AI 智能体的两种类型：个人助理与虚拟员工

Someone recently let Meta’s Muse AI agent sell a keyboard on Facebook Marketplace for him. Muse agreed to a lowball price, gave a buyer the seller’s home address, and replied “Yep I’m here!” when he wasn’t available. It didn’t check with the seller about any of it until the buyer was already standing outside, and all of it was done under the seller’s name and account.

Muse owned the sale (it has its own computer and works independently), but it put the seller’s name on it. He ended up accountable for actions he didn’t direct or fully understand.

This points to something more general: There will be two very different types of AI agents, and what separates them is who owns the work:

Personal assistant. This agent will be co-located with your personal data and will help you do your work. It can act, but it will do so on your behalf and under your name. You retain ownership of this work.
Virtual employees. These agents will be separate entities; they will have their own name, computer and accounts. They work for you but they own their own work; you hand them a responsibility, not a task.

Personal assistants are close to what people today already use ChatGPT and Claude for. But they will be much more capable: Instead of being constrained to a chat window, they will have access to your files, emails and calendar and can perform actions directly on your behalf.

Virtual employees will look a lot like human employees: You give them their own computer, provide them with a company email account, and invite them to Slack or Teams. From this point on, they own their area of responsibility, and you manage them by outcomes rather than checking every step. If they need access to something, you forward them an email, add them to a Google Drive folder, or send them a file via Slack just like you would with a human employee. Similar to human employees, you might hire multiple agents with different roles and personalities and you might fire them if they don’t work out.

The two types differ in the following ways:

	Personal assistant	Virtual employees
Ownership	You own the work	Agents own their own work
Accounts	Uses your existing accounts	Use their own, dedicated accounts
Data	Co-located with your personal data (file system, emails, calendar, …)	Have their own file system and access to shared data
Integration	Deep OS-level	Separate VM
Oversight	Approvals and permissions	Trust, scope and outcomes; may get fired
Number of instances	One	Many

I expect that almost every person will have one, and only one, personal assistant agent. Some people will have a few virtual employees and even fewer will have effectively a whole company of virtual employees. More commonly though, existing companies will start to hire these, and they will work alongside human employees. Lastly, your personal assistant will likely help coordinate your virtual employees.

Some of this is already visible today:

Apple is clearly building the personal assistant with Apple Intelligence and its semantic index for accessing all your personal data.
Meta’s Muse, while very cool and well-executed, ultimately is a confused product: It is built like a virtual employee but acts like a personal assistant.

Here’s what I expect to follow:

By default, OS vendors (Microsoft, Google and Apple) will win on the personal assistant front for two reasons. First, they have full access to everything: your data, accounts, and apps already live on their platform, which is exactly what today’s chatbots lack. Second, since the work is yours (it happens under your name and you’re accountable for every action), personal assistants need strong guardrails and permissioning to avoid catastrophic outcomes, and only the OS vendor can enforce access control across the entire system.
Enterprise software will continue to do well: Virtual employees are just another seat, and enterprise access control and auditability are important. In fact, some consumers might start to use these enterprise tools to manage their personal virtual employees.
Legal liability for virtual employees is an open question and resolving it will be important: If they work independently and under their own name, who is liable for their actions?

Finally, if I were to build something in this space, I’d 100% focus on the second type. Competing with Apple, Google, and Microsoft on personal assistants will be very difficult, but virtual employees neutralize both of their advantages. They don’t need access to everything; they get access the way a new hire does, through forwarded emails, shared folders and Slack invites, which all enterprise software already supports. They also don’t need OS-level guardrails, since they act under their own accounts, and those are governed by the same access controls that already apply to human employees. In short, virtual employees are a level playing field whereas personal assistants are not.