Automation or decision support?
Automation executes actions under defined rules. Decision support presents recommendations for a person to assess and act on. Depending on risk, automation may use pre-authorisation, exception review, monitoring or post-action review.
Acceptance data can identify candidate tasks, but safe automation also requires validation of outcomes, error impact, permissions and recovery paths.
Table of Contents
Summarize with:
Key facts
- Automation: the system executes actions within defined rules and controls.
- Reversibility, model accuracy, volume and regulation decide between them.
- Unchanged acceptance is one signal to investigate, not proof that a task is safe to automate.
- Redbrix provides undo or compensating actions where possible and prevents or explicitly authorises consequences that cannot be reversed.
Choosing
| Factor | Decision support | Automation |
|---|---|---|
| Reversibility | Hard to reverse | Easy to reverse |
| Model accuracy | Varies by case | Consistently high |
| Volume | Low to medium | High |
| Regulation | Sign-off required | Review after the fact allowed |
Execution visibility
Redbrix makes the progress of an AI operation visible and connects it to the plan and data the user approved.
Design implication
Why it matters
The choice changes the interface more than the model does. Decision support needs suggestion and review screens inside the workflow. Automation needs a log of what the system did, exception queues and undo. Products often need both for different cases in the same workflow.
How it shows up in the product
| Mode | Interface |
|---|---|
| Decision support | Suggestions with reasoning, accept or reject in one step, acceptance tracked |
| Automation | A log of actions taken, a queue of exceptions, review after execution, undo |
Options and trade-offs
- Manual approval kept for thousands of trivial items wastes time. Acceptance rates alone can hide errors or habitual approval; Redbrix evaluates the actual outcomes.
Common mistakes
- Automating before validating outcomes and the consequences of errors
- Manual approval kept for thousands of trivial items
- No prevention, recovery or compensation plan for automated actions
- No record of which actions the system took on its own
Where Redbrix has worked on this