An automated decision is one made by a system without a person deciding it at the moment it takes effect. The system applies its rules or judgement to the facts and the outcome follows. If a person reviews and approves before anything happens, the system has recommended; it has not decided.
Why does the distinction matter?
Automated decisions carry different duties from assisted ones. People affected may be entitled to an explanation, a route to challenge, or a human review. Staff who oversee agents need to know which of the agent's actions are decisions in their own right, because those are the ones they will be asked to explain after the fact rather than approve before it.
The distinction also decides where training effort goes. A reviewer approving held drafts needs to judge cases one by one. A person overseeing automated decisions needs to read patterns: samples, overturn rates, complaints, drift.
How does it work with agents?
Depth gives a plain way to tell. At depth 1 the agent reads and at depth 2 it drafts for a person to send, so a person decides. At depth 3 the agent executes routine work within bounds, and at depth 4 it moves money or takes actions that are hard to undo: those are automated decisions whenever no person approves first. A held action is not automated; it waits for a named person.
Every verdict, whether allowed, held, refused or withheld, is filed in an append-only record that names the bound that decided it. That record is what lets a person explain an automated decision later. Reading it is a skill, set out in what decision literacy is.
What is an example?
Take a lender whose agent is certified at depth 3 to approve payment-date changes within limits the lender set. A customer asks to move their payment by five days, within the limit; the agent agrees and updates the account. That is an automated decision. Another customer asks to move it by forty days, beyond the limit; the agent holds the request for a collections officer. That one is not. Same agent, same day, two different kinds of outcome.
How is it different from automated decision-making under GDPR?
"Automated decision" is a plain description. Under GDPR Article 22, and the UK GDPR equivalent, the legal term covers decisions based solely on automated processing that produce legal or similarly significant effects on a person, and those carry specific restrictions and safeguards. Many automated decisions, such as routing a ticket, fall outside it. The legal test is explained in automated decision-making under GDPR, with UK and South African angles in the ICO and automated decision-making by agents and POPIA and automated decision-making.
A token human glance does not turn a solely automated decision into a reviewed one. The review has to be real.
Which terms are related?
- Held action: drafted and waiting for a named person, so not automated.
- Autonomy decision matrix: sets which capabilities may decide alone; see the autonomy decision matrix template.
- Meaningful human review: what turns an automated outcome into a reviewed one.
For the people side, read AI literacy and certification for people who work with agents.
Questions people ask
Is an agent's refusal an automated decision?
It can be. If the agent refuses because the action is outside its bounds and nobody reviews it, the refusal is the outcome. Where a refusal affects a customer significantly, route it to a person who can explain and reconsider.
Who should be trained to explain automated decisions?
Anyone who handles complaints or challenges about them, and the agent's owner. They need to read the decision record, name the bound that decided and say what would have changed the outcome.