A task needs AI only when one step's input is unpredictable enough that a fixed rule can't cover it, and someone has to make a judgment call. If the logic fits a plain if/then statement, it just needs automation. Most real workflows mix both, so the real question isn't whether the whole task needs AI. It's which steps do.
The quick answer: it's not the task, it's the step
A real workflow is almost never all-AI or all-automation. It's a mix: five or six steps that just move data and follow rules, one or two that actually need a judgment call. So “does this task need AI” is the wrong question to ask. The one that matters is which steps in your workflow need AI, and which are just rules a $20/month automation plan already handles without you paying a cent extra.
This guide gives you a named test for any step (the Step Triage Test), a quick-reference table, and a worked example that labels every step of a real small-business workflow.
Why “does this task need AI” is the wrong question
As reported by Zapier's analysis of 375 mid-market and enterprise companies (published 2026-07-29), AI accounted for just 18% of workflow steps. The rest ran on ordinary rules and plain data-moving, nothing fancy. Workflows that reserved AI for the steps that actually needed reasoning cost 71% less to run than workflows that routed everything through a model. That sample skews mid-market and enterprise, not small business, so treat it as directional rather than a number that applies to you exactly. But the direction holds either way: even companies with real budgets aren't putting AI on most of their steps.
On the steps that do use AI, Zapier's data breaks the work into four repeatable jobs:
| AI job type | What it looks like in practice |
|---|---|
| Drafting for a human to review | A first-pass reply or summary a person reads before it goes out |
| Filling records from messy input | Pulling structured fields out of a free-text email, form, or PDF |
| Deciding where work goes next | Routing a request by tone, urgency, or intent |
| Turning a request into an owned task | Converting an inbound message into a ticket with the right owner |
This is separate from a question this site already answers elsewhere: the broader test for whether you need an AI agent at all. This piece is the narrower follow-up. It assumes you already have a workflow, or are about to build one, and just want to know which steps in it need AI, not the whole thing.
As one operator put it on X: “If a good old if/else is faster, cheaper, and more reliable, use that” (Pawel Huryn, July 2025). Hard to argue with that.
The Step Triage Test: 3 questions for any step
Run these three questions on one step at a time, never on the whole workflow at once. A workflow isn't one decision, it's a chain of them.
Can you write the rule?
If you can describe the logic in a plain if/then sentence covering roughly 90% or more of real cases, it is a rule, not a judgment call. “Dedupe by email address,” “route by zip code,” “flag anything over $500” are all rules. No model needed, no matter how official the step feels.
Does the input vary in a way a fixed rule can't handle?
Free text, inconsistent formatting, or genuinely ambiguous cases push a step toward AI, because you cannot enumerate every pattern in advance. An email that could be a complaint or a compliment depending on tone is this kind of step. A five-digit zip code field is not, no matter how you squint at it.
What does a wrong call cost you?
If a mistake is cheap and easy to catch, a lighter check is fine even with AI running the step. If it's expensive or hard to reverse, money sent to the wrong place, a customer promised something you can't deliver, that step needs a human checkpoint regardless of which one runs it.
Rule of thumb: two “yes” answers pointing toward a rule usually means automation. A genuine “no” on question one is the real signal for AI. Complexity, or a step that just “feels smart,” is not a signal on its own.
Step Triage Table: the quick reference
| Signal | Points to plain automation (rules) | Points to AI (judgment) |
|---|---|---|
| Input format | Structured fields (zip code, dollar amount, dropdown value) | Free text or messy, inconsistent input |
| Exception paths | One or two fixed paths you can enumerate | Many cases you cannot list in advance |
| Consistency requirement | Must behave identically every time | Some variation is fine, or even wanted |
| Cost of a wrong call | Cheap and easy to catch, easy to reverse | Expensive or hard to reverse |
| Example | Route a lead by zip code | Decide if a support message is urgent |
Land mostly in the left column, and it's a rule: build it as a filter or a fixed path, and move on. Land mostly in the right column, and that's where AI actually earns its cost, and where a human check still belongs too.
A worked example: triaging a real small-business workflow
Here's how a typical new-lead intake workflow breaks down, step by step. This is illustrative, a common shape for this kind of workflow, not one we built or tested ourselves.
- Capture the form submission. Rule. Structured fields, zero judgment involved.
- Check for a duplicate contact. Rule. Just matching on email or phone.
- Verify the email format. Rule. A pattern check that behaves the same way every time.
- Route the lead by territory or product line. Rule. A lookup against values you already know.
- Draft a first personalized reply.Judgment worth AI help, since the input is free text. But it's a drafting job, not a decision, so it still needs a quick human glance before it sends.
- Decide whether this lead needs an immediate follow-up call or can wait for the weekly batch. The one genuine judgment step in the whole list. Urgency comes from messy, inconsistent signals, not a field you can check.
Four of six steps are plain rules, one is AI-assisted drafting with a human check, one is the real judgment call. Mostly rules with a small AI core, close to Zapier's 18% figure at a larger scale.
What it actually costs to get this wrong
Workflows that reserve AI only for judgment steps cost 71% less to run than workflows that route everything through a model, per Zapier. Put AI on all six steps of the intake workflow above instead of just step six, and you're paying model cost on five steps a filter already handles for free.
“AI steps cost more” isn't automatically true, and that trips people up. A plain rule step and a bare-bones AI step can both cost 1 task on Zapier's system. What actually changes the price is the model tier (see how automation pricing actually works: tasks vs operations vs credits): Standard runs at 1x, Advanced at 3x, Premium (the default for new AI steps) at 5x, before tool calls are even counted. Zapier's own example makes the gap concrete: Standard with no tools costs 1 task; the same step on Premium with two tool calls costs 15 tasks. Gumloop prices the same shape in different units. Per Zapier's breakdown of Gumloop's pricing (June 8, 2026), a base workflow step costs 1 credit, a standard AI model call costs 2, an advanced one 20, and an expert-tier call 30 or more. Worth noting that breakdown is written by a competitor, so treat the price facts as reported and the verdict as vendor-biased.
There's a reliability angle too, and it's worth taking seriously. Even the current best model on Zapier's AutomationBench benchmark, Claude Opus 5 (Max), completes only about a quarter of real multi-step workflow tasks unaided (26.2%, July 2026). That's a reminder AI steps need a human checkpoint more than rule-based ones do, not less, one more reason why AI agents fail more often than a plain rule does. Platform reliability matters too: as reported by users on Trustpilot, “Zapier has had many outages, we have to fix the broken runs ourselves.” That's a general platform complaint, not an AI-specific one, but the underlying point still holds: more judgment-based, less predictable parts is more that can break.
We send one honest breakdown like this a week, including the “you don't need AI for that” verdicts, in the AgentsExplained newsletter.
Common mistakes when triaging steps
Treating the whole workflow as one yes/no AI decision. That's the broader question of whether you need an AI agent at all in the first place, already covered elsewhere on this site. Triage step by step instead, this piece is about the steps, not the whole thing.
Adding AI to a step “just in case it gets complicated later.” Wait for a real judgment need to actually show up. A rule you haven't needed yet isn't a reason to pay a 5x multiplier today.
Never revisiting the triage as the workflow scales.A rule covering 95% of cases at 50 leads a month may not hold at 500. Re-run the test as volume changes, this isn't a one-time label.
Skipping the human-review checkpoint on the one step that is AI. Once the earlier rule-based steps have run cleanly for a while, it's tempting to trust the whole chain. Don't. The judgment step still needs a glance, worth checking how to test an automation before it goes live before you trust it fully.
FAQ
What's the difference between AI and automation? Automation follows fixed rules you write in advance. AI makes a judgment call on input it has not seen in exactly that form before. The Step Triage Test above tells you which a given step actually needs.
Can I use AI and automation in the same workflow?Yes, and most real workflows do. A lead-intake process might run four rule-based steps and one AI step in sequence, that's the whole point of triaging by step instead of picking one approach for the entire workflow.
How do I know if a specific step needs AI or just a rule? Run the Step Triage Test: can you write the rule for roughly 90% of cases, does the input vary in ways a fixed rule can't cover, and what does a wrong call cost you. A genuine “no” on the first question is the real signal for AI.
Is AI automation more expensive than plain automation? Generally, yes. Zapier found workflows reserving AI for judgment steps cost 71% less to run, though the figure comes from a mid-market and enterprise sample, so treat it as directional for a small business rather than a guarantee.
What tasks are best suited for automation vs AI? Structured, rule-describable input suits automation. Free text or ambiguous cases suit AI. Use the Step Triage Table above to check.
Should a small business put AI on every step of a workflow? No. Most steps are rules, and running them through a model instead of a filter mainly adds cost and another point of failure. Save AI for the steps that need judgment.
The short version (and where to go next)
It's a step question, not a task question. Run the Step Triage Test on each step: can you write the rule, does the input vary too much for it, what does a wrong call cost. Most steps do not need AI: as reported by Zapier, AI made up just 18% of workflow steps studied, and reserving it for judgment steps cut cost by 71%. Triage again as the workflow scales. This isn't a set-and-forget call, it's a habit.
Still deciding whether an entire project needs an agent in the first place? That question already has an answer elsewhere on this site. Once a step is confirmed to need AI, the next decision is which AI skill or tool fits it.
For more honest, no-hype breakdowns, including the verdicts where the answer is “skip it,” the AgentsExplained newsletter sends one a week.