What are AI agent risk tiers, and which actions wait for my review?
I sort every agent action into low or high risk, and only the high-risk ones wait for my review. That one split decides whether something goes out directly or lands in front of me first. I call the sorting rule an AI agent risk tier. It is written down before the action happens, so the agent does not have to guess in the moment whether to ask.
Email is where I started. A reply that confirms a meeting time or sends a file someone asked for is low risk, so the agent sends it. Anything that commits money, makes a promise, goes to someone new or could be read the wrong way is high risk. For those the agent writes a draft and I read it before it goes out.
What is my test for low risk?
A mistake is cheap to notice and cheap to undo. Both halves have to hold.
The two halves catch different failures. A wrong time in a meeting confirmation is easy to see, and a correction costs one short message. A promise sent to a new contact might be noticed late and cannot be taken back, so it fails on both counts. An action that is easy to undo but that nobody would notice for a week fails the first half, so it is high risk.
How does the same split work for dictation?
I dictate prompts to coding agents with Wispr Flow, and the same two tiers apply there.
A short prompt to a coding agent is low risk. I end it with the spoken words “press enter” and it submits straight away. A misheard word costs me one follow-up line, which is cheap to notice and cheap to undo.
A message to a person is different, and so is a prompt that tells an agent to delete or deploy something. Those stay on screen until I have read them.
The spoken command depends on the recognizer hearing it. Wispr Flow has a press enter setting, and it is switched on in my configuration (the setup and its hit rate have their own page). On October 6, 2026 I added seven corrections to its personal dictionary for phrases the recognizer produced in place of “press enter”, such as “Presenter” and “in the present”, each mapped back to the command. I read that from the app’s local database on one Mac on October 9. The goal I gave myself was to stay off the keyboard apart from holding the key that starts dictation, and every missed command is a place where I would have to reach for it.
Which tiers do I already have written down?
The same idea already shows up in several places where I work.
My instructions for coding agents carry a three-level list, the same one I use to route AI code reviews to a human. Schema migrations, authentication logic, payments, deleting user data, changes to an API that outside consumers depend on, infrastructure changes and security-sensitive changes always stop and ask. Low confidence and ambiguous requirements proceed but get flagged. Well-defined tasks with passing tests, bug fixes with a clear reproduction, formatting and documentation updates go ahead without asking.
The same file lists what an agent may do without a prompt, such as committing and pushing to my own repositories, and what still needs my confirmation, such as a force push, a hard reset on a published branch or any action on a repository I do not own.
My executive assistant’s charter follows the pattern for people instead of code. Anything that commits my time or money, sends on my behalf, books travel or speaks for me externally is drafted for my approval and not sent. That line is stricter than the email example above, so the charter needs a low-risk exception before the two agree.
My LinkedIn pipeline has the clearest hard gate. Agents draft posts, but the command line cannot approve one and no script can. Only a listed approver can, through the page.
I am now writing a tier down for as much of my work as I can, including calendar invites, shared files, code merges and social posts.
What does a wrong tier cost?
A low-risk action filed as high risk costs me the time to read something that did not need me. A high-risk action filed as low risk is the expensive error, because it goes out unread.
That asymmetry is why the test has two halves, and why a failed half sends the action to high risk. It is also why I am writing each tier down before an agent acts in that category.
Does it save time?
I have no numbers on this yet. What I expect is less time approving things that never needed my eyes, and more attention for the few that do. A fair test would count my approvals per week before and after a tier goes live. I have not run it.
Who does this fit, and how do I start in 15 minutes?
It fits anyone whose agents already produce more output than they can read. It adds little if your agents only work on drafts you review anyway.
List ten things an agent did or drafted for you this week. Mark each one with the two questions: easy to notice, easy to undo. Write the ones that passed both as the low-risk list, in a file the agent reads at the start of every session, and write the rest under high risk with the instruction to draft and wait. Start with email, since it is the easiest to sort, and add a category only when you have written its line.
This page belongs to the Compounding CEO series.
Frequently Asked Questions
What is an AI agent risk tier?
It is a rule written in advance that says whether an agent acts on its own or asks first. I use two tiers. Low-risk actions go out directly, and high-risk actions wait for my review. The agent then knows before it starts whether to act or to ask.
How do I tell a low-risk agent action from a high-risk one?
I ask two questions: is a mistake cheap to notice, and is it cheap to undo? If both answers are yes, the action is low risk. If either fails, it is high risk. Anything that commits money, makes a promise, goes to someone new or could be read the wrong way fails the test.
Is it safe to let an agent send email without review?
For narrow cases, yes. A reply that confirms a meeting time or sends a file someone asked for is low risk in my setup, so the agent sends it. Anything that commits money, makes a promise or goes to a new recipient is drafted and waits for me.
Do risk tiers save time?
I have no numbers on this yet. What I expect is less time spent approving things that never needed my eyes, and more attention for the few that do. I will only claim a saving once I have counted one.
Ready to Transform Your AI Strategy?
Get personalized guidance from someone who's led AI initiatives at Adidas, Sweetgreen, and 50+ Fortune 500 projects.
