Grounding Gemini Enterprise: Which Systems to Connect First

Gemini Enterprise connects to the places your work already lives: Google Workspace, Microsoft 365 and SharePoint, Salesforce, SAP, ServiceNow, Box and Workday among them. Because that list is long and the connectors are not hard to switch on, the default plan is usually to switch on all of them and see what happens.

What happens is a search experience that is technically grounded and practically useless. The assistant starts confidently citing a pricing deck from 2023, a policy that was superseded in March, and four copies of the same document at different stages of edit. Users try it twice, conclude it is unreliable, and go back to asking a colleague.

The failure is not retrieval quality. It is that you indexed a corpus nobody had curated, and the model has no way of knowing which of your four documents is the one people actually follow.

The instinct is to start with the biggest repository, because that feels like the most value. The better instinct is to start with the one that has the highest ratio of answers to noise.

We score candidate sources on three things before connecting any of them.

Answer density. Of the questions people actually ask, how many does this source settle definitively? A well maintained policy space in Confluence scores very high. A shared drive that has accumulated for eight years scores near zero regardless of size.

Permission clarity. Does this system already model who can see what, accurately, today? If the permissions in the source are wrong, permission-aware retrieval will faithfully reproduce that wrongness at conversational speed, which is considerably worse than the current situation.

Change rate. How often does the truth here move? Fast-moving sources need a refresh story before they need an index, or the assistant becomes an efficient way of distributing stale answers.

Gemini Enterprise scopes what an agent can retrieve to what the person asking is already entitled to see. This is the right design, and it is the feature that makes enterprise deployment defensible at all.

It is also frequently misread as a guarantee that the platform provides. It is not. It is a faithful reflection of the access model in the system you connected. If a folder was shared with the whole company in 2021 because it was easier than managing a group, the assistant will treat that as intentional, because as far as any system can tell, it was.

Before connecting a source we run one deliberately unglamorous exercise: pick the three most sensitive documents in it, and check who can currently open them. In roughly every engagement, at least one answer surprises the owner. Finding that during connector planning is a good day. Finding it when an agent surfaces a compensation model to the wrong person is not.

A useful first wave is narrow enough to fix and broad enough to be worth using. In practice that has meant one curated knowledge source, one system of record, and nothing else.

The curated source answers the questions people ask constantly and gives the assistant an authoritative voice on policy, process and how things are done here. The system of record gives it live facts: the account, the ticket, the order, the case. Together those cover a surprising share of what people currently interrupt each other to ask.

What you deliberately leave out in wave one is the large uncurated document store. Not permanently. But it needs an owner, a retention decision and a pass at duplicates before it earns a place in the index, and that work is easier to fund once people can already feel the assistant working.

The part that gets underestimated is that this is not a project with an end date. Sources drift, teams reorganise, someone stands up a new space, a system gets replaced.

We treat the connected corpus the way you would treat any production dependency: named owners per source, a review on a fixed cadence, and a route for users to flag an answer that was grounded in something stale. The last one matters most, because your users will find the decayed corners of your corpus far faster than any audit will.