Different jobs, different trust models.
This page compares only what each company says publicly.
Can’t. Not won’t.
The others describe what their staff are told not to do. Qoncierge is designed so the people who run it cannot read your content at all.
Verifiable.
The app verifies the hardware and a signed release log before it sends anything. None of the others states that you can verify what code runs.
No training, ever.
We never train on your content. Elsewhere it is an opt-out, often on by default, and flagged content can still be used.
Delete means delete.
Deleting your account destroys the key that unlocks your data. Elsewhere deleted data lingers for weeks, or an assistant may still remember it.
| Question | Qoncierge | ChatGPT | Claude | Instinct | Muse |
|---|---|---|---|---|---|
| What it is | Personal agent: email, calendar, scheduling, research; approval-first | General assistant; “dots” always-on agents[6] | General assistant; consumer apps and API. Duet: proactive agent in its own cloud computer; morning brief; research preview[14] | Personal agent over text, iMessage and WhatsApp; invite-only[7] | Personal agent with its own cloud computer[5] |
| Stated use of your content for training | None. We cannot read content, and keep none in logs | On by default for personal accounts, can be switched off; feedback conversations may still be used; Business and API default to no training[1] | Your choice in Privacy Settings; retention five years if allowed, 30 days if not; flagged chats and feedback may still be used[2][3] | Opt-out only, forward only; flagged content still used; Google Workspace data excluded[13] | On by default; the toggle also covers past interactions; not shared with Meta ad systems[12] |
| Protection against the provider itself reading your data | Designed so it cannot: attested enclave, per-account keys. Staging today; production is pre-launch | No technical bar stated; policy says OpenAI may monitor content for safety and abuse[10] | By default staff cannot access conversations; exceptions are feedback you send and Trust & Safety review. A policy statement, not a technical bar[11] Duet: not stated for stored data. Passwords go through an encrypted form and tokens are kept out of the model’s view[14] | No technical bar stated; shares data with third-party AI model providers; terms authorise indexing of connected-service data[13][4] | Isolated cloud VM; a “Confidential” VM that keeps data private even from Meta is only being designed[12] |
| Can you verify what code runs? | Yes: the app checks hardware attestation and a signed release log before sending anything | Not stated publicly | Not stated publicly | Not stated publicly | Not stated publicly |
| Data after you disconnect or delete | Deleting the account destroys the key; Google grant revoked | Deleted data removed within 30 days, with exceptions for safety and legal reasons; Temporary Chats not used for training[1][10] | Removed from history at once and from back-end systems within 30 days; flagged content may be kept up to 2 years[2][11] Duet: deleting the account wipes everything; disconnecting an app keeps what it learned until you reset its memory[14] | Indexed data from a disconnected service may still be used unless you request deletion; disconnecting an app does not delete its data[4][13] | You can delete messages and files, or reset everything; Muse may still “remember” what it learned[12] |
| Acting on your behalf | One signed approval (a tap, or Face ID for payments) can cover an email, a restaurant booking at places you were shown, and the calendar event; it never pays; undo for five seconds. Booking is in testing | Agents run in the cloud; approval model not assessed here | Not assessed for chat. Duet: asks before sending, paying, signing or deleting; a check outside the model; trading and transfers barred[14] | Terms appoint it as your agent with authority to enter binding agreements; its own policy warns of unintended payments or messages[4][13] | Asks before important actions by default, changeable in Settings; a separate “Sentinel” agent gates what reaches the internet[5][12] |
| Availability in Europe | Europe first; by invitation | Dots for Pro excluded the EEA, Switzerland and the UK at launch[6] | Not assessed for chat. Duet: US waitlist only, Pro, Max and Ultra[14] | No statement about the EU found | Rolling out in the US[5] |
| Governing law and forum | Operating entity and terms not yet final | OpenAI Ireland Limited for the EEA and Switzerland; data processed in the US under standard contractual clauses[10] | Anthropic PBC (US); no Duet-specific terms named[14] | Spear Street Technology, Inc. (US); California law; JAMS arbitration; class actions barred[4][13] | Meta (US); terms follow the Meta Privacy Policy |
Where the big agents launch.
OpenAI
At the September 2026 launch of its dots agent, the Pro version excluded the EEA, Switzerland and the UK.[8]
Anthropic
Its Claude Duet agent is a research preview, US-only, by waitlist for Pro, Max and Ultra subscribers.[14]
Meta
Muse is rolling out in the US. Meta says a Confidential VM will come “later this year”.[9]
The largest agents launch in the US first, and Europe waits. European privacy rules are a large part of why. They are the starting point for Qoncierge: built for them from the first day, not adapted later.
Sources
- OpenAI Help Center, “Data controls in ChatGPT” (consumer: “Improve the model for everyone” setting; Temporary Chats not used to improve models; thumbs-up/down feedback can still be used; Business and API not used by default). https://help.openai.com/en/articles/7730893-data-controls-in-chatgpt (full page text checked 18 Nov 2026; it says feedback conversations may be used for training, Temporary Chats may be kept up to 30 days for safety, and “Trusted Access” sessions may be used even when the setting is off)
- Anthropic, “Updates to our consumer terms” (28 Aug 2025): choice to allow chats to improve Claude; retention five years if allowed, 30 days if not; commercial products and the API excluded. https://www.anthropic.com/news/updates-to-our-consumer-terms (read 17 Nov 2026)
- Anthropic privacy centre, “Is my data used for model training?” (as of 16 Mar 2026): training only on your choice, safety-flagged conversations or explicit programs; incognito chats excluded. https://privacy.claude.com/en/articles/10023580-is-my-data-used-for-model-training (read 17 Nov 2026)
- Instinct Terms of Service, last revised 26 Aug 2026: training on Materials permitted excluding Vault content; opt-out applies going forward and not to safety-flagged content; indexed data from connected services may still be used after disconnecting unless deletion is requested; California law; JAMS arbitration. https://instinct.com/terms (read 17 Nov 2026)
- Meta Newsroom, “Introducing Muse” (8 Sept 2026, updated 10 Nov 2026): dedicated cloud computer per person, no sharing of conversations or VM data with ad systems, training opt-out, US rollout, Confidential VM planned. https://about.fb.com/news/2026/09/introducing-muse-personal-ai-agent/ (read 17 Nov 2026)
- OpenAI release notes on dots (29 Sept 2026) as quoted by MIXED. https://mixed-news.com/en/openai-chatgpt-dots-exclude-pro-eea-switzerland-uk/ (read 17 Nov 2026)
- Vellum, “Instinct breakdown” (a competitor’s description: invite-only, operated by Spear Street Technology, Inc.; channels include iMessage and WhatsApp). Used only for what Instinct is and its channels, which we could not confirm on Instinct’s own pages. https://www.vellum.ai/blog/official-instinct-breakdown (read 17 Nov 2026)
- OpenAI release notes (29 Sept 2026) as quoted by MIXED: dots for Pro “excludes the European Economic Area, Switzerland, and the UK at launch”; Business Premium “across all supported ChatGPT regions”. https://mixed-news.com/en/openai-chatgpt-dots-exclude-pro-eea-switzerland-uk/ (read 17 Nov 2026; OpenAI’s own page blocked automated access)
- Meta Newsroom, “Introducing Muse” (8 Sept 2026, updated 10 Nov 2026): “Muse is rolling out in the US”; Confidential VM “later this year”. https://about.fb.com/news/2026/09/introducing-muse-personal-ai-agent/ (read 17 Nov 2026)
- OpenAI, “Europe privacy policy” (updated 24 August 2026): content may be monitored to prevent fraud and misuse; deleted data removed from systems within 30 days unless retained for safety, security or legal reasons; Temporary Chats deleted automatically within 30 days; controller for the EEA and Switzerland is OpenAI Ireland Limited; data processed outside the EEA, Switzerland and the UK, including in the United States, under adequacy decisions and Standard Contractual Clauses. https://openai.com/policies/eu-privacy-policy/ (full text checked 18 Nov 2026)
- Anthropic Privacy Center, consumer articles: “How long do you store my data?” (updated 1 Jul 2026), “Is my data used for model training?” (16 Mar 2026), “How does Anthropic protect the personal data of Claude users?” (16 Mar 2026) and “Can you delete data sent via Claude?” (16 Mar 2026). https://privacy.claude.com/en/collections/10663362-consumers (full text checked 18 Nov 2026; the GDPR page does not name the controller entity, processing location or transfers)
- Meta, “Muse Privacy Policy” (effective 16 Nov 2026): training setting on by default and also applies to previous interactions; conversations and VM data not shared with Meta ad systems; Confidential VM planned for the future; important actions need approval by default; deleted items may still be “remembered”; data may be combined across profiles if Muse is added to the same Accounts Center. (full text checked 18 Nov 2026)
- Spear Street Technology, Inc. d/b/a Instinct, “Privacy Policy” (revised 26 Aug 2026): training opt-out is forward only and flagged content may still be used; Google Workspace data not used for training; shares information with third-party AI model providers; disconnecting an integration does not delete its data; warns of unintended payments or communications and of misleading instructions aimed at agents. (full text checked 18 Nov 2026)
- Anthropic, “Claude Duet” (21 Oct 2026, updated 23 Oct) and “Claude Duet: Safety, Privacy, and Security” (23 Oct 2026): research preview for US Pro, Max and Ultra waitlist; own cloud computer; approval before sending, paying, signing or deleting; deleting the account wipes everything stored; the pages do not state training use, who at Anthropic can access data, or regions beyond the US. https://claude.com/resources/articles/claude-duet-safety-privacy-and-security (full text checked 18 Nov 2026)