Executive summary: Most "AI assistant" advice for small businesses is the same advice as last year: buy a chatbot, type at it, hope it helps. On October 21, 2026, Anthropic shipped something different: Claude Duet, a standalone AI agent app built specifically "for professionals and small business owners with more on their plates than they can track." It is proactive, not reactive — it sends you a morning brief on what needs attention, keeps checking your connected apps in the background, and can start work on a task while you are away, running on its own virtual computer. It runs on Claude Opus 5.5, comes with more than 6,000 marketplace connectors (including QuickBooks, Stripe, Gmail, Slack, and HubSpot), and during its research preview it is included at no extra cost in a Claude Pro plan ($20/month, or $17/month billed annually) for US subscribers on a waitlist. The real question is not "is a $20/month agent worth it?" but "can I turn it on, keep a human on anything that sends or spends, and measure whether it pays for itself in the first two weeks?" This playbook gives four steps to do exactly that.
Duet is a research preview with a waitlist, so access is by invitation. It is available only to US subscribers on a Claude Pro or Max plan during the preview, and it is not on Team or Enterprise plans yet. The entry point is the cheapest way in: a Claude Pro plan at $20/month (or $17/month when you pay the $200 annual bill up front) — Duet is included at no extra charge during the preview. Max plans start at $100/month (5x the usage) and $200/month (20x); you do not need Max to start, and you should not buy it to test the concept. Join the waitlist from your Claude account settings; invites go out in waves, and one Anthropic lead said US Max invitations go out before US Pro, so a Pro subscriber may wait a little longer. When you are admitted, you get one full usage-limit reset (it clears both your 5-hour session limit and your weekly limit) that is available until November 30, 2026 — a useful cushion for a first week of heavy testing, but not a substitute for scoping what you actually connect. The single most important decision before you touch a connector is this: pick one department or workflow where your time leaks are most visible — accounts receivable, payroll prep, or outbound sales follow-ups — and make Duet your test pilot for only that. A narrow scope is what makes the measurement in Step 4 meaningful; "let's try AI on everything" is how a pilot quietly fails.
Tools & pricing: Claude Pro — $20/month, or $17/month billed annually ($200 up front); Claude Duet — included at no extra cost during the preview; Max — from $100/month if (and only if) heavy background usage justifies it. The limit-reset welcome is available until November 30, 2026.
The value of Duet is that it already knows your business: it carries over the memory, skills, and connected apps in your Claude account and can draw on the 6,000-plus connectors and plugins in the Claude Marketplace. The trap is connecting all of them. Start with just the two categories an owner is comfortable an assistant sitting near: money (QuickBooks or Stripe) and communication (Gmail or Slack). Because Duet connects through OAuth-style app connectors in the Claude Managed Agents design, it never directly sees your passwords or sign-in tokens — your credentials stay with you and the app. Then enable the small-business plugin and let it do what Anthropic demonstrated: scan QuickBooks or Stripe for unpaid invoices and draft the reminders, prepare payroll in Gusto with hours, bonuses, and reimbursements filled in, draft a job post for you to edit before it posts to LinkedIn, prep Docusign agreements, draft HubSpot follow-ups from your deal history, turn a launch into a Canva social campaign, and set up or manage a Shopify store. For purchases, it pays through Stripe Link (a virtual card, so your real card number is never shared) or Shopify Shop Pay — and only after you approve the spend. Resist the urge to wire in every tool at once. Every additional connector is more context, more surface area, and more of your usage budget to watch.
Tools & pricing: QuickBooks / Stripe, Gmail / Slack, Gusto, HubSpot, Docusign, Canva, Shopify — your existing subscriptions; Claude Marketplace connectors — included. Duet's background work and larger jobs draw on your plan's 5-hour and weekly usage limits at the same rate as chats (Opus 5.5 runs at $4 per million input tokens and $20 per million output tokens on the API card).
This is the step that turns a demo into something you would actually leave running. Duet ships with hard guardrails already: before it sends a message, makes a payment, signs a document, or deletes anything, it shows you the action, the destination, and the content, and waits for your approval. A separate safety classifier — running outside the model doing the work, so hidden instructions in a page it reads cannot change the outcome — screens every outbound message and is built to refuse sharing the strictest data, like credit-card numbers and government IDs, and to release health or personal-finance details only when you explicitly approve it. Some actions are off-limits even with your approval: trading securities, moving money between accounts, accessing someone else's account, and outbound phone calls or text messages. That built-in layer is the floor, not the ceiling. Add your own thin policy so the approvals are yours and consistent: which drafts need a review before they go out, who signs off on payment and document actions, and whether a payment or account creation can ever be pre-approved or always needs a fresh tap. Write it on one page. The point is that an agent that drafts is allowed, but committing — sending, paying, signing, deleting — stops at a person, and the person is defined by your policy, not by chance.
Tools & pricing: The Duet approval workflow and safety classifier — included in the plan. The policy sheet itself — no cost; it is a one-page internal rule about who approves what.
"We turned on an AI agent" is not a result; the result you want to show your own budget is hours. Pick one repeatable workflow from Step 1 — for example, "review unpaid invoices, draft the reminder emails, and schedule the follow-ups." For two weeks, log how long that workflow takes when you (or your staff) do it manually. Then let Duet run the same tasks and log the minutes it takes you to review and approve its drafts. Two honest levers control the cost while you measure: tell Duet to check less often, and watch fewer topics, because its background checks and bigger jobs (a financial model, a slide deck) consume more model turns. After the two weeks, convert the average weekly hours saved into a dollar value at your own hourly cost, and compare it against the $20/month Pro seat. If the workflow does not clear the seat cost, that is information, not failure — narrow the scope, pick a different workflow, or step back before scaling. The realistic upside is not "10x output"; it is a measurable chunk of hours a week back in one workflow, at a per-seat price you can verify and a human gate you control.
Tools & pricing: A spreadsheet or your existing time-tracking — free. The discipline is a two-week measurement before you scale, not a dashboard.
The shift is from "AI is a chat I type into" to "AI is a colleague that checks my accounts and tells me what needs my attention, then waits for me before it acts." Until recently, the credible always-on AI agent story for a small business was an enterprise price tag — OpenAI's comparable always-on agent (ChatGPT Dots) starts around $100/month, and Google added a small-business hub to Gemini on October 20, 2026, free with a Google account. Duet changes the price floor to $20/month on a Pro plan for US owners, with the strictest built-in approval and hard-limit design of the three. That does not remove the two things that still matter. First, the waitlist and the shared usage limits: invites go out in waves, and an always-on agent that checks your accounts every few hours draws on the same 5-hour and weekly limits as your chats, so scoping what it watches is a real decision, not a formality. Second, the guardrail: an agent that reads your email and drafts your replies is only as safe as the boundary you set before it runs. The honest ceiling is not a full replacement of your staff — it is a dependable front-line for the repetitive money-and-message work, with a human on every action that counts, at a price you can actually test this weekend.
Your next twenty minutes should be a scope decision, not a strategy debate. Join the Claude Duet waitlist from your Claude account settings, start a Pro plan, write down the one workflow you will measure, and sketch the one-page policy that decides who approves what. Then connect QuickBooks/Stripe and Gmail/Slack — and only those — set the approval gate, and start the two-week clock before the welcome limit reset expires on November 30, 2026. To get the full SMB AI agent strategy — including how Duet stacks against the other 2026 always-on agents and where the guardrails actually pay for themselves — start here:
Get the SMB AI agent strategy