The single most important fact for any law, architecture or finance firm: the tier you pay for matters more than the brand you pick. On consumer accounts, Gemini, Claude and ChatGPT can all train on your conversations and keep them for months or years. Their business, enterprise and API tiers don’t train on your data by default, and they come with contracts, admin controls and retention settings. Pick the right tier and configure it, and all three can be defended for confidential work. Get the tier wrong and no setting will save you.
TL;DR
- Consumer accounts are the risk, whichever vendor you use. Personal Gemini keeps activity for 18 months by default, and human-reviewed chats for up to three years. Claude Free/Pro/Max keeps chats for up to five years if training is on. ChatGPT Free/Plus/Pro trains on chats unless you switch it off.
- Business tiers fix training, not every retention problem. All three exclude business data from training by default, but Claude Enterprise and ChatGPT Enterprise keep chats indefinitely until an admin sets a retention period.
- Law firms: courts have forced OpenAI to preserve deleted chats and ruled a defendant’s consumer Claude chats unprivileged. ABA Opinion 512 requires informed consent before client data goes into “self-learning” tools.
- Finance firms: the problem is reversed. You may have to keep AI chats as business records, which in practice requires an enterprise tier with a compliance API feeding your archive.
Last verified November 3, 2026. These policies change often. Confirm against each vendor’s current documentation. This is not legal or compliance advice. Disclosure: Streamline Agency is an Anthropic Claude Select Service Partner and certified n8n partner; we’ve applied the same standard to every vendor and note where Claude falls short.
Architecture firms: drawings, planning and design IP
Architects usually ask whether drawings and planning submissions will be kept, used for training, or surfaced to someone else.
- Training is the risk you can control. Since September 2025, Google may use a sample of uploads to personal Gemini accounts (with Keep Activity on) to improve its services, so a floor plan or permit set uploaded there is in scope. Consumer Claude and ChatGPT can also train on uploads unless training is switched off. On business tiers, none of the three trains on your content by default.
- Retention is the risk you have to configure. Business-tier defaults are often “keep forever.” Unreleased designs sitting in a chat archive for years are a breach and discovery exposure even if no model trains on them.
- Exposure mostly comes from people. The realistic leaks are staff using personal accounts, shared projects left open, and connectors that can reach more of SharePoint or Drive than intended. Clean up folder permissions before enabling connectors.
Law firms: confidentiality, privilege and discovery
Consent (ABA Formal Opinion 512). Because “self-learning” tools could disclose client information, informed client consent is required before inputting it, and boilerplate engagement-letter language isn’t enough. A business tier that contractually excludes training is a strong argument the tool isn’t self-learning, but check your state bar; Alaska’s Opinion 2025-1, for example, requires anonymized inputs absent consent.
Privilege (United States v. Heppner). In February 2026 Judge Rakoff (SDNY) held that documents a defendant generated with consumer Claude were protected by neither attorney-client privilege nor work product: the platform was a third party, Claude isn’t a lawyer, and the defendant didn’t act at counsel’s direction. The ruling concerns a consumer account, not enterprise use by lawyers, but it shows how courts will reason. Warn clients against running legal questions through consumer chatbots.
Discovery (NYT v. OpenAI). A 2025 court order made OpenAI keep deleted consumer and standard API chats until September 26, 2025 (Enterprise, Edu and zero-retention API customers were excluded), and in January 2026 OpenAI was ordered to produce 20 million de-identified logs. Your vendor’s retention can be overridden in someone else’s case, and your own AI logs are ESI your legal-hold process must cover.
Supervision. Model Rules 5.1 and 5.3 make partners responsible for staff use. Meet that with a written policy and admin controls.
Finance firms: you may have to keep the chats
Recordkeeping. FINRA’s 2026 Oversight Report applies existing supervision and recordkeeping rules to generative AI and suggests storing prompt and output logs. Whether AI chats are records under SEC Rules 17a-3 and 17a-4 is still debated, and FINRA has asked members about it, so the defensible position is to capture AI use tied to clients, trades and recommendations.
Vendor retention isn’t your books and records. Rule 17a-4 requires broker-dealers to keep business communications for at least three years in a tamper-proof (WORM or audit-trail) format. A vendor’s 30- or 90-day retention, or an “indefinite” default, doesn’t meet that. Export conversations to your archiving vendor and keep the AI vendor’s retention short.
Consumer AI is off-channel. The SEC’s off-channel sweep produced about $2.3 billion in penalties since 2021, no books-and-records rule has been repealed, and FINRA still lists off-channel use as an exam priority. An advisor drafting a client email in personal ChatGPT is today’s version of texting on WhatsApp.
Regulation S-P and NYDFS. Amended Reg S-P (larger firms from December 3, 2025, smaller firms from June 3, 2026) requires oversight of every service provider handling customer information, including AI vendors, and contracts requiring breach notice within 72 hours. NYDFS-regulated firms should also follow its 2024 AI letter and 2025 third-party letter, which says Part 500 compliance can’t be delegated to a vendor.
Platform by platform
Google Gemini. Personal accounts keep activity for 18 months by default (3 or 36 months, or never, are options), human-reviewed chats are kept up to three years even after deletion, and chats are held 72 hours even with activity off. In Workspace, content isn’t used for training or human review outside your domain without permission; admins set history at 3, 18 or 36 months. Google Vault now supports retention rules, holds and exports for Gemini conversations (not the Gemini API or Vertex AI), and Workspace integrates with Smarsh, Global Relay and Mimecast. Google has the broadest certification list (SOC 1/2/3, ISO 27001/27701/42001, FedRAMP High), though Gemini Notebook is carved out of most of it. The free Gemini API can use prompts for improvement; the paid tier doesn’t.
Anthropic Claude. Consumer users chose in 2025 whether chats train Claude; with training on, data is kept up to five years, otherwise deleted chats are purged within 30 days. Team, Enterprise and the API don’t train on your data. API inputs and outputs are deleted within 30 days; Enterprise owners can set retention (minimum 30 days), otherwise it’s indefinite. Watch two exceptions: feedback ratings are kept five years, and since June 2026 prompts to Anthropic’s most capable “covered” models are kept 30 days for safety monitoring, even under zero data retention. Enterprise has a Compliance API (Smarsh integrates with it), SCIM, audit logs and a BAA; Team has neither the Compliance API nor a BAA. There’s no first-party EU data residency.
OpenAI ChatGPT. Consumer chats train the model until you turn off “Improve the model for everyone”; deleted and Temporary Chats are removed within 30 days. Business, Enterprise and Edu don’t train on your data, but chats are kept indefinitely by default; Enterprise Owners can set retention (minimum 90 days). Enterprise adds SCIM, key management, data residency in ten regions, and a Compliance Logs Platform with integrations including Global Relay, Microsoft Purview and Relativity. Business has no compliance export and no BAA. The API keeps abuse-monitoring logs up to 30 days; zero retention is by approval.
Side-by-side comparison
| Gemini | Claude | ChatGPT | |
|---|---|---|---|
| Consumer trains by default? | Yes (Keep Activity on) | If toggle on (pre-set on) | Yes, until switched off |
| Consumer retention | 18 mo default; reviewed chats up to 3 yrs | 30 days after deletion; up to 5 yrs if training on | Until deleted; 30 days after |
| Business tiers | Workspace with Gemini | Team, Enterprise | Business, Enterprise, Edu |
| Business trains by default? | No | No | No |
| Business retention control | 3/18/36 months; Vault | Enterprise: min. 30 days (default indefinite) | Enterprise: min. 90 days (default indefinite) |
| API default retention | Paid tier not used for improvement | 30 days | Up to 30 days |
| Zero data retention | Vertex AI, configured | By approval (30-day covered-model carve-out) | By approval, eligible endpoints |
| HIPAA BAA | Workspace | Enterprise, API | Healthcare, API |
| Data residency | US/EU (top editions) | US only | 10 regions |
| Archive export | Vault; Smarsh, Global Relay, Mimecast | Compliance API (Enterprise); Smarsh | Compliance Logs (Enterprise); Global Relay, Purview, Relativity |
No vendor wins outright. Google leads on certifications and eDiscovery, OpenAI on residency and archiving integrations, Anthropic on default API retention and consumer/commercial separation. For most firms, the deciding factor is the ecosystem you already run on.
Automation layers change the data flow
When an n8n, Zapier or Make workflow pulls a file, sends it to a model API and writes a result back, your data also passes through execution logs and intermediate storage. Use paid commercial API keys, never free-tier or personal ones. Turn off or prune saved execution data in n8n. Use zero data retention where you qualify, and strip client names, account and matter numbers the task doesn’t need. Finance firms should note that zero retention means the vendor keeps no copy, so the workflow itself must write prompts, outputs and metadata to your archive.
Practical checklist
- Ban personal AI accounts for firm work and block them where you can.
- License a business tier; choose Enterprise for sensitive litigation or regulated finance. Sign the DPA, and a BAA if you handle health data.
- Set an explicit retention period that matches your records schedule, never “indefinite.”
- Enforce SSO and SCIM so leavers lose access the same day.
- Scope connectors and memory to projects, matters or clients.
- Connect AI logs to your legal-hold and eDiscovery process.
- Finance: connect the compliance API to your archive and test it with a mock exam request; add AI vendors to your Reg S-P inventory with 72-hour breach notice; cover consumer AI in your off-channel policy.
- Write an acceptable use policy with a “never upload” list, engagement-letter AI language, and client advisories about consumer chatbots.
- Ask every vendor: Is no-training in the contract? What’s retained, for how long, and can we shorten it? Who can read our data? What happens under a third-party preservation order? Can we export everything, including deleted chats?
The bottom line
All three vendors offer enterprise tiers a careful firm can use for confidential work, and consumer tiers no firm should use for client matters. The real risks are staff on personal accounts, retention left on “forever” (or, for finance firms, records that never reach the archive), and connectors that can see too much. Choose a business tier, configure it deliberately, write the policy down, and recheck the settings every quarter.