Bunny Honey ClubBunny Honey/blog
Work with us
← back to indexblog / ai / claude-duet-for-small-business
● AI

Claude Duet for Small Business: What to Hand It

Claude Duet checks your apps every few hours and asks before it acts. Here is what an owner should hand it, and what should run on a trigger instead.

AH
Arthur HofFounder, Bunny Honey Club AI
publishedOct 29, 2026
read7 min
Claude Duet for Small Business: What to Hand It

Anthropic just built an assistant that messages you first. Every morning it sends a brief, and during the day it pings you when something needs a human. That is Claude Duet for small business, and the pitch is good. The detail that decides

Anthropic just built an assistant that messages you first. Every morning it sends a brief, and during the day it pings you when something needs a human.

That is Claude Duet for small business, and the pitch is good. The detail that decides whether it fits your company is one sentence in the launch post about how often it looks.

Duet is a chief of staff that works from your logins

Anthropic opened the Claude Duet research preview on October 21, 2026. It is a standalone app on iOS, Android, and claude.ai, built on Claude Opus 5.5, aimed at "professionals and small business owners with more on their plates than they can track."

It carries over your Claude memory, skills, and connected apps. It has its own virtual machine with a browser, so it works while your laptop is closed.

Anthropic's examples are ordinary office work. Check QuickBooks or Stripe for unpaid invoices and draft the reminders. Prepare payroll in Gusto. Draft HubSpot follow-ups from deal history. Send agreements through Docusign.

The testimonials are small and believable. A broker's Duet noticed a showing was on her calendar for the wrong day. A brand strategist's Duet saw a client call landed right before her flight and offered to move it.

That is the product: a second pair of eyes on your week. This is a read of Anthropic's published material, not a hands-on test. We are not going to pretend otherwise.

Oct 212026 research preview launch, waitlist, US Pro and Max subscribers first
6,000+connectors and plugins in the Claude Marketplace that Duet can use, per Anthropic
Every few hourshow often Duet checks your connected accounts, per Anthropic
0extra cost in the preview: included in Pro and Max, sharing their usage limits

The every-few-hours check decides what you can hand it

Here is the sentence. Anthropic says Duet "checks your connected accounts every few hours, because checking nonstop would use up your plan's limits much faster."

Honest engineering. It also sets the shape of the product.

Duet is good at things that stay true for an afternoon. A forgotten invoice. A double-booked Thursday. A contract waiting on a signature. A brief that tells you what to do first.

Duet is not built for things that go stale in minutes. A web form lead who is shopping three vendors. A missed call from someone with a burst pipe. A customer whose payment just failed at checkout.

On inbound leads, minutes matter. The test for any answering setup is whether someone replies within minutes, and a polling assistant that looks every few hours fails that test.

Two more limits from the launch post, both worth reading twice:

  • Duet's background checks count against your plan's usage limits, and bigger jobs like a financial model or a deck take more model turns.
  • Duet lives in its own app, separate from the main Claude app, and Anthropic says it still might "misread a thread or miscalibrate what's urgent."

That last one is the sentence to build your trust around. Urgency is exactly the judgment you are renting.

Approvals are the product, and Anthropic says so

Duet reads, drafts, and prepares on its own. Before it sends a message, makes a payment, signs a document, or deletes anything, it shows you what it is about to do and waits.

Anthropic's safety post adds the detail that matters. The approval card shows the action, the destination, the content, and every recipient. A separate automated check, built on the auto mode classifier from Claude Code, runs outside the model doing the work, so instructions hidden in a web page cannot talk it out of the decision.

Some things are off-limits even with your approval: trading securities, moving money between accounts, accessing someone else's account. Paying or sending money, deleting an order, and creating an account never get a "don't ask me again" shortcut. Routines that run while you are away have the same safeguards.

That is a sensible design. Now read the part Anthropic wrote about its own design.

Approval fatigue is a serious issue, because too many requests teach people to tap through them, sometimes without adequate consideration.

— Anthropic, Claude Duet Safety, Privacy, and Security, October 23, 2026

A vendor admitting its own safeguard has a human failure mode is rare. Take it seriously.

Run the numbers on your own week. If Duet drafts ten invoice reminders, three follow-ups, and a payroll run, that is a queue. On a good Monday you read each card. On a Friday at 5:40 you tap through.

We see the same pattern in our own agent fleet. In our OpenClaw Factory report, 33 production agents showed a 4.1% reversal rate over six months: how often a human had to undo something an agent did. Our rule is to shut things off above 10%. Even at 4.1%, someone is reading the queue, and that is the part that does not scale to a busy owner.

Your inbox is the data you are handing over

Duet reads your email, Slack, calendar, accounting, and CRM. For a service business that means customer names, addresses, health details, quotes, and card disputes, all in one place.

Anthropic describes real controls. Duet never sees your passwords, cookies, or sign-in tokens; they pass through separate machines the model cannot access. A new classifier checks every outgoing message. It is designed to refuse the strictest tier of information, including full card and bank numbers and government ID numbers, even if you ask. Health and finance details go out only if you explicitly instruct it.

Three things to know before you connect a business inbox:

  1. Memory sticks. Disconnecting an app stops new access, but what Duet already learned stays until you reset its memory in settings. If a team member leaves and you cut their app access, remember to clear what the assistant picked up.
  2. It shares your Claude memory. Details from earlier Claude chats are visible to Duet. Check what is in there before you connect a work account.
  3. It is a personal plan. The preview is for Pro and Max subscribers. A tool tied to one person's subscription is one person's context, not a company system with shared rules and an owner when that person leaves.

Anthropic itself says the safeguards "will occasionally miss something or flag a harmless message." Fine. Plan for the occasional miss.

If you are already weighing agents against workflows, our comparison of n8n and Claude agents is the longer version of this argument. The permissions angle on a rival product is in our Muse for Small Business breakdown, and OpenAI's always-on equivalent is covered in the ChatGPT dots rules post.

A sorting rule for what goes to Duet and what gets built

Three vendors now sell a proactive agent to the owner of a small company. Meta has Muse, OpenAI has dots, Anthropic has Duet. The pitch is the same each time: tell it what you want, it gets on with it.

Our sorting rule works for all three.

Hand it to the agent when the job needs judgment, happens a few times a week, and one person owns the result. Prioritizing your inbox. Prepping for a client call. Pulling a weekly sales number out of four tools. Drafting the reminder for the invoice that is 30 days late.

Build it as a workflow when the job is the same every time, the clock matters, and more than one person depends on it. New lead arrives, enrich it, text it, book it. Payment fails, notify the customer, flag the account. Review requested, send it, route the one-star answer to a human.

Keep it human when the action is irreversible and the stakes are personal. Refunds above a number. A termination email. Anything with legal wording.

The honest summary: Duet is a chief of staff, not a front desk. A chief of staff makes your Monday better. A front desk answers at 7:12 on a Tuesday evening when the water heater floods.

We build the second kind. Event triggers, defined steps, a human gate where money or customers are involved, and a log you can read. If you want that around your own stack, our automation service is where it starts. Duet on the owner's side and a workflow on the company's side is a combination we would happily build around.

Where we would start if we were on the waitlist

Not with the payroll.

Connect Gmail and your calendar, nothing else. Let the morning brief run for two weeks and count how often the "urgent" call was wrong. If it picks the right three things most days, add QuickBooks or Stripe read-only and let it draft reminders you edit and send yourself.

Leave payments and payroll for last. Anthropic's own materials say the wallet pays with a one-time card limited to the merchant and amount, which is a good guardrail. It still takes a person to read the amount.

Keep a short list of what Duet may never touch: customer health information, anything in a dispute, anything with a lawyer attached. Write it down on day one, because by week three nobody remembers deciding.

And keep your response times off the list entirely. The thing that wins you a lead should not depend on a check that happens every few hours.

— share
— keep reading

Three more from the log.