Three work assistants, compared for regulated firms
On 22 October, Anthropic launched Claude Duet, the third serious work assistant regulated firms now have to weigh against OpenAI’s ChatGPT Work and the Microsoft Copilot most already pay for. Two of them need you to bring your data to them. The one you already pay for sits inside it, and for a regulated firm that is most of the argument. Claude Duet is built on Opus 5.5; Anthropic says it runs morning briefings, watches your connected apps, and builds dashboards and web apps while you get on with something else[2]. ChatGPT Work is a desktop agent with its own virtual computer and browser[1]. Copilot lives inside the 365 tenancy, which matters more than any feature list.
There will be plenty written this week about which model is cleverer. The Artificial Analysis Intelligence Index, an independent benchmark, reports workflow-specific gaps: one tool leads on some tasks, another elsewhere[3]. No single model holds a general lead. For regulated admin work, cleverness is the second question anyway. The first is where your client data has to go for the work to happen.
Where your data already sits
Judged on data residency first, Copilot starts ahead. Copilot runs inside your existing Microsoft 365 tenancy. The client data it reads stays inside the boundary you already govern. Its activity is logged through Microsoft Purview, which records audit trails and screenshot-based session replay for traceability[4]. That is the bar the other two have to clear, and it holds because the data never moves.
The newcomers ask it to move. OpenAI says its enterprise customers can set EU and UK data residency and switch on admin audit logging, so ChatGPT Work is not starting from nothing[1]. Claude Duet is days old, and its enterprise controls are as announced rather than independently audited. Anthropic’s Opus line, though, has shipped with no mandatory data retention, the thing regulated buyers kept stalling on[2]. On paper both can be configured safely. With Copilot, the safe set-up is where your data already lives.
With Copilot the safe configuration is the default. For the newcomers it is something to configure, and then prove.
What each is allowed to touch
An agent is only as safe as the systems it can reach. Nufar Gaspar, an AI implementation specialist, puts it plainly: “A team agent only knows what’s written down in a place that it can reach.” So the live risk is what you have let it open, and whether anyone mapped that before switching it on.
Giving a model its own browser raises the stakes. A read-only chatbot that drafts an email carries far less risk than an agent that can log into systems and fill in forms. The US law firm Cooley has built ‘GO Public’, a ChatGPT-integrated tool that drafts IPO documentation and flags issues[5]. It is a US firm, but it shows the direction: the more an agent can reach, the more the firm is exposed.
Open questions for all three tools
None of the vendors has yet closed these.
- Liability when an agent goes wrong. There is still a legal vacuum over who answers for a rogue agent’s actions, which may push cautious firms to keep agents in sandboxed, low-consequence work for now[6]. That gap applies to Copilot as much as to the newcomers.
- An auditable record inside the isolated browsers. Both ChatGPT Work and Claude Duet run isolated browser environments that research, fill forms and execute tasks on their own[1]. Whether that produces a tenant-level, replayable record to match Purview’s session replay is not yet demonstrated for either.
- Retention guarantees you can evidence. A no-retention tier in a contract is not the same as a logged, checkable boundary your compliance function can inspect on demand.
Isolated environments are useful, but they are not yet the same as an audit trail held in your own tenancy.
The verdict for a firm already in Microsoft 365
For most advice firms, the ones already working inside Microsoft 365, the verdict turns on where the data sits rather than which model is cleverer. Data that never leaves a governed tenancy, logged where your compliance function already looks, starts the conversation ahead. The rival agent needs the data to travel first, whatever its measured capability.
A firm on Google Workspace, or one whose client data lives in a cloud case-management system rather than a Microsoft tenancy, faces a different calculus. For them Copilot’s home advantage disappears, and residency has to be assessed from scratch against whichever tool sits closest to the data. (Our Copilot vs ChatGPT comparison for regulated firms goes into the configuration detail.)
For a firm in a Microsoft tenancy, Copilot wins on governance before the comparison starts, and the newcomers have to earn their way past it with evidence, not features.
This article is for informational purposes only and does not constitute regulated financial advice or a compliance opinion. Consult a qualified compliance professional for advice specific to your firm.
Sources
[1] OpenAI, ‘ChatGPT Work’, OpenAI, October 2026. [Cited for ChatGPT Work’s virtual computer and browser, isolated environment, and enterprise residency and audit options.] [2] Anthropic, ‘Claude Duet’, Anthropic, 22 October 2026. [Cited for the Claude Duet launch, its Opus 5.5 base, and the Opus line’s no mandatory retention position.] [3] Artificial Analysis, ‘Artificial Analysis Intelligence Index’, 2026. [Cited for workflow-specific performance gaps between models.] [4] Microsoft, ‘Microsoft Purview audit logging and session replay’, Microsoft, 2026. [Cited for tenant-level audit trails and screenshot-based session replay.] [5] OpenAI, ‘Cooley GO Public’, OpenAI, 20 September 2026. Available at: https://openai.com/index/cooley-gopublic. [Cited as a US, directional example of agent reach into document drafting.] [6] MIT Technology Review, ‘Who’s liable when AI agents go rogue?’, MIT Technology Review, 28 September 2026. Available at: https://www.technologyreview.com/2026/09/28/1145197/whos-liable-when-ai-agents-go-rogue/. [Cited for the unresolved liability position for autonomous agents.]