Claude Duet for Small Business: What It Handles and What It

By Lazar Milicevic · Published October 22, 2026 · 10 min read

A customer email sits unread for six hours. An invoice goes 30 days overdue because nobody looked at the QuickBooks aging report. A meeting lands on the calendar at the wrong time. None of these were on your to-do list, and each one costs you something. Anthropic's new Claude Duet is built to catch exactly this class of problem, and it's worth understanding precisely what it does, what it doesn't, and where a custom setup still wins.

What Claude Duet actually is (and isn't)

Claude Duet is a proactive agent from Anthropic that monitors your connected apps and acts on what it finds, with your approval for anything consequential. Anthropic introduced it on October 21, 2026 as a research preview, and VentureBeat's coverage positions it as Anthropic's answer to the "Muse" class of personal agents.

The facts that matter if you're deciding whether to care:

• Availability: Research preview, waitlist only, limited to US Pro and Max subscribers. Per Anthropic's help article, it is not yet on Team or Enterprise plans. Anthropic says those are coming later, with no published date. Regions outside the US have no announced timeline either.
• Model: It runs on Claude Opus 5.5. You can't pick a model; Duet decides how much effort a job needs.
• Surfaces: Sources mention iOS, Android and claude.ai on the web. The headline says "mobile and desktop," but I found no source confirming a dedicated desktop app, so don't plan around one.
• Where it runs: In a virtual machine with its own browser, so it keeps working after you close your laptop or turn off your phone. During the preview you can't watch it work in that browser or take it over.

It is not a real-time watcher, and it is not a fully autonomous employee. Those two points shape everything below.

The proactive loop: morning brief, interval checks, alerts

Duet's core behavior is a three-part loop: a daily brief, periodic checks of your connected accounts, and alerts when something needs attention.

According to MacRumors, Duet sends a short brief each morning on the important items across your connected apps, then re-checks those apps every few hours and alerts you when something needs a decision. VentureBeat reports the same "every few hours" cadence and adds that Anthropic says tighter monitoring would burn through usage allowances faster.

That cadence has a practical consequence. If a customer emails at 10:05 and the next check is at noon, Duet won't tell you at 10:06. The pitch of "spots problems before you think to ask" is accurate for slow-burn problems: the overdue invoice, the unanswered thread from yesterday, the calendar conflict next week. It's not accurate for anything that needs a response in minutes.

A rough way to sort your own business problems:

If your pain is mostly the first three rows, a periodic agent is a reasonable fit. If it's the fourth row, you want event-driven automation instead.

What it can do with your invoices and customer email

Anthropic says Duet can investigate unpaid invoices, draft customer follow-ups, build financial forecasts and create web applications. VentureBeat's examples include checking QuickBooks or Stripe records for outstanding invoices and preparing reminders, and using HubSpot customer data to draft outreach.

Read that carefully: these are Anthropic's stated capabilities, relayed by the press. There are no published benchmarks, no task-level cost estimates and no measured productivity gains for Duet. Anyone quoting you "saves 10 hours a week" for this product is making it up.

What you can reasonably expect for the overdue-invoice case:

1. Duet pulls open invoices from your connected accounting or payments tool.
2. It identifies which are past due and drafts a reminder for each.
3. It shows you the intended email and waits.
4. You approve, and it sends from your connected account.

Step 3 is not optional behavior. It's a design constraint, covered next.

For a sense of the workflow Duet is approximating, here is the deterministic version of "find overdue Stripe invoices" that engineers have written for years:

import stripe
from datetime import datetime, timezone

stripe.api_key = "sk_live_..."  # use a restricted, read-only key

def overdue_invoices():
    now = int(datetime.now(timezone.utc).timestamp())
    invoices = stripe.Invoice.list(status="open", limit=100)
    return [
        {
            "id": inv.id,
            "customer_email": inv.customer_email,
            "amount_due": inv.amount_due / 100,
            "days_overdue": (now - inv.due_date) // 86400,
        }
        for inv in invoices.auto_paging_iter()
        if inv.due_date and inv.due_date < now
    ]

That script is exact and cheap, but it can't read the email thread where the customer said "paying Friday." That judgment layer, reading context before nudging, is what an LLM agent adds. And it's also where misreads happen.

The approval model and the hard limits

Duet shows you what it intends to do and asks for approval before it sends a message, makes a payment, signs a document or deletes anything. Anthropic's help article also lists a few actions as off-limits even with approval: trading securities, moving money between accounts, and accessing someone else's account.

Other boundaries worth knowing:

• No phone or SMS. Duet doesn't make calls or text other people. It can send email from connected accounts, with your approval.
• Purchases go through Stripe's Link wallet or Shopify's Shop Pay. In the Link flow, payment uses a virtual card that keeps your real card number private.
• Sensitive data: Anthropic describes a classifier that blocks sharing of payment-card numbers and government-issued IDs. Health and personal financial information is shared only after an explicit instruction. Duet never directly sees your passwords or sign-in tokens.

This is a sensible permission design, and it's the same pattern I'd insist on in any agent I ship: read broadly, write narrowly, and put a human gate in front of anything irreversible. The tradeoff is that approval gates mean you're still in the loop. Duet reduces the finding and drafting work; it doesn't remove the deciding work. For a solo owner, that's often exactly the right split.

Anthropic is also candid about quality. Per CNBC, the company says the early version "might misread a thread or miscalibrate what's urgent," and that personal errands like booking tickets are weaker than business tasks. Treat that as the spec, not fine print. Duet is designed for small businesses and professionals, and the approval step is your protection against the misreads.

Connectors, gaps, and the Microsoft 365 problem

Anthropic says Duet can draw on more than 6,000 connectors and plugins in the Claude Marketplace. VentureBeat's examples include Gusto payroll, Docusign agreements, Canva campaigns and Shopify store maintenance.

A big connector count is a marketing number until you check your own stack. Before you join the waitlist, list the five tools your business actually runs on and check each one:

• Microsoft 365 is not yet available in Duet, per Anthropic's help center, regardless of the app logos shown on the product page. If your email and calendar live in Outlook, the core "missed email" and "wrong meeting time" scenarios don't apply to you yet.
• Slack and some other connectors work with limits. Check what "limits" means for your use before assuming parity with the native app.
• Announced-but-not-shipped items such as 1Password credential handling and improved Shopify design tooling are described as forthcoming. Don't build a plan on them.
• Payroll and e-signature (Gusto, Docusign) are announced scenarios. Anthropic says Duet can prepare and run Gusto payroll with your approval, but that's a vendor-described flow, not something independently tested. If payroll is on your list, test with a dry run and read every approval screen.

A quick compatibility check you can run today, regardless of Duet:

# duet-fit-check.yaml - fill this in before you join the waitlist
email_provider: gmail        # outlook = not supported yet
calendar: google             # check connector status
accounting: quickbooks       # named in Anthropic's examples
payments: stripe             # named in Anthropic's examples
crm: hubspot                 # named in Anthropic's examples
chat: slack                  # works with limits
plan: pro                    # team/enterprise not yet available
region: US                   # preview is US-only

If more than two lines come back as "unsupported" or "unknown," Duet isn't your tool this quarter.

Cost: what's known and what isn't

Duet is included in Pro and Max at no extra cost during the research preview, but it consumes your plan's allowances. Per the help article, it uses the plan's five-hour session and weekly limits, and background work like the morning brief and check-ins counts toward them.

That last clause is the one to internalize. A proactive agent spends tokens when you're not looking. If you also use Claude interactively for real work, Duet's background activity competes with you for the same limits. Admitted users get one full reset of their five-hour and weekly limits, available until November 30, 2026, which gives you a free window to measure how much headroom background monitoring eats.

On pricing, VentureBeat reports Anthropic lists Pro at $20 a month or $200 for an annual subscription, and Max starting at $100 a month. Pricing pages change, so check Anthropic's current pricing page before you decide. The sources I found say nothing about what Duet costs after the preview ends or whether it stays in the Pro plan, so check Anthropic's help center and pricing page for updates. Anthropic also gives no per-task cost estimates, so you can't calculate a cost per invoice chased. You have to measure it yourself during the preview.

A simple evaluation plan for the preview window:

1. Connect two or three accounts only (say Gmail, Stripe, HubSpot).
2. Note your usage meter before and after a normal week.
3. Count how many alerts were useful, wrong, or late.
4. Count how many drafts you sent unchanged versus edited.

That gives you your own numbers, which are worth more than any vendor claim.

Where Duet fits and where custom agents still win

Duet is a generalist assistant with a fixed interval and a fixed approval model. That's a strength for owners who want something running in an afternoon, and a limit for anyone with specific rules.

Duet is a good fit when:

• You're a solo owner or tiny team on Gmail, Google Calendar, Stripe, QuickBooks, HubSpot.
• Your pain is noticing things: the thread you forgot, the invoice nobody chased.
• A check every few hours is fast enough, and you're happy to approve each action.
• You're on a US Pro or Max plan and can wait for the waitlist.

A custom or hybrid setup wins when:

• You need real-time reactions. Event-driven triggers (a webhook on invoice.payment_failed) fire in seconds, not hours.
• You have business rules. "Remind at day 7, escalate at day 21, never contact enterprise accounts without account-manager sign-off" is policy, and policy belongs in code or a workflow, not in an LLM's judgment.
• You're on Outlook/Microsoft 365, or on a plan or region Duet doesn't cover yet.
• You need an audit trail. Logging every decision, input and output in your own system is easier when you own the pipeline.
• Volume is high. Deterministic steps are cheaper and more predictable than agent tokens for repetitive work.

Most real setups end up hybrid: deterministic code or workflow tools for the rules and the real-time triggers, and an LLM only for the fuzzy parts, such as reading a thread, classifying intent and drafting a reply. Duet is a credible option for the "owner's inbox" layer. It's not a replacement for the rule-bound plumbing underneath your billing and ops.

Where this leaves you

I build AI automations for small teams, and the pattern Anthropic describes for Duet is the one I'd use in any agent: connect the accounting, payments, CRM and inbox tools, let the model do the reading and drafting, and keep deterministic logic and human approval around anything that sends, pays or deletes. The difference with a custom build is that you control the schedule, the rules, the triggers and the logging.

The practical first step is triage: which of your problems are fine for an off-the-shelf agent like Duet once it's available to you, which need an event-driven workflow, and which genuinely call for a custom agent. Run the fit-check file above against your own stack and you'll know quickly where you land.

Work with BizFlowAI

If you'd rather have this built for you, that's what we do: production AI automation for solo founders and small teams — agents, integrations, and document pipelines that actually ship.

Request a free discovery call — 30 minutes, we map the highest-ROI automation in your workflow. No pitch deck, just engineering.

More guides like this on the BizFlowAI blog.