An AI assistant answers your requests one at a time and waits for your next instruction. An AI agent takes a goal, works out the steps, uses your apps, accounts and websites to carry them out, and can keep going while you are away, stopping only to ask for approval. Chatbots and classic voice assistants are assistants; Meta Muse, OpenAI dots and Claude Duet are agents. In 2026 the two are converging, which is why the terms get confused.

What is an AI assistant?

An AI assistant is reactive. You ask, it responds. Classic examples include:

  • Chatbots like standard ChatGPT, Claude or Gemini chat: they answer questions, write drafts and analyse files inside a conversation.
  • Voice assistants like the original Siri, Alexa and Google Assistant: they set timers, play music and run single commands.

Assistants can be very capable, but each turn ends when they reply. If a task needs ten steps across three websites, you do most of the clicking.

What is an AI agent?

An AI agent is goal-driven. You describe an outcome and it plans and executes the steps itself. The 2026 personal agents share a few traits:

  • Their own computer. Muse, dots, Claude Duet, Grok Bot and Gemini Spark each run on a cloud computer with a browser, so they work while your phone is off.
  • Connected accounts. They link to email, calendars, payment tools and business apps, often through MCP connectors.
  • Memory. They remember preferences and ongoing tasks across days.
  • Proactivity. Claude Duet sends a morning brief; Muse suggests actions from things you mentioned once.
  • Approvals. They pause before sending, paying, signing or deleting.

Anthropic's description of Duet captures it: an app where Claude "comes to you with what needs your attention, completes tasks you ask of it, keeps working when you step away, and checks with you before it acts on your behalf" (Claude). For a fuller definition, see what is a personal AI agent.

AI assistant vs AI agent at a glance

AI assistant AI agent
Starts work When you ask When you ask, on a schedule, or on its own
Scope One question or command A goal with many steps
Tools Mostly text, some single actions Browser, connectors, files, payments
Runs while you're away No Yes, usually on a cloud computer
Memory Often limited to the chat Persistent across days
Main risk Wrong answers Wrong actions in your accounts
Examples (as of November 2026) ChatGPT chat, classic Siri, Alexa Meta Muse, OpenAI dots, Claude Duet, Grok Bot, Gemini Spark, OpenClaw

Where do Siri, Alexa+ and Copilot fit?

The line is blurring, and the big assistants are moving toward agents.

  • Siri AI (iOS 27) can search personal context and run multi-step commands across apps. But it has no documented always-on computer or long background tasks. Analyst M.G. Siegler called it a "dark horse" that is "not on par with Muse or Dots" (Spyglass). See our Siri AI review.
  • Alexa+ has gained agent-style features for orders and bookings, covered in Alexa+ vs personal AI agents.
  • Microsoft Copilot rebuilt itself in September 2026, folding in Autopilot, which grew out of an OpenClaw-based project. Autopilot is a business agent inside Microsoft 365, in private preview and requiring Microsoft 365 Copilot ($30 per user a month) plus Copilot Credits; it is not a consumer feature.
  • ChatGPT has agent mode for one-off tasks, and dots for persistent, always-on work. The distinction is explained in OpenAI dots vs ChatGPT agent mode.

A useful rule of thumb: if it keeps working after you close the app and comes back to ask permission, it is an agent.

Why does the difference matter?

Usefulness

Agents can take on whole jobs: chasing invoices, rescheduling meetings, comparing and buying products. That can save real time on admin. Assistants save time per answer, but you still drive.

Risk

Agents read untrusted content such as web pages and emails, which exposes them to prompt injection. Meta has acknowledged the problem is not fully solved, and Anthropic says "no defense is perfect." Our prompt injection explained guide covers it.

Cost

Assistants are often free or bundled. Agents cost more to run because they use a cloud computer and many model calls. As of November 2026, entry prices range from Meta Muse's free tier and Claude Duet on Claude Pro ($20/month) to OpenAI dots on ChatGPT Pro ($100/month). See personal AI agent pricing.

How do people feel about letting AI act?

Willingness lags behind capability. A Gartner survey of 322 US consumers in January 2026, reported by secondary sources, found 31% willing to let AI narrow their choices for household supplies, but at most 11% willing to let AI make the purchase decision. That gap explains why every major agent leans on approval gates: companies are betting that people will delegate more once they can review each consequential step. It also explains why proactive features, such as Claude Duet's morning brief and Muse's suggestions, are framed as offers rather than actions. If you are unsure, start with an agent in a mostly read-and-suggest role and widen its permissions only as it earns trust.

Which should you use?

  1. Mostly questions and draftsan assistant like standard chat is enough
  2. Simple phone actionsSiri AI or Alexa+ cover quick multi-step commands
  3. Repetitive admin across appstry an agent with approvals on
  4. Business workflowscompare Claude Duet, dots and Grok Bot
  5. Full control and any modelconsider self-hosted OpenClaw

Most people will use both: an assistant for quick help and an agent for delegated work. Our how personal AI agents work explainer covers what happens under the hood once you delegate. UK readers can start with personalagent.uk's guide.