Use this whenAI agents are multiplying, and you need to decide which ones each person owns and which ones the organization manages.
This one looks ahead. It comes from my essay on company-issued AI agents, and describes where I think agents in large organizations are heading.
Why it exists
Give everyone the same data and tools and you still get wildly different results. Some setups expose information, or feed bad answers back into company knowledge. Without a clear split, nobody knows which agents the company actually depends on, or who is responsible when one gets something wrong.
The framework
- Contracts
- Finance
- Design
- Engineering
Personal agent
Who does this agent work for?
It learns how one person works: their priorities, who they work with, and what they want flagged. It prepares, follows up, and works in systems on their behalf. That person owns it and checks its work.
Shared expert
Who is accountable for its answers?
A managed specialist in an area like contracts, finance, design, or engineering, with deep company context. Internal experts test it, look into its failures, and improve the version everyone uses.
Delegated authority
What can it do without asking?
This isn’t a third kind of agent. It’s the rule that sits between the other two. Calling something a legal expert doesn’t let it approve a contract. The organization sets permissions, and they’re earned through use. Drafting a message well doesn’t earn the right to send it. The trust ladder lays out the rungs.
How to use it
- List the agents people already use or are building, and sort each one into personal or shared.
- Give every shared expert an owner who knows what good looks like and will take it offline when it’s wrong.
- Write down what each shared expert can do, what needs review, and who owns the result.
- Let personal agents call shared experts instead of everyone rebuilding them.
- Expand permissions only when real use shows they’re earned.
What it looks like
Take a vendor proposal. Your personal agent already knows why you need the vendor and what’s still missing. It asks the shared contract expert to check the terms against current guidance, and a finance expert to check the commercial assumptions. You make the decision.
Where it breaks down
Shared experts concentrate risk as well as quality. When one gets something wrong, the mistake spreads further than any personal setup could. Owning one is a standing job, not something you finish at launch.
Try it with AI
Help me sort the AI agents in my organization into personal agents and shared experts. For each agent I describe, ask: 1. Who does it work for? 2. Who is accountable for its answers? 3. What can it do without asking, and what needs review? Then flag shared experts with no clear owner, personal setups other people now depend on, and permissions that haven't been earned through use.