By the end of this guide, you will know exactly what a proactive AI agent is, how it differs from the chatbot you already use, what it costs, where it goes wrong, and whether it makes sense for your business right now.
The timing is not an accident. On 21 October 2026, Anthropic opened a research preview of Claude Duet, an app it describes as "a new app where Claude works as a proactive agent." OpenAI's Dots and Meta's Muse for Small Business are chasing the same idea. Within a few months, "proactive agent" will be a phrase every software salesperson uses. It helps to understand it before someone tries to sell it to you.
What is a proactive AI agent?
A proactive AI agent is an AI assistant that starts the conversation instead of waiting for you. It watches the apps you connect, such as email, calendar and invoicing, decides what needs your attention, prepares the next step and asks your approval before acting. A normal chatbot answers questions. A proactive agent brings you the questions you forgot to ask.
Think of the difference between a calculator and a good office manager. The calculator is useful, but only when you pick it up. The office manager walks into your room at 9 AM and says, "Three invoices are overdue, the supplier changed Friday's delivery time, and you have two meetings booked at the same hour."
That second behaviour is what the new category is selling. The AI still uses the same kind of language model underneath. What changes is who moves first.
How does a proactive AI agent work?
A proactive agent runs on a loop: connect, check, decide, prepare, ask. It links to your accounts, checks them on a schedule, uses the AI model to judge what matters, drafts the reply or task, then waits for your yes before sending, paying or deleting. Claude Duet, for example, checks connected accounts every few hours rather than nonstop.
Step 1: Connect. You give the agent access to the tools your business runs on. Anthropic says Claude Duet can use the more than 6,000 connectors and plugins in its marketplace, including Gmail, Slack, QuickBooks, Stripe, HubSpot and Shopify.
Step 2: Check. The agent looks through those accounts at intervals. Anthropic's announcement explains why Duet does this every few hours: "checking nonstop would use up your plan's limits much faster."
Step 3: Decide. This is the hard part. The model has to judge which of 200 emails is urgent and which can wait. Anthropic says Duet was designed to "decide what needs you, what can wait, and when to check with you first."
Step 4: Prepare. The agent drafts a reply, a payment reminder or a summary. Duet sends a short brief each morning on the most important items across your connected apps.
Step 5: Ask. Before it sends a message, makes a payment, signs a document or deletes anything, Duet shows exactly what it is about to do and waits for approval. Some actions, such as trading securities or moving money between accounts, are blocked even if you approve.
How is a proactive agent different from ChatGPT or a chatbot?
A chatbot is reactive: nothing happens until you type. A proactive agent is persistent: it keeps working in the background, often on its own cloud computer, and messages you when something changes. The model can be identical. The difference is memory of your business, access to your apps, a schedule, and permission to act within limits you set.
Here is the practical comparison for a small business owner.
--- Who starts: With a chatbot, you do. With a proactive agent, the agent does, usually through a morning brief or a notification.
--- Where the work happens: A chatbot works inside one conversation. Claude Duet has a virtual machine with its own browser, so it keeps working when your laptop and phone are off.
--- What it knows: A chatbot knows what you paste in. A proactive agent reads your connected accounts and carries over saved memory and skills.
--- What it can do: A chatbot writes text for you to copy. A proactive agent drafts inside your tools and, with approval, sends or files it.
--- What it costs: A chatbot uses your plan when you chat. A proactive agent also uses it in the background. Anthropic warns that Duet's background checks count toward plan limits and larger tasks need more model turns.
If you want the deeper story of the chatbot-to-agent shift, our explainer on ChatGPT Dots covers OpenAI's version of the same idea.
What can a proactive AI agent do for a small business?
The best early uses are the jobs that slip because nobody remembers them: overdue invoices, unanswered enquiries, double-booked calendars and follow-ups that went quiet. Early testers named by Anthropic include a property broker whose agent spotted a showing booked on the wrong day, and a restaurant owner with about 110 staff who used it to work through invoices store by store.
Those examples translate directly to Hong Kong businesses.
A trading company. The agent checks your accounting software each morning, lists the five customers more than 30 days overdue and drafts polite reminders. You read them over coffee and approve four.
A tutoring centre. The agent notices that a parent asked about Saturday classes on Tuesday and nobody replied. It drafts an answer with the available slots and flags it as waiting.
A renovation contractor. The agent sees that a site visit on your calendar clashes with a supplier delivery and suggests moving one. Anthropic describes a brand strategist whose agent noticed a client call fell right before her flight and offered to reschedule it.
A boutique retailer. The agent reviews the weekend's online orders, spots two with missing addresses and prepares messages asking customers to confirm.
Notice the pattern. None of these jobs is clever. Each is the kind of small task that costs money when forgotten. That is where proactive agents earn their keep.
How much does a proactive AI agent cost?
Right now most proactive agents are bundled into existing AI subscriptions rather than sold separately. Claude Duet carries no extra charge on Pro and Max plans during its research preview, but it shares those plans' usage limits. The real cost is therefore usage: background checks and big tasks use up your allowance faster than chatting does.
This matters more than the headline price. 9to5Mac reported that Anthropic warns users could reach their limits "sooner than with chat alone." If your agent checks five accounts every few hours and you also hand it a sales forecast, you can hit a cap mid-week. Our guide to AI usage limits explains how those caps work and why your tool can suddenly stop.
Availability is the other cost. As of late October 2026, Duet is waitlist-only for US Pro and Max subscribers, and Anthropic says it is not yet on Team or Enterprise plans. A Hong Kong business cannot plan next week's customer follow-up around it yet.
What are the risks of letting an AI agent act on its own?
The three main risks are wrong judgement, too much access and false confidence. An agent can misjudge what is urgent, read accounts it does not need, or miss a message because it checks every few hours rather than instantly. Anthropic itself says Duet "might misread a thread or miscalibrate what's urgent." Approval steps reduce the damage but do not remove it.
Wrong judgement. If the agent decides a complaint is low priority, you may never see it in the morning brief. Check its list against your inbox for the first few weeks.
Too much access. Every connected account is a door. Anthropic says Duet never directly sees passwords and uses a classifier that refuses to share credit card numbers and government ID. Even so, connect only the accounts the first job needs.
False confidence. An agent that checks every few hours is not a security guard watching every message the second it arrives. An urgent WhatsApp at 2 PM may not surface until the next check.
Governance is still catching up. OneTrust's AI-Ready Governance Report, published in September 2026 from a survey of 1,200 executives, found 87% of companies encourage AI agents but only 47% have governance in place. Our explainer on AI guardrails covers the safety rules every agent should run under.
What do people get wrong about proactive AI agents?
The most common mistake is assuming "proactive" means "autonomous." Today's agents prepare work and ask permission; they are designed not to send, pay or delete on their own. The second mistake is assuming they replace staff. They replace the remembering, not the relationship, the judgement call or the phone call that saves an account.
Misconception 1: "It runs my business while I sleep." It prepares a to-do list while you sleep. You still approve each action. Anthropic's design keeps important decisions with the user.
Misconception 2: "More connected apps means more value." More apps mean more noise for the agent to sort and more doors to secure. Start with one inbox and one calendar.
Misconception 3: "It never misses anything." It checks on a schedule and can misjudge urgency. Treat it as a second pair of eyes, not the only pair.
Misconception 4: "It works the same everywhere." Features, connectors and availability vary by country and plan. A tool launched in the US may not reach Hong Kong for months.
How should a Hong Kong SME start with a proactive agent?
Start with one recurring job that costs you money when forgotten, connect only the accounts that job needs, keep every send and payment on approval, and compare the agent's list with your own for two weeks. If it catches things you missed and invents nothing, widen the scope. If it creates more checking than it saves, stop.
Step 1: Pick the leak. Write down the one task that most often slips: overdue invoices, unanswered enquiries or calendar clashes.
Step 2: Limit the doors. Connect the one account that task needs. Use a staff account with limited permissions if you can.
Step 3: Keep approval on. Never let the agent send, pay or delete without your yes during the trial.
Step 4: Run a shadow test. For two weeks, compare the morning brief against your own check. Count hits, misses and false alarms.
Step 5: Watch the meter. Note how much of your usage allowance the background checks consume before you add a second job.
Frequently asked questions
Is a proactive AI agent the same as an AI employee?
Not quite. A proactive agent is a general assistant that notices and prepares work across your apps. An AI employee is usually set up for one defined role, such as customer service or admin, with fixed procedures. The two ideas overlap and are converging.
Can I use Claude Duet in Hong Kong today?
As of 26 October 2026, Anthropic lists Claude Duet as a waitlisted research preview for Pro and Max subscribers in the US, and not yet available on Team or Enterprise plans. Check Anthropic's availability page before planning around it.
Will a proactive agent send messages without asking me?
Claude Duet is designed to show you exactly what it will do and wait for approval before sending a message, making a payment, signing a document or deleting anything. Some actions, such as moving money between accounts, are blocked even with approval.
The bottom line on proactive AI agents
A proactive AI agent is a chatbot that moves first. It watches the apps you connect, tells you what needs attention and prepares the next step, but it still waits for your approval. For a small business, its real value is catching the small jobs that cost money when forgotten.
The technology is early, the best-known version is still US-only and waitlisted, and every background check uses part of your plan. That is a reason to learn the idea now and test it carefully later, not a reason to ignore it.
The businesses that benefit first will be the ones that already know which tasks keep slipping. Write that list today. We understand AI. UD stands with you.
Reviewed by the UD AI team.
Not sure which jobs in your business keep slipping, or whether your team is ready to let an AI agent prepare them? Take the free AI Ready Check and we will walk you through it step by step, from finding the tasks worth handing over to setting the approval rules that keep you in control.