Claude Duet for small business: should owners use it?

Anthropic opened the waitlist for Claude Duet on October 21, and it built the product for the owners I write for. The announcement says Claude Duet is for “professionals and small business owners with more on their plates than they can track.” The product page demos a morning brief for an 85-person café company.

One job on Anthropic’s list is checking QuickBooks or Stripe for unpaid invoices and drafting the reminders. An agent of mine has run a version of that job on a Monday-morning schedule since July: it matches payments to open invoices and drafts the reminders I send. I run hundreds of AI agents across my own companies and client companies, so I read the launch with one question: which of an owner’s jobs belong in Claude Duet, and which don’t?

My answer: Claude Duet looks like a good assistant for one person, the owner. It runs on one person’s Claude account, on Pro and Max plans only, and on one vendor’s model. A company’s recurring jobs need an agent with its own login, a record the company owns, and the freedom to change models. For now, it offers none of those.

What is Claude Duet?

Claude Duet is Anthropic’s proactive AI agent, in its own app. It sends you a morning brief from your connected apps, takes tasks by text or voice, keeps working on its own cloud computer while you’re away, and asks before it sends, pays, signs or deletes anything. It’s a waitlisted research preview for US Claude Pro and Max subscribers.

It runs on Claude Opus 5.5 and has its own virtual machine and browser, so your laptop can be closed while it works. It carries over the memory, skills and connected apps from your Claude account. Anthropic says it can use more than 6,000 connectors and plugins, and its examples are the owner’s stack: Gmail, Slack, QuickBooks, Stripe, Gusto, HubSpot, Docusign, Shopify and Canva.¹

The approval rule is written plainly: before Claude Duet “sends a message, makes a payment, signs a document, or deletes anything, it shows you exactly what it’s about to do and waits for your approval.” A few things are off-limits even with your approval, “like trading securities or moving money between accounts.”

Anthropic gave early access to some small business owners. Amin Ebrahimi owns Shahs of Kabob in South Florida, with about 110 employees, and used it for accounting, customer support and marketing. “It brings it all together,” he told Anthropic. “I don’t need to open up 100 tabs.”¹

Anthropic is candid about the limits, too. Its FAQ says Claude Duet “may misread a thread or miscalibrate what’s urgent, and it can miss something that arrives between checking in with you.”¹

How much does Claude Duet cost?

Claude Duet has no separate pricing: it’s included in Claude’s Pro and Max plans during the research preview. Pro is $20 a month, or $17 a month billed annually; Max starts at $100 a month.² The real limit is usage: it shares your plan’s allowance with the rest of Claude, and its background checks count against it.

That’s why it checks your accounts only “every few hours, because checking nonstop would use up your plan’s limits much faster,” in Anthropic’s words. Larger tasks, like a financial model or a deck, “require more model turns and background work.”

I’d plan around the limit before the price. Anthropic doesn’t say what Claude Duet does when a shared limit runs out, so ask before you rely on it for anything with a deadline.

On price alone, Claude Duet wins. My AR (accounts receivable) agent costs less than $200 a month to run. For one owner’s inbox and calendar, $20 a month is the better deal.

Can Claude Duet chase unpaid invoices in QuickBooks and run payroll?

Anthropic says yes: Claude Duet can check QuickBooks or Stripe for unpaid invoices and draft reminders, and prepare and run payroll in Gusto. It drafts, and you approve each send. My own collections agent works the same way: it drafts each reminder, and I send it. Which reminders can safely go out without you is in should you let QuickBooks AI send your invoice reminders?

The difference is everything around the job. Here’s Claude Duet as I read Anthropic’s description, next to my AR agent’s record:

	Claude Duet	My AR agent
Whose account it uses	Yours: your Claude account and the apps you connect	Its own key to the bank, read-only except for one note per payment
When it works	Every few hours, and when you message it	Every Monday morning, on a schedule
What waits for a person	Sending, paying, signing, deleting	Every reminder, plus four exceptions it never guesses at: an unknown payer, a payment larger than the open balance, a payment with no open invoice, anything that looks like a refund
The record	Its activity in your app	A dated weekly digest, and a note on each matched payment that it rereads before counting the payment as matched
When the person who set it up leaves	It goes with their account	The job stays with the company

For the owner of a five-person shop who does the collections personally, the left column may be enough. Once collections belong to a role, someone besides the owner needs to see what ran, what was held and why.

None of the agents at Fifth Nine (fifthnine.com), my company, run payroll, so I have no record to offer there. Claude Duet asks before it runs payroll, and that click should come after someone has checked the hours.

Is Claude Duet safe for company data?

Claude Duet’s safety design is serious. It never directly sees your passwords or sign-in tokens, a classifier approves its actions, a new one guards against leaking sensitive information, and red teams and an outside firm tested it.¹ The catch for a company: it runs only on personal Pro and Max plans, and “Team and Enterprise plans are not included yet.”

A personal plan changes three things.

Training. On Free, Pro and Max, Anthropic’s privacy center says it uses your chats to improve its models if you choose to allow it. If you allow it, Anthropic keeps that data for five years; if you don’t, the standard is 30 days.³ Anthropic’s pricing page lists Team plans with “No model training on your content by default.”²

Admin. Team plans come with central administration and “Admin controls for remote and local connectors.” Enterprise adds audit logs. A personal plan has neither, so the company can’t see which of its systems an employee’s Claude Duet is connected to.

Responsibility. Anthropic’s consumer terms say: “You are responsible for all Inputs you submit to our Services and all Actions.” When the account is personal, so is that responsibility, even when the inbox belongs to the company.

For the owner’s own accounts, that’s a choice the owner can make.

Should employees use personal AI accounts for work?

Not with company systems. A personal Pro or Max plan leaves the training choice to the employee, gives the company no admin view of what’s connected, and puts the responsibility on the individual. My rule for staff: nobody connects the company’s QuickBooks, payroll or shared inbox to a personal AI plan, Claude Duet included.

Every agent at Fifth Nine that needs a key holds its own. I wrote up why in how to give AI agents access safely.

Claude Duet vs ChatGPT dots: which should a business pick?

Claude Duet and ChatGPT’s dots are the same idea: an always-on agent with its own cloud computer, working across apps you connect and bringing decisions back to you. The practical difference for a business today is the plan. Dots are available on ChatGPT Business Premium as well as Pro. Claude Duet is on Pro and Max only.

OpenAI’s dot help page says dots run on GPT-6 Astra and are “available to Business Premium users in supported ChatGPT regions, subject to workspace eligibility requirements.” A dot lets you set custom rules for some actions, choosing one of four behaviors, from “Take action without asking” to “Hand off to you.”⁴ Anthropic describes no custom rules for Claude Duet: it asks before every send, payment, signature or deletion.

If your company is on ChatGPT Business Premium, a dot can live inside the company workspace, subject to OpenAI’s eligibility rules. If you already use Claude, Claude Duet starts with your memory and connectors on day one. I compared ChatGPT’s agents with agents that hold a job in AI agent vs ChatGPT.

I haven’t used Claude Duet or a dot myself yet. Claude Duet is waitlist-only, and I’d rather judge either one on a real week of work than on a launch demo.

Does Claude Duet lock you in?

Yes. Claude Duet runs on Anthropic’s Opus 5.5, a dot runs on OpenAI’s GPT-6 Astra, and neither page mentions a choice of another vendor’s model. What the assistant learns about your business (its memory, connections and history) lives in that vendor’s account.

Meta’s Muse runs on Meta’s own Muse Spark model and has had a small business version since September: the same bet, placed on Meta.

Lock-in matters more when the vendor can turn into your competitor. Anthropic sells its models to software companies and also builds products of its own. On April 17 it launched Claude Design, for “designs, prototypes, slides, one-pagers, and more,” the work many people open Figma for. Three days earlier, Anthropic’s Mike Krieger resigned from Figma’s board, according to Figma’s SEC filing, which gives no reason. Claude Code, Anthropic’s coding agent, competes with Cursor for developers.

Access can also be cut on short notice. In June 2025, while OpenAI was reportedly buying Windsurf, Windsurf’s CEO, Varun Mohan, wrote that “with less than five days of notice, Anthropic decided to cut off nearly all of our first-party capacity to all Claude 3.x models.”

None of that means Anthropic is coming for a 50-person company’s line of business. It does mean that an owner who connects the books, the inbox and payroll to one vendor’s assistant hands that vendor a detailed picture of the business, and the vendor decides what it builds next.

I run Fifth Nine’s agents the other way around. The system they run on can put any agent on Anthropic’s Claude, on OpenAI’s Codex, or on open-source agent software like OpenCode, which supports 75 or more model providers and local models.⁵ Most of the agents that run Fifth Nine today are on Anthropic’s models. Some run on OpenAI’s, including the bookkeeper and the agent that writes this website’s code.

Sometimes Claude doesn’t do a good job on a task, and I want that agent on a Codex model instead. My system lets me move it. The job, its rules, its tickets and its history stay where they were. The agent setups Fifth Nine builds for clients work the same way, so the company keeps the job, whichever model does it.

Which jobs belong in a personal AI agent, and which need their own?

A job belongs in a personal agent like Claude Duet when it’s yours: your inbox, your calendar, prep for your meetings, a dashboard you’ll read yourself. A job needs its own agent when it belongs to a role: it has to run when you’re away, other people rely on its record, or it touches the company’s money on a schedule.

The job	Personal agent (Claude Duet, a dot)	An agent with its own login
Morning brief of your inbox and calendar	Yes	Overkill
Drafting replies you’ll send yourself	Yes	Overkill
Prep for a client call	Yes	Overkill
A one-off forecast or dashboard	Yes, after you check the numbers	Only if it refreshes for the team
Matching each week’s payments to invoices, and the overdue-invoice report	No	Yes
Collection reminders across every client	Only for the owner’s own books	Yes, with written rules for exceptions
Anything a second person has to audit	No	Yes

Claude Duet is built for one person: your account, your memory, one thread of history. There’s no team version yet, so its work lives in one person’s account.

The promise Claude Duet makes, that it “goes quiet until something needs you,” is the one I built my fleet around. The agents that run Fifth Nine don’t send me a morning brief. Once a week, on Friday afternoon, my Chief of Staff agent posts a digest listing every open decision that waits on me, oldest first, with its deadline and what happens by default. Between digests, an agent reaches me only on the ticket that needs my decision.

That works for a company because the record sits in the company’s own system, where anyone who takes over the job can read it.

Is Claude Duet worth it for a small business?

For an owner whose day lives in the inbox, yes: join the waitlist. Start it on the morning brief and your own drafts, set the model-training choice first, and keep every approval. It isn’t worth it yet for the company’s recurring jobs: keep those out, keep company systems off staff’s personal plans, and look again when it reaches Team.

Whatever you hand it, keep the company’s jobs somewhere you can change the model without starting over.

Before you connect anything, write down the three jobs you’d hand it. Next to each, write whose job it is. Where the answer is you, Claude Duet is a reasonable bet at $20 a month. Where the answer is a role, that job needs its own login, its own rules and a record that outlives whoever set it up. How to set that up is in how to hire your first agent.

If the list has a role’s job on it and you want that job mapped before anyone builds anything, that’s what Fifth Nine’s AI Audit does, before any AI automation gets built.

---

Sources

How this was written: drafted with AI from my notes and Fifth Nine’s operating records, then edited and approved by me.

Footnotes

1. Anthropic, “Introducing Claude Duet”, October 21, 2026, and the Claude Duet product page and FAQ, both read October 22, 2026. Anthropic says purchases through Stripe’s Link wallet use a virtual card, and that signing in to sites with logins saved in 1Password is “coming soon.” Testimonials in the announcement are Anthropic’s, from early testers. ↩ ↩² ↩³ ↩⁴

2. Anthropic, Claude pricing, read October 22, 2026. Prices are before tax. Team is “for teams of 2 to 150” at $25 a seat per month billed monthly, or $20 billed annually. ↩ ↩²

3. Anthropic, “Updates to Consumer Terms and Privacy Policy”, August 28, 2025. The update applies to Free, Pro and Max, and not to services under Anthropic’s Commercial Terms, which include Team and Enterprise. ↩

4. OpenAI, “Getting started with your dot”, read October 22, 2026. The page says dots for Pro users are “rolling out” to adults “in markets excluding the European Economic Area, Switzerland, and the UK,” and that Enterprise workspaces can try a beta that is “initially turned off by default.” It also warns: “Your dot can make mistakes, including when following your rules.” ↩

5. OpenCode, “Providers”, read October 22, 2026: “OpenCode uses the AI SDK and Models.dev to support 75+ LLM providers and it supports running local models.” Which models Fifth Nine’s agents run on is from its own agent roster, read October 23, 2026. ↩