The consequence boundary
What are consequential AI actions?
A consequential AI action is an action by an AI agent that changes something outside the model and matters if it is wrong, such as moving money, changing a record or contacting a customer. These actions need a decision before they execute. Everyday read-only work, such as looking up information, usually does not.
AI Acts. Crelis Decides. · Runtime authorization for AI agents
- What it changes
- Money, records, customer contact
- When it is decided
- Before the action executes
- Outcome
- Allow, require human approval, escalate, or block
- Everyday reads
- Usually no decision needed
Why the line matters
Reading is not the same as changing
When an AI agent looks something up, the enterprise is left as it was. When it moves money, changes a record or sends a message in your name, something outside the model is now different.
That difference is the consequence boundary. On one side is everyday read-only work. On the other side are the actions that change something outside the model and matter if they are wrong. Those actions need a decision before they execute.
Each enterprise draws the line in its own policy. A simple test helps: ask how hard the action would be to put right if it were wrong.
A read leaves things as they were; a consequential action does not.
Where the line falls
Not every action carries the same weight
An AI agent can read, draft or change something. The consequential action is the kind that is decided before it executes.
Kind of action
What it does
When a decision is needed
Reading information
Looks something up and leaves things as they were
Usually none
Drafting or suggesting
Prepares something for a person to review
The person who reviews it decides
Routine change
Changes something small that is easy to put right
When your policy says so
Consequential action
Moves money, changes a record or contacts a customer, and matters if it is wrong
Before the action executes
Actions that matter, not every action.
What counts
What a consequential action looks like
Each of these changes something outside the model, and each one matters if it is wrong.
Move money
A payment, a transfer, a refund.
Change a record
An account, a limit, a customer file.
Contact a customer
A message sent in your name.
What regulators point to
Controls at the moment an agent acts
We publish an independent, primary-sourced reference mapping SAFR, NIST AI RMF and EU AI Act requirements to controls at AI-agent execution time. It cites the source documents and implies no endorsement.
If you are mapping your agents
Questions to ask about your agents' actions
A short list for the teams that run your AI agents. These are questions, not answers.
- 01Which of our agents' actions change a system of record, and which just read?
- 02Has anyone written down which of those actions are consequential?
- 03If a given action were wrong, who would have to put it right?
- 04Which consequential actions should wait for a person?
- 05What happens to a consequential action if the authority cannot be reached?
- 06After a consequential action, can we show what was decided?
Read next
Go deeper
Pages and articles that take one part of this further.
- What is runtime authorization for AI agents?the decision itself, explained
- What is an AI agent audit trail?what the record of a decision should prove
- GREENLIGHTthe product page
- Agentic AI Controls: A Working Definitionwhat counts as a control on an agent action
- AI Agents in Regulated Compliance Workwhere assisting ends and signing begins
FAQ
Consequential actions, in short answers
The questions people ask about which AI agent actions need a decision.
What is a consequential AI action?
A consequential AI action is an action by an AI agent that changes something outside the model and matters if it is wrong, such as moving money, changing a record or contacting a customer. These actions need a decision before they execute. Everyday read-only work, such as looking up information, usually does not.
Which AI agent actions need approval?
The consequential actions that your policy says should wait for a person. Policy decides the outcome: allow, require human approval, escalate, or block. How an approval is decided and recorded is explained on the runtime authorization page.
Why do AI agent writes differ from reads?
A read retrieves information and leaves things as they were. A write changes something outside the model: a payment is sent, a record is altered, a customer is contacted. The two carry different risk, so the decision is placed where an agent does something consequential.
Does every AI agent action need a decision before it executes?
No. The decision is for the moment an AI agent does something that matters, not for every step it takes. Everyday read-only work is untouched. The decision itself is explained on the runtime authorization page of this site.
Are all AI agent actions that change systems consequential?
No. Consequential actions are a subset of the actions that change something. A change is consequential when it matters if it is wrong: it moves money, alters a record people rely on, or reaches a customer. Each enterprise sets that line in its own policy.
Can reading data be a consequential action?
Usually not. A read retrieves information and leaves things as they were, so nothing outside the model changes. Each enterprise draws the line in its own policy, and that policy can name any action the enterprise decides matters.
What does GREENLIGHT decide about a consequential action?
Before an AI agent does something that matters, GREENLIGHT decides whether it should happen: allow, require human approval, escalate, or block. Every allowed action earns an execution visa, and every decision is recorded as proof. Crelis authorizes or routes the action; your own system executes it.
Give your agents a green light — safely.
Start in shadow mode on non-production traffic. See every decision GREENLIGHT would have made — before you let it make one.
