Until now, using AI at work meant opening a chat and asking for something. Within one month, the three largest AI companies launched the same different idea: an assistant that does not wait to be asked. OpenAI announced dots on September 29, Anthropic opened a preview of Claude Duet on October 21, and Google has been rolling out Gemini Spark since the spring.
For a small business this changes the question. It is no longer what should I ask the AI, but what should I let it do while I am busy with something else. This guide covers what these agents have in common, where they help, where they cannot, and what to settle before you switch one on. Details below are as of late October 2026, and they are moving quickly.
What these agents have in common
The three products share one design. Each agent has its own computer in the cloud with a browser, so it keeps working when your laptop is closed. It connects to the apps you already use, email, calendar, accounting, CRM, chat, and checks them in the background. It reports back with a brief or a message when something needs you, and it asks before anything that sends, pays, signs or deletes.
- ChatGPT dots (OpenAI): reachable in ChatGPT, Slack and Teams, connected to over 4,000 apps through plugins. Background research uses read-only tools, and Custom Rules let you allow, require approval for, or block specific actions. Rolling out to Pro and Business Premium plans in eligible countries, with an Enterprise beta an admin must enable.
- Claude Duet (Anthropic): a separate app on iOS, Android and the web, built for professionals and small business owners. It sends a morning brief, checks connected accounts every few hours, and waits for approval before sending, paying, signing or deleting. A research preview for US Pro and Max subscribers, by waitlist.
- Gemini Spark (Google): a 24/7 agent running in Google Cloud and working across Gmail, Calendar, Drive and Docs. According to press coverage, it is rolling out to AI Ultra subscribers and to AI Pro in the US, and is not yet available in the European Economic Area, the UK or Switzerland.
Where they help a small business
The best early uses are the small, recurring jobs that nobody owns and everybody forgets. They have clear inputs, a person can check the result in a minute, and the cost of a miss is a delay rather than a disaster.
- A morning brief across email, calendar and accounting, so the day starts with what actually needs you.
- Unpaid invoices: checking the accounting tool and drafting reminders for you to approve.
- Customer follow-ups drafted from the history in your CRM, instead of from memory.
- Catching mistakes nobody asked about, such as a meeting booked on the wrong day or a call right before a flight.
- Keeping a document current when something changes, such as a proposal after the scope moves.
What they cannot know on their own
- Your unwritten rules. Which client never gets chased in December, which discount needs your sign-off, how you speak to a supplier you have known for ten years. If it lives only in your head, the agent will guess.
- Systems without a connector. An agent reaches your tools through plugins and connectors. A custom application, an old internal database or a store with no integration is invisible to it.
- When to stop. These are early products. Both OpenAI and Anthropic say plainly that their agents still make mistakes and that consequential work needs review.
- The real cost. Background work consumes your plan's usage. With Claude Duet, background checks count toward your limits; OpenAI includes a first dot with eligible plans and plans to charge for more capacity later.
Before you switch one on
- Start read-only. Connect the apps, let it watch and draft for two weeks, and keep every send, payment and change behind your approval.
- Write your rules down. A short page on how the business works, who decides what and what must never happen is what turns a generic agent into a useful one. Our guide on setting up AI with a folder of plain files shows how.
- Give each agent one owner. Someone has to read the brief, approve the drafts and correct the agent when it is wrong.
- Check the plan's data terms. OpenAI says it does not use content from Business, Enterprise or Edu workspaces to train its models by default; personal plans have their own settings.
- Check availability where you work. Several of these launches exclude some countries or start in the US only.
Which one to choose
Choose the one that lives where your work already is. A business on Google Workspace will get more from Spark; a team already working in ChatGPT, Slack or Teams from dots; a team already using Claude from Duet, which carries over its memory and connectors. Changing your whole office suite to get a particular agent is rarely worth it.
Then judge it on one recurring task for two weeks. Measure the time the task took before and after, the same way you would audit any manual step, and keep the agent only if the number moves.
Where custom work still matters
An agent is only as useful as what it can reach and the rules it is given. For most small businesses, the important data sits partly in mainstream tools and partly somewhere an agent cannot see: a custom application, a spreadsheet that became a system, an online store with no connector.
OpenAI is reaching the same conclusion for larger companies. Its specialist dots, with their own identity and integrations into a company's systems of record, start as pilots where its engineers define each dot's responsibilities, tools and approvals with the client. That definition work is what we do for smaller businesses: writing down the rules, connecting the systems an agent cannot reach, and putting approvals in the right places, so whichever agent you choose has something solid to work from.