AI Guides › Skills & Agents
The Difference Between Automating A Task And Automating A Decision
By Nigel Guy · 3 min read
A task is "format this into a table." A decision is "which of these three candidates should we move forward with." Both can be handed to a skill, and both will produce a confident-looking output — which is exactly the problem, because only one of them should be trusted the same way the other is.
The rule: automating a task speeds up something you'd have done the same way regardless; automating a decision replaces judgement — the second needs far more scrutiny than the first, and treating them the same is how a fast, useful skill turns into a quietly bad one.
The mechanism
- Ask what the output actually is. A formatted document, a sorted list, a drafted email — these are tasks; the "right" version is largely mechanical, and there's usually one correct-ish shape. A recommendation, a ranking, a "should we" — these are decisions, and reasonable people looking at the same information could land somewhere different.
- Check whether the skill is being asked to weigh things, not just arrange them. Sorting by a stated criterion is a task. Choosing the criterion, or trading off several criteria against each other, is a decision. A skill can do either, but only one of them should run without a human actually reading the reasoning.
- Require the reasoning to be visible for anything decision-shaped. A task's output can reasonably stand on its own. A decision's output needs its reasoning shown alongside it — not because the answer is necessarily wrong, but because judging whether it's right requires seeing how it got there.
- Match the review to which one it is. Spot-check a handful of task outputs and move on. Read every decision output in full, at least until you've built enough evidence of where it tends to go wrong, in what kind of case, and why.
- Watch for decisions disguised as tasks. "Just summarise the applications" sounds like a task. If the summary quietly determines who gets shortlisted, a decision is happening inside what was framed as formatting — and it deserves the scrutiny of one.
What to skip
Skip trying to remove all judgement from a genuinely decision-shaped process just because it would be convenient to automate fully — some decisions are supposed to be slow and considered, and a skill that speeds up the mechanics shouldn't be mistaken for one that's made the decision safely faster too. And skip applying decision-level scrutiny to genuinely mechanical tasks — that's just wasted caution that makes the useful, low-risk automation feel not worth the setup.
Guardrails
- The line between "task" and "decision" isn't always clean — many real requests are a task wrapping a small decision, or vice versa. When in doubt, treat it as decision-shaped; the cost of over-caution here is much lower than the cost of under-caution.
- A skill can present a decision's reasoning clearly and still be wrong — visible reasoning helps you judge the output, it doesn't verify it for you.
- Anything that affects a person — hiring, credit, access, discipline — is decision-shaped by definition, whatever it looks like on the surface, and deserves review that goes beyond what this guide covers; that's a conversation for whoever owns that process, not a skill-design question alone.
All 751 AI guides · JulieMango plans from £17/mo