Pick by how your firm works
| If this is true | Start here |
|---|---|
| You can choose freely | A Claude agent: Duet if you are in the US on a personal plan, otherwise one we build on Claude Managed Agents. |
| Your firm has to stay inside Microsoft 365 | Microsoft Autopilot when it reaches you. Budget for Copilot Credits. |
| Your team already works in ChatGPT | ChatGPT dots on Business Premium seats. |
| You run on Gmail and Google Workspace | Gemini Spark where it is offered to you; watch the Workspace agent. |
| Most of your sales come through Instagram, Facebook or Shopify | Test Meta Muse on marketing and order follow-ups. |
| The work is in websites or apps with no connector | Grok Bot. |
| Data must stay on machines you control | Hermes Agent, self-hosted, with someone to maintain it. |
For the bigger choice of which AI your whole team should use day to day, read Copilot vs Claude vs ChatGPT: which is safer for business?
What we recommend, and why
An always-on agent keeps working after you close the app. You give it a goal and the apps it may use, and it checks in, runs recurring jobs and brings finished work back to you. Most of the big AI companies launched one between August and October 2026, so the choice is new for everyone.
Our first choice is Claude, for four reasons:
- The biggest share of business AI spending goes to it. Anthropic, which makes Claude, now earns 55% of what large US companies spend to build on AI models, ahead of OpenAI at 31% and Google at 9% (Menlo Ventures survey, November 2026; Menlo is an investor in Anthropic). In practice, that makes connectors, add-ons and people who know how to set it up easier to find.
- Staff find it the easiest to use. That is what we see across our clients. An agent only helps if the person it works for is comfortable checking its work.
- It stays on task. In our client work, Claude holds a long job together from start to finish, while GPT models drift off track more often on long tasks. An agent spends most of its life on long tasks.
- It gives the most for the money. A fixed monthly Claude plan buys far more work than the same spend on ChatGPT, when we price the usage you get at each company's own pay-per-use rates. Microsoft is moving the other way: Copilot's agent features are billed per use in Copilot Credits on top of the seat (Computing, reporting Microsoft's plan), so costs rise with use.
The catch today: Claude Duet, Anthropic's ready-made agent, is a US-only research preview (an early test version) on personal plans. For firms outside the US, or a team that needs one agent shared across people, we build the agent on Claude Managed Agents instead. It takes setup work, which is where we come in.
The eight compared
Checked on each maker's own page on November 14, 2026. These products change monthly, so check the link for your country before you plan around one.
| Agent | Best fit | Where and on which plans | Asks before it acts? | How you pay |
|---|---|---|---|---|
| Claude Duet Anthropic | An owner or professional who wants one agent across email, calendar and files. | US only, waitlist. Personal Pro, Max and Ultra plans. Research preview (early test version). | Yes. Shows you the action and waits before it sends, pays, signs or deletes. | Included in the plan, from the same usage allowance. |
| Claude Managed Agents Anthropic | A firm that wants an agent built around one job, such as follow-ups or reporting. | Anywhere the Claude API is offered. Beta. | You set its permission rules when it is built. | Pay per use through the Claude API. Needs someone to build it. |
| ChatGPT dots OpenAI | A team already working in ChatGPT, Slack or Teams. | Pro, outside the EU area, Switzerland and the UK. Business Premium (OpenAI's higher-usage business seat) in every ChatGPT region. Enterprise: on by default from November 18, 2026, unless the admin opts out. | Your rules: act, act if pre-approved, ask first, or hand off to you. | Included on Pro and Business Premium (US$100 a seat a month billed yearly, US$125 monthly). Enterprise: billed per use from November 18, 2026. |
| Grok Bot xAI | Work in websites and apps that have no connector (a ready-made link an agent can use), signed in the way a person would. | Personal X Subscription plans (Premium and up) and an enterprise version. | Prepares the work and waits for your okay before it goes out. | Included in Premium, Super and Ultra. Plus does not include it. |
| Meta Muse for Small Business Meta | A shop that sells through Instagram, Facebook, Shopify or Stripe. | US and Canada. | Yes. Nothing publishes, sends or spends without your approval. | Free for most needs, with paid plans for more use. |
| Microsoft Autopilot Microsoft | A firm that has to stay inside Microsoft 365. | No rollout date yet. Will need a Microsoft 365 Copilot licence and a spending policy set by your admin. | Admin controls not yet published. | Billed per use in Copilot Credits (Microsoft's pay-per-use units), on top of the Copilot seat. |
| Gemini Spark and Gemini agent | A person or firm that runs on Gmail and Google Workspace. | Spark: personal Google AI plans, as reported (Pro in the US, Ultra in more countries). The Workspace agent: private preview. | Not stated on Google's pages we could read. | Spark: included in Google AI Pro or Ultra. Workspace agent: not published yet. |
| Hermes Agent Nous Research | A firm that wants to host its own agent and choose its own AI model. | Anywhere. Runs on your own Mac, Windows or Linux machine, or Nous Portal's cloud. | Whatever you configure. The maker's page does not describe an approval step. | Free and open source. Nous Portal plans run from US$0 to US$200 a month. |
Not sure which fits? The free AI Opportunity Map asks 10 short questions about how your firm works and shows where an agent would save the most time.
Where each one wins
Claude Duet
The most careful of the ready-made agents. It shows you each action before it sends, pays, signs or deletes, and some money moves are off-limits even with your approval (Anthropic). It uses your existing Claude memory, skills and connected apps. Its limit is reach: US only, by waitlist, on personal plans.
Claude Managed Agents
The way to get a Claude agent built around one job for your whole firm, wherever you are. Anthropic runs the computer it works on; someone still has to define the job, connect the apps and set the rules (Anthropic docs). It suits recurring work like client follow-ups, weekly reports or sorting a shared inbox. Anthropic notes it is not yet covered by zero data retention or HIPAA agreements, which matters for health and some regulated work.
ChatGPT dots
The strongest pick for a team already living in ChatGPT. A dot gets its own cloud computer, connects to the apps you choose and can be reached in ChatGPT, Slack or Teams. You set custom rules for what it may do alone and what needs your okay (OpenAI help centre). One thing to know before you connect your email: OpenAI says disconnecting an app does not delete what your dot already learned from it.
Grok Bot
Wins when the work lives in websites and apps with no connector. Bots keep their logins and work the way a person would, then bring drafts back for your okay before anything goes out (xAI docs). It comes with personal X Subscription plans, Premium or higher (xAI plans), and firms with stricter needs can use the enterprise version. Keep client data out of a personal plan.
Meta Muse for Small Business
Built for small shops that sell through Instagram and Facebook. It connects to Meta ad accounts and Pages plus QuickBooks, Shopify, Stripe, Slack, Canva and other apps, with custom connectors for the rest, and Meta says nothing publishes, sends or spends without your approval (Meta). It is free for most needs.
Microsoft Autopilot
The natural fit for a firm that has to stay inside Microsoft 365. It is an always-on agent in the Copilot app, managed by your Microsoft 365 admin. You need a Microsoft 365 Copilot licence, your admin has to set a spending policy, and use is billed per task in Copilot Credits (Microsoft admin notice). Microsoft has not announced when it reaches everyone.
Gemini Spark and the Gemini agent
Google has two. Gemini Spark is the personal one: it runs around the clock on Google's cloud and works across Gmail and the rest of Workspace. It reached Google AI Pro subscribers in the US this summer, with Ultra subscribers in more countries, except the EU area, Switzerland, the UK and Nigeria (Digital Trends; Tech2, both reporting Google). The Gemini agent is the company version: one agent that plans work across your business systems and keeps going for hours or days (Google), in private preview with wider release promised for some Workspace Business and Enterprise plans (9to5Google). If your firm runs on Google, start with Spark for one person and wait for the Workspace agent before a team rollout.
Hermes Agent
The pick for control. It is free, open source and can run on your own machine, so conversations and memories stay in a folder you own, and you choose which AI model it uses from more than 200 (Nous Research). The trade is upkeep: you install it, update it and decide what it may do. We ran Hermes on our own server for a time before switching it off. For most small firms, we find the upkeep outweighs the control. OpenClaw is the other well-known open-source option and carries the same trade.
Personal errand agents such as Instinct are left out on purpose. They are built for booking and buying for one person, and none is meant for company accounts yet.
What hands-on tests show
The fullest public test so far is Claire Vo's: nine real tasks across Grok Bot, ChatGPT dots, Meta Muse and Claude Duet (How I AI, October 29, 2026). Grok Bot ranked first overall. Muse was her favourite for personal and creative jobs. Dots and Duet did the strongest coding work. Duet was slow and asked for permission four times where the others asked once, which she says can turn delegating into supervising. She is clear about the limits: the judging was not blind, and each agent had different apps connected.
We read that as a fair warning and a real trade. Duet's extra check-ins are the same caution that makes us comfortable pointing it at a client inbox, and they cost you time. For a firm, we would rather start careful and loosen the rules than clean up after an agent that acted too freely. If speed on messy web work matters most, Grok Bot is the one to test.
Before you switch one on
- Pick one job. A recurring task with a clear finish, such as chasing missing client details every Monday. An agent with a vague goal makes vague work.
- Connect only what that job needs. Start read-only, with sending and spending turned off. Agents can form memories from everything they read, so less access is less to clean up later.
- Keep a person approving what goes out. Emails, posts, payments and anything a client sees. Loosen it only after a few weeks of work you would have sent as is.
- Use a business plan for company work. Personal plans have different data terms. What changes when you connect AI to your business covers the settings.
Common questions
What is an always-on AI agent?
An AI assistant that keeps working after you close the app. You give it a goal, connect the apps it may use, and it checks in, runs recurring jobs and brings work back for your review. A chat tool waits for your next message; an agent keeps going on its own computer in the cloud.
Which always-on AI agent is best for a small business?
We recommend a Claude-based agent whenever you can choose. Where Claude Duet is not offered yet, we build a custom Claude agent for you on Claude Managed Agents. If your firm lives in Microsoft 365, ChatGPT or Google Workspace, the agent from that company is usually the easier start. If you sell mostly through Instagram and Facebook, Meta's Muse is worth a test.
Will an AI agent send emails or spend money without asking?
It depends on the agent and how you set it up. Claude Duet and Meta's Muse say they wait for your approval before anything is sent, published or paid. ChatGPT dots lets you set rules for what needs your okay, and Grok Bot brings drafts back before they go out. Self-hosted agents like Hermes do what you configure them to do. Start every agent with sending and spending turned off.
Can I use these outside the United States?
Only some of them. Claude Duet is US only for now. ChatGPT dots on the Pro plan skip the European Economic Area (the EU plus Iceland, Liechtenstein and Norway), Switzerland and the UK, though Business Premium seats get it in every region ChatGPT supports. Meta says Muse is in the US and Canada, and Google's Gemini Spark skips the same European countries and Nigeria. Check the maker's page for your country before you plan around one.
Is it safe to connect an agent to company email and files?
It can be, with care. An agent reaches further into your accounts than a chat tool, and it can form memories from what it reads. Use a business plan, connect only the apps the job needs, start read-only, and keep a person approving anything that leaves the building.
If you want help
We set these agents up for clients. We recommend Claude whenever it fits, and we support firms on Microsoft, ChatGPT and Google. We start by mapping the job the agent will do, connect only what it needs, set the approval rules with you and your IT provider, and coach the person it works for until they trust its output. Agents can be set up on their own or as part of Fractional AI leadership.
Start with the free AI Opportunity Map, about 3 minutes, or book a free 30-minute call to talk through which agent fits.
Sources
Checked November 14, 2026: Anthropic, Claude Duet (published October 21, 2026); Anthropic, Claude Managed Agents overview; OpenAI, Getting started with your dot; xAI, Grok Bot overview (updated November 7, 2026); xAI, Grok Bot plans (updated November 11, 2026); xAI, Grok Bot for Enterprise (September 3, 2026); Meta, Muse for Small Business (September 29, 2026, updated October 8, 2026); Microsoft admin notice MC1479277, Autopilot (updated November 9, 2026); Google, Gemini at Work 2026 (October 8, 2026) and 9to5Google on its preview status; Nous Research, Hermes Agent. Market share: Menlo Ventures, 2026: The State of Enterprise AI (November 3, 2026; Menlo invests in Anthropic). Copilot billing: Computing (November 2026). Our views on ease of use, staying on task and value come from our own client work.


