This is the page the decision memo promised. Everything between the August 21 decision and the team actually having both tools working properly, in order, with who holds each piece. Claude is a build from nothing. Gemini is already running and needs its settings looked at, which is a different job, so the two are kept apart below.
Two halves. The first is the setup work, which is mostly Katrina, your IT partner and me. The second is for everyone, and it is short. If you are not doing the setup, skip to what changes for you and you will have what you need.
The reasoning behind every setting recommended here, ie. what each vendor actually commits to and why we are choosing what we are choosing, lives on the data posture page. This page is the sequence and the owners.
How we are starting. Everyone gets invited, and we watch what actually happens. Building an opt-out list before anyone has touched the tool would mean guessing, and the account reality in the staff diagnostic was that ten of thirteen already use ChatGPT, nine of them on personal accounts, so for most people this is a change of address rather than a new thing to learn. We turn seats off as people ask, on real usage. The one thing that follows from starting this way is the seat review at step 7, because a seat starts costing from the moment its invitation goes out.
Standard Claude Team pricing, ie. $20 per user per month billed annually. As Emily clarified in the room on August 21, Echidna is legally an LLC funded by charitable dollars, without the 501(c)(3) status the vendor discount programmes verify against, so the discounted rates the pre-read had quoted do not apply. At 18 seats that is roughly $4,300 a year, close to what Emily had budgeted. The decision memo carries the monthly and premium-seat options.
Each step opens for the detail. Owners are a proposal, not an assignment, so please move them around.
The one decision here is the Primary Owner. There can be only one, they hold every permission including data exports, and the account itself consumes a licence. It can be a service account rather than a person, ie. something like ai-admin@echidnagiving.org.
My take: make it a service account your IT partner and Katrina both hold, rather than an individual's login. When the Primary Owner is a person, the workspace quietly depends on that person still being there, and this is meant to outlast the engagement.
Four roles: User, Admin, Owner, Primary Owner. Users chat and use projects. Admins invite and remove people. Owners handle billing, integrations and features, and can change roles. Primary Owner does all of it plus data exports.
One quirk worth knowing. On Team plans, Admins cannot see usage analytics, only Owners and the Primary Owner can. So if Katrina is going to be the person reporting on how much the team is actually using this, she needs Owner rather than Admin.
Invite links are on by default for new Team organisations. Anyone holding the link can join. Domain discovery is the other one: colleagues whose email matches your domain can request to join on their own.
Neither is dangerous, and both sit awkwardly with paying per seat. Since we are inviting the whole team anyway, we should send those invitations ourselves and turn invite links off, so the seat count only ever moves when Katrina moves it.
Turn off Rate chats. Organization settings > Data and Privacy. By default Anthropic does not use Team or Enterprise inputs and outputs to train its models. The one route by which a conversation could be used is a person clicking the thumbs up or thumbs down button, which opts that conversation in and stores it for up to five years. Turning the setting off closes that route for everyone, so nobody can opt Echidna's data in by accident.
This matters more here than it would elsewhere. Two people in the staff diagnostic raised privacy posture unprompted, and this is the concrete answer to them.
Custom data retention is an Enterprise feature. On Team, chats are kept until someone deletes them, and we cannot set an automatic expiry. Enterprise allows a retention period as short as 30 days. Audit logs are Enterprise-only too, so on Team there is no record of who did what in the workspace.
Neither is a reason to change course. They are both reasons to decide deliberately whether the retention control is worth the Enterprise step, and to be clear with the team about what the plan we are on does and does not give them. My read is that this belongs with the working group rather than in the setup, so I have kept it as a decision rather than a task.
An Owner enables or disables web search for the whole workspace under Organization settings > Capabilities. Once it is on, people can still toggle it per chat.
The trade-off is real in both directions. Web access on makes Claude far more useful for scanning the field and checking a grantee's public reporting. Web access off is what keeps an answer anchored to the document in front of it, which is exactly the discipline that made the accuracy risk manageable in the bake-off. My take: leave it on, and teach the habit of turning it off for document work, because a setting people cannot reach is a setting they route around.
Connectors are what let Claude reach into Google Drive, email and the rest. An Owner has to enable them for the organisation before anyone can use one, and Owners can set each action to Always allow, Needs approval, or Blocked.
We should start with them off and open them one at a time, after the drive sort below has told us what is safe to reach. Connecting Drive before we know which folders are AI-excluded would undo the sort before we have finished it.
Team plans allow spending caps at both the organisation and the individual level. Worth setting at the start, mostly so that nobody has to think about it again.
The catch, and it matters given how we are starting. Removing a member frees their seat for reassignment, but it does not reduce the plan's total seat count. Turning someone off is two actions rather than one, ie. remove the member, then reduce the seat count. Pending invitations also hold a seat from the moment they go out, whether or not anyone accepts. So invite-everyone-then-turn-off only stays honest if someone is watching: a monthly look at who has actually signed in, with the seat count reduced to match. Katrina is the natural owner, and it wants a recurring diary entry rather than good intentions.
Gemini has been in the building since before this project started, which means its settings are wherever the defaults left them and nobody has been back since. This is a review rather than a deployment, and it comes with a wrinkle Claude does not have: people already have real conversations in there, so a retention change acts on work that already exists.
Ten minutes in the Admin console, under Generative AI, recording the current state of each setting below rather than assuming it. My expectation is that most sit at their defaults, which would mean conversations are being kept indefinitely and the Gemini app's connections to non-Workspace Google services are on. I would rather we knew than guessed.
The reason this goes first: turning on auto-deletion acts on conversations that already exist. If anything in there matters to somebody, we want to have seen it coming.
Generative AI > Gemini for Workspace > Conversation history & deletion. The choices are letting each user decide, 90 days, 540 days, 1080 days, or never deleting. We recommend 90 days, the shortest available.
This is the single most valuable setting on either tool, because it is the one thing Claude Team cannot do at all. Retention was the concern the team raised, and Gemini is where we can actually answer it. Leaving it on user choice reads as flexible and means in practice that almost nobody changes it.
Keep Allow users to delete conversations on, on the same screen. People should be able to remove something they pasted by mistake without filing a ticket.
Generative AI > Gemini app > Apps. The Gemini app can connect to Google services outside Workspace, ie. Maps, YouTube, Flights and similar, and those run under the standard Google Terms of Service rather than your Workspace agreement.
Little upside for this team, and turning them off keeps one clean line: everything Gemini touches sits inside the agreement that already covers your Gmail and Drive.
On the same screen, Gemini's access to Gmail, Drive, Calendar, Docs, Keep and Tasks is on by default. It is almost certainly right to leave on, and it should be a decision rather than an inheritance.
Gemini reaches exactly what the signed-in person can already open, so the shared-drive sort is what actually decides its reach. That is the same sort the Claude connectors wait on, which is why it is the one piece of work gating both tools.
Which Workspace edition Echidna is on decides which Gemini features are included, what the retention options are, and whether Google Vault comes with it. I have not confirmed the edition, and it is a question for your IT partner rather than a guess worth publishing.
If Vault is included, Echidna can set retention rules and place legal holds on Gemini conversations itself, and those rules take precedence over both admin and user settings. That is a capability we have no equivalent for on the Claude side, and it changes what we are able to promise in the policy session.
The workspace is the smallest piece. These are the five that carry more weight, and they run in parallel.
We are not opening a negotiation with either vendor, so this is education rather than procurement. The team should understand what Anthropic and Google publish about retention, training on your data, and how each responds to a legal hold or a government request, and understand it well enough to make their own calls. The data posture page is the material, and it belongs in the training rather than sitting on the hub waiting to be found.
Being straight about the limits of that: working from the published record means some things stay unanswered, ie. Claude Team's exact retention commitment and whether Echidna would be told if a litigation hold ever reached its data. We are choosing to accept that rather than chase it, which is a reasonable call at this size, and it is worth the team knowing that is the choice we made.
Drives mapped into AI-allowed and AI-excluded, leadership drive first. HR, personnel, accounting and founder files stay outside what any AI tool can reach. This is the piece that has to be finished before connectors get switched on, and it is the one that takes real hours rather than clicks.
Ellen, Arjun, Katrina, Dennis and Ali, with Emily as the senior-team sponsor and Nat joining for grants questions as needed. We will propose the shape and cadence first so everyone can size the commitment before saying yes. The role and expectations draft is on the hub for senior staff to look over.
Personal accounts for work use wind down, with an offboarding path for saved work and chat history. Nine of the thirteen have been running work through personal accounts, so this needs to feel like help rather than enforcement. We will give people a window and a way to bring their saved work across.
What was decided, that it came from the team's own survey input, and when everyone will feel it. Draft coming for your edit.
The green, yellow and red data rules, account rules, and the go/no-go framework for new tools, built on the policy pre-read. Setup should be far enough along by then that the rules land on something real.
Gemini stays where it is and keeps doing what it does inside Workspace, with two changes behind the scenes: Gemini conversations will start deleting themselves automatically after 90 days, and the Gemini app will stop connecting to Google services outside Workspace, ie. Maps, YouTube and the like. Claude joins it for the heavier reading and writing work, ie. going through a long grantee report, drafting a synthesis, working through a set of documents together. You will get an invitation by email.
We are inviting the whole team rather than asking people to sign up in advance. If you would rather not have it, say so and we will turn your seat off, no explanation needed and no permanent decision made. You can pick it up again later.
If you have been running work through a personal ChatGPT or Claude login, that winds down, and we will give you a window and a hand to bring saved work across. Nobody is in trouble for having done this. Nine of thirteen of you told us you were, which is precisely why the workspace exists.
Your chats live in an account Echidna administers rather than a personal one. The Primary Owner can request a data export that includes conversations and uploaded files. In practice nobody is reading your chats, and the reason to know this is simple: treat it like your work email rather than your personal notes app.
The flip side is the part worth having. Anthropic does not train its models on this workspace's conversations, and we are turning off the one setting that could opt a conversation in.
Until the rules land on September 1, hold to the red lines you already drew in the diagnostic: no grantee personal data, nothing from HR or personnel files, nothing from accounting. If you are unsure, that is a good question for a Guide rather than a reason to avoid the tool.
Ellen, Arjun, Katrina, Dennis and Ali are stepping into the Guide role, with Emily sponsoring and Nat picking up grants questions. They are your first stop, ahead of me, and that is deliberate: this needs to work after October.
Shout if any of this looks off, and I will fix the page.
Sources: Anthropic help centre and privacy centre, retrieved 28 August 2026. What is the Team plan · Roles and permissions · Manage members on Team and Enterprise plans · Manage user feedback settings on Team and Enterprise plans · Is my data used for model training · Configure custom data retention controls for Enterprise plans · Access audit logs · Enable and use web search · Use connectors to extend Claude's capabilities. Prepared by SA Advisory as the follow-on to the August 21 tool decision.