What “Proactive” Actually Means — and What These Agents Cannot Do

All three products carry the word “proactive,” but they mean different things. Claude Duet sends a morning brief of priority items across your connected apps and works on tasks in the background. OpenAI’s Dots run persistently in their own cloud computers, scanning connected apps when you are inactive and returning completed work for approval. Meta Muse remembers your preferences and fires unprompted suggestions, while keeping its agent in an isolated virtual machine per user.

None of them operate autonomously. All three require user approval before consequential actions — sending a message, making a payment, deleting a file, signing a document. The autonomy is real for research, drafting, and monitoring tasks; it stops at the last irreversible step. This matters because most enterprise risk questions dissolve once you understand where the guardrail sits.

What separates them is not the concept but the architecture: who runs the compute, where the data goes, who holds the credential, and what the enterprise admin can see. Those four questions should drive your evaluation more than the demo video.

How the Three Platforms Work

Claude Duet (Anthropic, October 21, 2026)

Duet is a research preview, US-only, available to Claude Pro and Max subscribers on a waitlist. It runs as part of the existing Claude conversation thread — one persistent thread that remembers earlier requests. A built-in classifier approves agent actions, flags sensitive information before it leaves, and confirms important steps with the user. Anthropic describes this as the same trustworthy-agent framework that governs Claude Code and Claude in Chrome.

The scope is explicitly work-focused. Duet prepares a morning brief of priority items, monitors messages and alerts only when something needs attention, builds dashboards, creates forecasts, and produces documents, spreadsheets, or slides for review. It accepts text or voice inputs and runs through Claude’s connected skills and apps ecosystem, which the company lists at over 6,000 connectors.

What Duet does not yet have: a published admin console, team or enterprise plan availability, or EU data residency. The waitlist opened October 21; Anthropic has not said when it will exit preview or what happens to the price. Per Layer3 Labs, usage shares Claude Pro’s weekly and five-hour session limits, meaning heavy background activity cuts into the same budget as your regular Claude sessions.

ChatGPT Dots (OpenAI, September 29, 2026)

Announced at OpenAI DevDay 2026, Dots are persistent agents — each gets its own cloud computer and browser, runs on GPT-6 Astra, and can handle multiple projects simultaneously without separate conversation threads. They connect to over 4,000 applications through OpenAI’s plugin ecosystem and communicate through ChatGPT, Slack, and Microsoft Teams. According to VentureBeat, Dots can also connect directly to a laptop when given permission.

The safeguard architecture is layered. Proactive research tools are read-only and cannot send messages, edit content, or control the computer. An auto-review system checks consequential actions against user instructions, Custom Rules, and built-in safety requirements before deciding whether to proceed or ask. Users get an Activity View to review everything the Dot did in the background.

Dots also introduce ChatGPT Space — a collaborative layer where employees, ChatGPT, and Dots work from shared context. Specialist Dots get their own organizational identity, credentials, and system access for defined business roles. OpenAI is testing these internally for procurement, invoice processing, and customer support, and working with Microsoft to integrate them into Microsoft Agent 365.

Meta Muse (Meta, September 8, 2026)

Muse launched September 8, 2026 in the US on iOS, Android, and muse.ai, powered by the Muse Spark model. The key architectural differentiator is the Muse Secure VM: each user gets a dedicated virtual machine that holds both the agent and their data, isolated from all other users. A separate Sentinel agent runs on the same machine at system level and must approve any Muse activity that reaches the internet, per Meta’s announcement.

Credentials are stored inside the VM so Muse can use them without seeing them. Checkout uses Link by Stripe, generating one-time cards so real payment details stay hidden. A full audit trail of past and planned actions is available. Users can also instruct Muse to “forget” specific items, and the company says a Muse Confidential VM — where data is encrypted with a key only the user holds — is planned “later this year.”

What Muse does not have is an enterprise tier. It is a consumer product. There are no admin controls, no team policies, and no org-level visibility. Meta has confirmed no European launch date; the most recent market expansion was Singapore on October 27, 2026.

The Numbers: Pricing and EU Access Side by Side

Here is what you actually pay and what you get in each region, based on published pricing as of October 31, 2026:

Feature	Claude Duet	ChatGPT Dots	Meta Muse
Entry price	$20/month (Pro)	$100/month (Pro 100)	Free (usage-capped) / $20/month Power
Team / enterprise tier	Not available yet	Business Premium ($100/seat/mo annual)	None
Minimum EU-accessible setup	Not confirmed	$120/month annual (1 Premium + 1 Standard seat)	Not available in EU
EU data residency	No (all US infrastructure)	Yes for Enterprise/Edu (in-region GPU since Jan 2026); unclear for Dots specifically	No EU availability
GDPR transfer mechanism	SCCs + EU-US DPF (Team/Enterprise/API only)	SCCs + EU-US DPF; DPA available	N/A (not in EU)
Admin controls / audit log	None published	Activity View, Custom Rules (Business)	Per-user audit trail (no org layer)
Model	Claude (unspecified version)	GPT-6 Astra	Muse Spark
Approval model	Classifier + user confirmation for major actions	Auto-review + Custom Rules + user approval gates	Sentinel VM agent approves outbound actions
Verdict	Watch (no enterprise controls yet)	Test (EU-accessible via Business Premium)	Skip for professional use

Sources: Layer3 Labs Duet pricing; Eesel Dots pricing; Layer3 Labs Muse pricing; OpenAI EU residency page; Sonomos Claude EU residency, Sep 2026. All prices USD. October 31, 2026.

The Permission Model Is the Real Architecture Decision

Enterprise security teams spend time on model quality benchmarks and miss the more important question: what can the agent do without asking, and how does your org control that?

Dots give the clearest answer. Custom Rules let an administrator specify which actions a Dot can take autonomously, which require user confirmation, and which are banned. The Activity View gives every user a timestamped record of background activity. For Business and Enterprise plans, workspace content is not used for model training by default — an explicit opt-out is not required. VentureBeat’s DevDay reporting notes that sensitive actions such as changing passwords stay with the human by design.

Claude Duet’s permission model is less documented today. Anthropic publishes that a classifier checks actions for sensitive information and confirms important steps, but without admin controls or team plans, there is no way to set org-level rules. A Duet running on a personal Claude Pro subscription is outside your DLP perimeter entirely. This does not make it unsafe — Anthropic’s trustworthy-agent framework is well-regarded — but it makes it unauditable for enterprise purposes.

Meta Muse’s Secure VM architecture is the most interesting from a technical standpoint. An isolated VM per user with a dedicated Sentinel agent at the system level is a stronger isolation model than either Dots or Duet on paper. But there is no org layer: no admin can see what a team member’s Muse is doing, can set policy, or can revoke access from a management console. That is a fundamental gap for enterprise IT governance, and it reflects Muse’s consumer focus.

The Data Privacy Reckoning for Enterprise Teams

When your Duet reads your email, that email content is processed by Anthropic’s US-based infrastructure. Per Sonomos’s September 2026 analysis, Anthropic offers no EU data residency across any tier — Free, Pro, Team, Enterprise, or API. Every inference call goes to US infrastructure. EU-to-US transfers of personal data via Anthropic require a valid Chapter V GDPR mechanism; Anthropic uses Standard Contractual Clauses and participates in the EU-US Data Privacy Framework, but only customers on Team, Enterprise, or the API get a Data Processing Agreement.

OpenAI’s story is more nuanced. ChatGPT Business and Enterprise customers can have data processed in-region in Europe, and OpenAI’s EU residency announcement extended in-region GPU inference to eligible Enterprise customers in January 2026. The catch: OpenAI has not explicitly confirmed whether Dots workloads are covered by this residency guarantee, or whether they use the same infrastructure path. Given that Dots run on dedicated cloud computers per agent, it is a reasonable question to raise with your OpenAI account team before deploying.

For Claude Duet, the training data question is also open. Anthropic does not train on API data and excludes Team/Enterprise conversations by default — but Duet runs on consumer plans (Pro and Max), where training terms differ. Check Anthropic’s current consumer data terms before connecting a Duet to shared business inboxes.

For Swiss & EU teams

EU teams have exactly one path to a proactive agent with addressable data controls today: ChatGPT Business Premium. According to Eesel’s Dots pricing breakdown, Pro plan access to Dots is explicitly excluded in the EEA, Switzerland, and the UK — only a Business workspace qualifies there. The minimum cost is $120/month annual for one Premium seat plus one Standard seat.

That is not a full solution. OpenAI has not confirmed that Dots’ background cloud-compute workloads fall inside the EU data residency guarantee that covers ChatGPT Enterprise conversations. Until OpenAI publishes a DPA amendment covering Dots specifically, treat the data residency as unconfirmed for agent workloads.

Meta Muse is not available in Europe, and Meta’s track record on EU AI data processing — including GDPR challenges over training data — makes an imminent compliant EU launch unlikely. Per The Model Wire, the platform has already drawn regulatory attention over its opt-out consent approach to using Instagram photos in agent tasks, which conflicts with GDPR’s requirements for explicit consent for new processing purposes.

Claude Duet has no published EU timeline or Swiss-specific data residency. Anthropic does have FADP coverage through its SCCs (which adequately cover CH-to-US transfers given Switzerland’s adequacy with the EU framework), but no EU infrastructure option is available. The AWS Bedrock and Google Vertex AI routes for Claude API do offer EU region selection, but those do not currently include Claude Duet’s consumer product features.

In short: no proactive consumer AI agent is fully ready for enterprise-grade EU data processing today. The closest option is ChatGPT Business Premium with Dots, but with unconfirmed residency for agent workloads and a $120/month floor.

What Enterprises Actually Need That None of These Provide

Reading the enterprise requirements that engineering teams and IT leads actually bring to vendor discussions, three gaps stand out across all three platforms.

First, outbound data classification. None of the three let an admin define which data categories — personal data under GDPR, health information, financial records, source code — cannot flow through the agent’s cloud infrastructure. Dots’ Custom Rules let you prohibit specific actions but not specific data types. Until DLP-aware permission models arrive, the safest deployment is connecting agents only to low-sensitivity inboxes or project channels, not to your main engineering Slack or your M365 tenant.

Second, credential governance. All three store credentials or use delegation tokens. For Duet and Muse, there is no org-level credential revocation — if an employee leaves, IT cannot revoke the agent’s access without the employee revoking it themselves. Dots on Business Premium does better here, because existing ChatGPT workspace permissions and app controls govern what a Dot can access; IT can deprovision the account and the Dot loses access.

Third, incident attribution. When a Dot or Duet sends a message or modifies a file, the audit trail needs to state whether it was the agent or the human. ChatGPT’s Activity View surfaces this at the user level. None of the three yet export this data to a SIEM or a centralized audit log in a standard format. This is a gap for any team subject to NIS2, the EU AI Act’s high-risk provisions, or FADP audit requirements.

These gaps do not make proactive agents unusable. They make clear which workloads they are ready for: individual productivity (scheduling, research, summarization, drafting) on low-sensitivity contexts. Connecting an agent to production systems, customer data pipelines, or HR inboxes is a different and higher-risk configuration that none of the three products has earned yet.

For the broader context on shadow AI agents and governance, see our earlier analysis: Shadow Agents: 81% Deployed, Only 14% Secured. And for security hardening of the coding-agent subset, the 2026 coding agent pipeline security guide covers credential handling and network controls in detail.

What to Do on Monday

1. If you are a US-based team that wants to trial proactive agents: Request Duet access on your personal Claude Pro account and test it against a non-sensitive mailbox or calendar for two weeks before connecting anything work-critical. Note where it hits session limits — those limits are shared with your regular Claude usage.
2. If you are evaluating Dots for enterprise use: Open a ChatGPT Business workspace, add one Business Premium seat and one Standard seat ($120/month on annual billing), and ask your OpenAI account team directly: “Does our EU residency commitment cover Dots’ background compute workloads?” Get the answer in writing before broader rollout.
3. Write a one-page agent data policy before connecting anything. Define: which data categories may not flow to a proactive agent’s cloud; which inboxes and calendars are in scope; who can grant and revoke agent access; and how incidents will be attributed in your audit log. This takes one afternoon and it forces the right conversations.
4. Map your credential exposure now. For any account you connect to Duet or Dots, document what that account can read and write. OAuth tokens issued to AI agents are a new attack surface — see Microsoft Guardian Agents for the emerging tooling to detect and scope these.
5. Skip Meta Muse for any professional evaluation in 2026. No EU availability, no enterprise controls, active regulatory scrutiny in Europe. Revisit in 2027 if a business tier launches.

Further Reading

• OpenAI Dots and ChatGPT Space, VentureBeat — The most detailed account of Dots’ architecture, use cases, and the Specialist Dots enterprise pilot.
• Claude EU Data Residency in 2026, Sonomos — A thorough breakdown of the EU transfer mechanisms and what “no EU residency” means in practice for each Claude tier.
• OpenAI EU Data Residency — OpenAI’s official residency page; check the January 2026 update for the in-region GPU inference announcement.

Your turn: Has your team connected any proactive AI agent to a company inbox or calendar yet? What did your data team say when you asked about the privacy terms? Reply to our newsletter or send us a note — we feature the best answers in the Friday Scorecard.

How this article was made: AI researched and wrote this article under the Christian persona, using the sources linked above, and it was published automatically without a human edit. Editorial guidelines are set by Adi. Spotted an error? Tell us and we will correct it.