements are explicit: the server must be exposed to the internet over HTTPS, the identity provider must be publicly reachable for the OAuth handshake, and Google points you to Cloud Run or to a partner to build it. In other words, the missing connector is custom development, just hosted at Google.
Writing reliably also takes more than a connector. An invoice created twice after a retry, a customer matched to the wrong record: those errors are prevented by checks specific to your business, tested before go-live. The same reasoning applies to Microsoft's flows, as we covered in our comparison of Power Automate and an AI agent.
Where your data goes, edition by edition
On training, Google commits for the Business, Standard and Plus editions: your prompts, responses and data are not used to train its models. In the Business edition, chats are kept for 60 days.
On where processing happens, the editions differ sharply. To the question "Can I restrict where my team's Gemini Business data is stored and processed?", the Business edition's security page answers no: it runs on Google's global infrastructure, with no setting or regional endpoint, and Google points to a higher edition for data residency. Google Cloud audit logging, customer-managed encryption keys and data residency appear among the compliance controls for Standard and Plus.
Those editions offer US and EU multi-regions, plus a Canadian region available through an allowlist. Several of the newest models aren't offered there, though: requests then go through the global endpoint, after the user confirms a warning.
Under Quebec's Law 25, the region doesn't settle everything: section 17 applies as soon as personal information is communicated "outside Québec", and Google doesn't say whether its Canadian region is in Quebec. If your agents are going to read customer files, the privacy impact assessment and the written agreement come before the first connection, whether the tool is Google's or ours. The requirements are laid out in our article on AI agent data outside Quebec.
Keep Gemini Enterprise, or move to custom
Gemini Enterprise remains the right tool as long as the agent searches and drafts for the person using it, from sources that are already connected, and someone reviews before acting. For that work, a license costs less than a project.
The math changes in four situations. The process has to write into software missing from the action list, such as your accounting or management system. It has to run on the 1st of the month even when the person who built it is on vacation. Someday you will have to explain to a customer why the agent made a given decision, with the log to back it up. Or its data has to be processed where you choose, which the Business edition doesn't allow, and cleared of personal information before it reaches the model.
Custom work has a price too: with us, a fixed monthly subscription, no development fee up front, set after a first call. What that subscription pays for is an agent the company owns: it has its own service account, its rules are versioned in a repository, and every step is recorded.
Many companies will keep both: Gemini Enterprise licenses for the team that searches and drafts, a custom agent for invoicing or follow-ups, where a mistake costs money. As for Gems, the lighter option, they change shape this fall: see our article on their replacement.
The 30-day test
Take the 30-day trial of the Business edition and build the simplest agent that would save you time. Write down what stops it: a missing action, an approval that waits, a schedule to refresh. If nothing stops it, keep Gemini Enterprise: it does the job.
If the list of blockers describes your most important process, bring it to a first conversation; sometimes our advice is to stay with Google.
→ First conversation, no commitment
Product names, editions, features and terms read on Google's pages and help centers on October 15, 2026. Google changes them often, and several features mentioned here are in preview. This article also explains a Law 25 obligation in plain language and is not legal advice.
Written byXavier Peich