Always-On AI Agents for Finance Teams: dots, Muse and Copilot Autopilot Explained

What Always-On AI Agents Are, and What Changed

For three years, AI assistants have worked like a very capable colleague who only acts when spoken to. You write a prompt, it answers, and nothing happens until you write the next one.

Always-on agents work differently. You give them a goal rather than a single instruction. They keep working towards it over hours or days, using their own computing environment, a browser and the apps you connect. They remember what they have done, follow up on their own, and only come back to you when they need a decision or approval.

Several launched within weeks of each other. Axios described it as the start of a race between the major AI companies to own the personal assistant. For finance teams, three matter most: OpenAI's dots, Meta's Muse and Microsoft's Copilot Autopilot. Others, including Google's Gemini Spark and the startup assistant Instinct, follow the same pattern.

The shift matters because it moves AI from drafting to doing. An assistant that drafts a supplier email creates no risk until a person sends it. An agent that sends the email, updates a record or chases an approval on its own acts on your behalf, and that is a very different question for a function built around controls.

OpenAI dots, Meta Muse and Copilot Autopilot Compared

OpenAI dots. Announced on 29 September, dots are always-available agents you assign work to through ChatGPT, Slack or Microsoft Teams. Each runs on OpenAI's GPT-6 Astra model with its own cloud computer and browser. They learn your preferences, ask permission before sensitive actions, and follow rules you set about what they can and cannot do. One dot is included with ChatGPT Pro and Business Premium plans. Important for UK readers: at launch, Fortune reported that dots are not available in the UK, the European Economic Area or Switzerland, citing data privacy concerns.

Meta Muse. Launched in the US in September, Muse runs from a dedicated virtual machine and works across the apps you connect to it. It is free up to a usage limit, with paid plans for heavier use. It is aimed mainly at personal use rather than finance operations.

Microsoft Copilot Autopilot. This is the one most UK finance teams are likely to meet first. Autopilot, previously known as Scout, lives inside your Microsoft 365 tenant with its own identity, memory and workspace. It can watch conversations, pursue updates, handle recurring responsibilities and pick a project back up after several days. In Microsoft's example, an agent pulled information from email, Teams, Dynamics 365 inventory records and spreadsheets to spot a supply problem. It needs a Copilot licence plus usage charges in Copilot Credits, and its private preview was expanding at the end of September.

Where Always-On Agents Could Help a Finance Team

The best early uses are tasks that are mostly chasing, checking and coordinating, where the agent gathers and prepares and a person still decides. Good candidates:

• Month-end coordination: tracking which journals, reconciliations and sign-offs are outstanding, and reminding owners as deadlines approach.
• Approval chasing: following up purchase orders or invoices that have sat in someone's queue, with a summary of what is waiting.
• Supplier statement checks: comparing statements against the ledger each week and listing differences for review.
• Board pack preparation: collecting the inputs each department owes, and flagging what is missing two days before the deadline.
• Monitoring: watching for a covenant metric, a cash balance or a budget line moving outside a range, and alerting the owner.

Notice what is not on the list: posting journals, changing supplier bank details, releasing payments or editing master data. Those are exactly the actions where an error or a manipulated instruction costs real money, and where your controls exist for a reason. Agents can prepare those actions for a person to approve, but in our view they should not perform them unattended, at least until your organisation has a track record and a tested control framework.

This mirrors how we approach any automation with clients. Start with work where the agent's output is checked before it matters, prove it on a contained process, and widen its permissions only when the evidence supports it.

Five Controls to Set Before an Agent Touches Finance

Microsoft's chief executive put the principle well when launching Autopilot: every agent needs an identity, and everything it does needs to be observed. In practice, that means five controls.

1. A named owner. Every agent belongs to a person who is accountable for what it does, just as a system or a spreadsheet would.

2. Least-privilege access. Connect only the systems and folders the task needs. An agent chasing approvals does not need write access to the ledger.

3. Human approval for anything that moves money or changes records. Payments, bank details, journals and master data should always require a person to approve, ideally outside the agent's own interface.

4. An audit trail. You need to be able to answer, months later, what the agent did, when and on whose instruction. Check what logging your platform provides before relying on it.

5. A defined scope and an off switch. Write down what the agent is for, review it monthly, and know how to stop it immediately.

Prompt injection deserves a specific mention. An agent that reads emails and documents can be manipulated by instructions hidden inside them, such as a fake supplier email asking it to update payment details. Approval controls are your defence. Our AI governance framework for finance covers this in more depth.

What UK Finance Teams Should Do Now

For most UK finance teams, the practical position this quarter is straightforward.

If you are on Microsoft 365, ask IT whether your organisation is in the Autopilot preview or plans to join, and make sure finance is involved in setting its permissions before it is switched on. That conversation is far easier before an agent has access than after.

Do not wait for agents to get value from AI. Most of the benefit available today still comes from people using assistants well on everyday finance tasks. The skills that make an agent safe and useful, writing a clear goal, defining what good looks like and checking the output, are the same skills that make a prompt work. Our guide to outcome-first prompting is a good place to start.

Write your agent policy now. A one-page statement of what agents may and may not do in finance, who approves their access and how their actions are reviewed will save you from making those decisions under pressure later.

Watch UK availability. Several agents are launching in the US first, and some, like dots, are not yet available in the UK. Expect that to change, and expect your colleagues to try them on personal accounts before your organisation approves anything. A clear policy gives them a safe alternative.

Frequently Asked Questions