How we use AI: our approval model
This page describes how we work when AI agents do part of the job, for us and for clients.
This page describes how we work when AI agents do part of the job, for us and for clients. An agent is software that can read, decide and act through your tools. That is useful and it is a risk, so we treat every agent the way we would treat a new employee with a narrow job: a named supervisor, limited access, and a record of everything it does.
These are our working principles, not a guarantee of any outcome, and they are applied to each engagement during scoping. We publish them so you can hold us to them, and so you have something to compare with any provider that offers you automation.
What an agent may do on its own
Every task is placed in one of three levels before an agent touches it. The level decides who has to say yes.
- Level 1
Runs on its own
Reading and sorting: classifying incoming documents, drafting a reply for review, summarizing a ticket queue, preparing a report from data the agent can only read.
Green: no approval needed
- Level 2
Needs approval
Anything with an effect: changing a security setting, deleting data, sending an external message on your behalf, disabling an account, or filing an entry in a financial system. A person reviews and approves each one.
Amber: a person approves
- Level 3
Never automated
Approving payments or moving funds, handling credentials or recovery codes, resetting a password on the strength of a voice or email request alone, and signing contracts.
Red: never automated
Eight rules we hold ourselves to
Every agent has a named owner
A person at Clayton, and one on your side, answers for what the agent does.
Least privilege per agent
Each agent gets its own identity with only the access its task needs, never a shared or administrator login.
Read-only by default
An agent starts by reading and drafting; write access is added one permission at a time.
Changes need human approval
Anything that alters a system, sends a message outside the company or moves money waits for a person to approve it.
Every action is logged and reviewable
What the agent saw, decided and did is recorded in a form a person can read.
The client can see the log
You get access to the record of what agents did in your environment.
A kill switch for every agent
A named person can stop an agent immediately and revoke its access.
Regular review
We check each agent's access, results and logs on a schedule and remove what it no longer needs.
Three approval tiers for AI agents
Every action an agent can take falls into one of three tiers: it runs on its own, a named person approves it first, or a human always does it. The examples are illustrative; each firm draws its own lines during scoping.
What you can and cannot see about AI agents
Each log or tool shows part of the picture. Some questions, such as the prompt text or the model version behind a copilot, are not answered by any of the four. Coverage differs by product and configuration, so it is verified per tenant. Qualitative view.
Frequently asked questions
Direct answers to the questions we hear most. Something else on your mind?
Send us a messageWhat can an AI agent do on its own at Clayton Cybernetics?
Agents run on their own only for reading and sorting work. Examples are classifying incoming documents, drafting a reply for a person to review, summarizing a ticket queue and preparing a report from data the agent can only read. Anything with an effect on a system moves to the approval tier.
What needs a person's approval?
Anything with an effect. That includes changing a security setting, deleting data, sending an external message on your behalf, disabling an account and filing an entry in a financial system. A person reviews and approves each one before the agent acts. Agents start read-only, and write access is added one permission at a time.
What is never automated?
We never automate approving payments or moving funds, handling credentials or recovery codes, resetting a password on the strength of a voice or email request alone, or signing contracts. Those actions stay with named people, whatever the agent recommends.
Who is responsible for each agent?
Every agent has a named owner: a person at Clayton Cybernetics and a person on your side. Each agent also has its own identity with only the access its task needs, never a shared or administrator login. The owners answer for what the agent does.
Can we see the log of what an agent did?
Yes. Every action is logged in a form a person can read: what the agent saw, what it decided and what it did. You get access to the record of what agents did in your environment, so you can check the work yourself.
How do you stop an agent that misbehaves?
Every agent has a kill switch. A named person can stop it immediately and revoke its access. Because each agent has its own identity, stopping one does not disable the others or any person's account.
How often are agents reviewed?
We check each agent's access, results and logs on a schedule and remove any access it no longer needs. The review cadence is set for each engagement during scoping, and the principles on this page are applied to each engagement there.
Is this a guarantee of how agents will behave?
No. These are our working principles, not a guarantee of any outcome. We publish them so you can hold us to them and compare them with any provider that offers you automation. How they apply to your environment is settled during scoping.