What each tool retains, where it lives, what a court or government can reach, and what training touches. Everything here comes from the public record, ie. the vendors' own policies, help documentation and transparency reports, plus one court case that tested the promises in practice. Each tool gets a section, and each section ends with the settings we recommend and why.
UPDATED AUG 28
Posture changes completely between a vendor's consumer product and its business product, so the product name matters more than the brand. These two are what Echidna runs.
The consumer versions of both behave differently from what is described here, and that gap is the whole argument for moving people onto Echidna accounts. ChatGPT was assessed alongside Claude on August 18 and not chosen; that comparison is kept at the bottom of this page for anyone who wants it.
The honest summary in one line: the two tools protect Echidna in different places, and neither of them protects it from what we choose to put in. That is what the September 1 policy is for.
Anthropic's commercial terms bar training on Claude Team and Claude Enterprise content. The single route in is a person clicking the thumbs up or thumbs down button on an answer, which opts that one conversation in and stores it for up to five years. There is an org-wide setting that removes the button entirely, and we are turning it off, so nobody can opt Echidna's data in by accident. Two people raised privacy posture unprompted in the diagnostic, and this is the concrete answer to them.
Custom data retention is a Claude Enterprise feature, and Claude Team has none. On our seats, chats are kept until somebody deletes them, with no automatic expiry available at any price on this plan. Enterprise allows a window as short as 30 days and adds audit logs, which Team also lacks, so there is no record of who did what inside the workspace.
That is worth stating flatly rather than softening: on the retention question the team cared about most, the plan we are on has no answer, and the mitigation is behavioural rather than technical. It is a fair trade at eighteen people, and it is a reason to keep the sensitive material out rather than to manage it once it is in.
Anthropic's chat product is US-hosted, with regional options existing only through cloud APIs rather than the product we deploy. So Claude Team means data processed and stored in the United States, reachable by US legal process. Anthropic publishes a government-requests report and maintains a public transparency hub, and has had no equivalent of the court-ordered preservation event described at the bottom of this page.
The full sequence, with owners, is on the setup checklist.
Prompts and responses are Customer Data under Google's Cloud Data Processing Addendum, the same agreement that already covers Echidna's Gmail and Drive. Google's commitment is that this content is not reviewed by humans and not used to train generative AI models outside your domain without permission. Gemini carries ISO 42001, BSI C5 and FedRAMP High, and supports HIPAA.
An administrator sets auto-deletion in the Admin console, choosing between letting users decide, 90 days, 540 days, 1080 days, or never. There is a separate switch for whether users can delete their own conversations. For organisations with Google Vault, retention rules and legal holds can be applied to Gemini app conversations directly, and Vault rules take precedence over both admin and user settings.
Which means Echidna holds its own legal-hold capability on the Gemini side, and holds nothing equivalent on the Claude side. That is the single sharpest difference between the two tools on this page.
Gemini respects existing Workspace access controls, ie. it reaches exactly what the signed-in person can already open, and nothing more. That is the reassuring half. The other half is that it makes the shared-drive sort the control that matters, because anything left readable in a drive is readable by Gemini on behalf of anyone who can open it, and Gemini is the one tool of the two that can also write back into Drive. The leadership drive coming out first, with HR, personnel, accounting and founder files, is what makes this safe.
One boundary worth knowing: the Gemini app can also connect to non-Workspace Google services such as Maps, YouTube and Flights, and those run under the standard Google Terms of Service rather than your Workspace agreement.
On a personal Google account, Gemini Apps Activity defaults to eighteen months, a subset of chats is reviewed by human reviewers, and chats are used to improve Google's services unless Keep Activity is turned off. Even with it off, chats are held for 72 hours. None of that applies to Gemini signed in with an Echidna account. Same brand, same interface, materially different deal, which is exactly the trap the account migration closes.
The strongest legal protections in this space live in Enterprise contracts. Claude Enterprise carries seat minimums, and ChatGPT Enterprise requires a 150-seat minimum on an annual contract. At eighteen people the natural tier is Claude Team at standard commercial pricing, and Echidna is legally an LLC funded by charitable dollars, so vendor nonprofit programmes do not apply either. We are not opening a negotiation with either vendor at this size, so the contract is what it is.
There is one piece of evidence about what that means in practice, and it is worth reading carefully. In the New York Times copyright litigation, a US court ordered OpenAI to preserve user conversations, deleted ones included, regardless of individual settings. Enterprise, Edu and zero-data-retention API customers were excluded from the hold. Lower tiers were not. The order was later narrowed, and in January 2026 a court ordered production of roughly 20 million de-identified consumer chat logs.
We want to be careful with this, because it is one lawsuit rather than a settled pattern, and the durable lesson points at no particular vendor: protection followed the contract tier, not the settings. Deletion promises sit downstream of courts, and what decided who was protected was the paper their customers had signed.
Echidna sits below the strongest carve-outs on both tools, which puts the protection back where we can actually reach it: what enters the tools at all. That means the red, yellow and green data sort coming out of the September 1 session, de-identification as the working habit, and the shared-drive sort that decides what Gemini can see. Those three do more for grantees than either vendor's terms do.
This was the evidence for the August 21 session, written on August 18 when the choice was still open. It is kept because the reasoning still holds and because ten of thirteen staff use ChatGPT today, nine of them on personal accounts. Nothing in it changes what we do now.
Both vendors excluded business-tier customer content from model training by default. OpenAI's policy states that ChatGPT Business and ChatGPT Enterprise data is not used for training unless the workspace explicitly opts in. Anthropic's commercial terms bar it outright. The practical difference was small but real: on ChatGPT an administrator could turn training on, and on Claude's commercial plans that switch does not exist.
In normal operation the two were comparable: deleted business-tier content purges within about 30 days on both, and ChatGPT Business gives workspace admins retention controls that Claude Team does not have. The legal-pressure evidence is in the section above.
OpenAI offers data residency in ten regions, India among them, but only on Enterprise and Edu plans. Anthropic's chat product is US-hosted. So at the tiers Echidna could buy, both meant data processed and stored in the United States. Residency did not separate them at this price point.
Both publish transparency reporting on government demands for user data: OpenAI issues semiannual reports and a law-enforcement policy, effective January 2026, requiring valid legal process; Anthropic publishes a government-requests report and maintains a transparency hub. Neither breaks out numbers for business tiers.
On training, transparency, certifications and residency the two were effectively level at the tier we could buy, so the data questions did not pick the winner. That was itself the useful finding: it freed the decision to rest on accuracy demonstrated on our own material, on what the team already uses, and on the values read the room made on the day.
Gemini gives us admin retention control, Vault holds if the edition carries them, and permissions it inherits from Drive. Claude Team gives us none of those, and its protection is that it only ever sees what a person deliberately puts in front of it. That is a reasonable pair, as long as both halves are actually configured and the sorting rules are real.
The settings above are the short version of the work. The drive sort and the September 1 data rules are the long version, and they matter more.
What the public record cannot answer, and we are choosing to accept rather than chase: Claude Team's exact retention commitment beyond what the settings allow, and either vendor's customer-notification commitment for legal holds and government requests. Open and answerable in-house: which Google Workspace edition Echidna is on, which decides the Gemini features included, the retention options available, and whether Vault is there.
Compiled by SA Advisory from public sources. The Claude and ChatGPT comparison, August 18, 2026, from vendor privacy policies and help documentation, OpenAI's government-request reports (2025 H1 and H2) and law-enforcement policy (effective January 2026), Anthropic's government-requests report and transparency hub, and published court orders and reporting in the New York Times litigation. The Gemini section and the Claude Team retention, audit-log and settings detail, August 28, 2026, from the Generative AI in Google Workspace privacy hub, Google's Workspace AI privacy pages, the Gemini in Workspace conversation history admin documentation, the Gemini app admin controls, Google Vault's Gemini retention documentation, the Gemini Apps privacy hub, and Anthropic's help centre on retention, audit logs, roles, connectors and feedback settings. Independent summaries were cross-checked where vendor pages were ambiguous. Please shout if anything looks off.