Most people building AI automations in Zapier do not know that the AI step they added last week is probably billed at five tasks every time it runs. Not one. Five. According to a pricing teardown that DevToolLab checked against Zapier's help pages on 6 October 2026, the Premium model tier is now the default for new AI by Zapier steps on paid plans, and Premium costs 5x.
If your task count has jumped and you cannot see why, this is the first place to look. This guide shows you how Zapier counts AI steps, a four-move audit that usually cuts the bill without losing quality, a copy-paste prompt that turns three AI steps into one, and the point where switching tools makes more sense than tuning.
How does Zapier count tasks for AI steps?
Zapier bills one task for every successful action step, but AI by Zapier steps are multiplied by model tier: 1 task on Standard, 3 on Advanced and 5 on Premium. Premium is the default for new AI steps on paid plans. A step using your own API key costs 1 task, and triggers, filters, paths and formatter steps are free.
The rules worth memorising, from Zapier's task usage documentation as summarised by DevToolLab and Zapier's own guide to minimising AI spend:
--- Free steps: triggers, Filter, Paths, Formatter and failed steps do not use tasks.
--- Normal actions: posting to Slack, adding a Google Sheets row or creating a CRM record is 1 task each.
--- AI by Zapier: Standard 1x, Advanced 3x, Premium 5x. Zapier says you can swap the model on a step at any time without rebuilding the Zap.
--- Bring your own key: 1 task in Zapier, but you then pay the AI provider for tokens separately.
--- Other AI features: a Zapier MCP tool call costs 2 tasks, and Zapier Agents use a separate activity quota.
So a simple Zap of trigger, one Premium AI step and three actions costs 8 tasks per run, not 4. Do that 2,000 times a month and you need 16,000 tasks.
How much does the Premium default actually cost you?
On DevToolLab's worked example, a support-ticket triage Zap running 2,000 times a month needs 16,000 tasks with a Premium AI step, which lands on the US$189 Professional tier. The same Zap on a Standard step or your own key needs 8,000 tasks and fits the US$129 tier. That is US$60 a month from one dropdown.
The numbers come from Zapier's Professional plan list prices, billed annually, as DevToolLab recorded them on 7 October 2026: US$19.99 for 750 tasks, US$49 for 2,000, US$89 for 5,000, US$129 for 10,000 and US$189 for 20,000. The free plan stops at 100 tasks a month and two-step Zaps.
At 20,000 runs a month the gap widens: 160,000 tasks on Premium (US$769) against 80,000 on Standard or your own key (US$489). For a Hong Kong team that is roughly HK$2,200 a month of difference, before you even ask whether the Premium model was needed.
The bigger lesson is that AI steps do not just add cost, they multiply it. Every extra AI step on Premium is five tasks per run, so a Zap with three AI steps on the default tier burns fifteen tasks before any normal action runs.
How do you cut Zapier AI step costs without losing quality?
Run a four-move audit on your busiest Zaps: count billed tasks per run, filter before the AI step, merge several AI steps into one structured call, then test a cheaper model tier on real samples before switching. Most Zaps lose a third or more of their task count, and you only keep Premium where a test proves it earns its cost.
Move 1: count before you change anything
Open Task History, set the range to last month, and sort Zaps by usage. For the top three, write down every step and its multiplier. You now know which Zap to fix first.
Move 2: filter before the AI step
Filters are free. If only half your form submissions are real leads, put a Filter step before the AI step so spam, test entries and existing customers never reach the model. Zapier's own advice is the same: if a decision follows fixed rules, use if/then logic, not an LLM.
Move 3: one AI step, several jobs
Many Zaps have one AI step to classify, another to summarise and a third to draft a reply. Ask one step to do all three and return labelled fields. Formatter steps, which are free, can then pick out each field for the actions that follow.
Move 4: test the cheaper tier on 20 real samples
Run 20 recent inputs through the step on Standard and on Premium, and compare the outputs side by side. If Standard gets 19 out of 20 acceptable, switch. If it fails on the cases that matter, keep Premium on that step only.
What does a single multi-job AI step prompt look like?
A multi-job prompt asks the model for several labelled outputs in one fixed format, so one AI step replaces two or three. The key is a strict output contract: exact field names, allowed values and a rule for uncertain cases. That keeps the result easy for free Formatter steps to split and stops downstream actions from breaking.
Here is the prompt I use in a support-inbox Zap. It replaces three Premium steps (15 tasks per run) with one step, and on Standard tier that is 1 task per run.
Try this prompt:
You are triaging one customer email for a small Hong Kong online shop.
EMAIL:
{{Email Body}}
Return exactly these five lines and nothing else:
CATEGORY: one of [order_status, refund, product_question, complaint, other]
PRIORITY: one of [high, normal, low]. High only if the customer mentions a payment problem, a legal threat or an order more than 7 days late.
LANGUAGE: one of [en, zh-Hant, zh-Hans]
SUMMARY: one sentence, under 25 words, in English.
DRAFT: a reply under 80 words, in the same language as the email, polite and specific. Do not promise refunds or dates.
If you are unsure of the category, use "other" and start SUMMARY with "UNSURE:".
Add a Formatter step after it with "Text, Split" or "Extract Pattern" to pull each line into its own field. Route high-priority tickets with Paths, which are free, and only the actions you actually need will use tasks.
Before you rebuild your own Zaps, you can also ask ChatGPT or Claude to do the counting for you:
Try this prompt:
Here are the steps of my Zap, in order: [paste step names and app actions].
It runs about [number] times a month.
Zapier rules: triggers, Filter, Paths, Formatter and failed steps are free; each other action is 1 task; AI by Zapier steps cost 1 (Standard), 3 (Advanced) or 5 (Premium) tasks; assume Premium unless I say otherwise.
1. Calculate tasks per run and per month.
2. Suggest which steps could be merged, filtered earlier or moved to a free step type.
3. Show the new monthly task count after your changes.
When should you stop tuning and switch platforms?
Switch when your Zaps are long, run thousands of times a month and are already tuned. Zapier bills every action step, so step-heavy AI workflows stay expensive. On DevToolLab's same example, Make cost US$16 and n8n Cloud US$20 at 2,000 runs, against US$129 to US$189 on Zapier. Migration still takes real time.
Be honest about the trade-offs before you move:
--- Zapier still wins on reach and ease. It connects 9,000-plus apps and is the fastest path from idea to working automation for a non-developer.
--- Make bills every module, including the trigger. A scenario polling every 15 minutes uses about 2,880 credits a month before it finds any data, so use instant webhooks.
--- n8n bills per whole run, so a 12-step agent costs the same as a 2-step one, but it has fewer ready-made integrations and a steeper learning curve.
--- Bring-your-own-key is not free. On every platform you then pay the AI provider for tokens on top.
--- Prices change often. Zapier's tier rules and all three price lists here were checked in early October 2026. Re-check before you budget.
If you want the full cross-platform cost picture, see our Zapier vs Make vs n8n cost breakdown.
Try it now: a 20-minute AI step audit
Pick your single busiest Zap, count its tasks per run, then apply the four moves in order: filter, merge, test the cheaper tier, switch. Twenty minutes is enough for one Zap, and the saving shows up in next month's Task History. Repeat on the next Zap only after the first one runs cleanly for a week.
--- Minutes 0 to 5: open Task History, find the top Zap, list every step with its multiplier.
--- Minutes 5 to 10: add a Filter before the AI step for anything that should never reach the model.
--- Minutes 10 to 15: merge your AI steps using the five-line prompt above, and split the output with Formatter.
--- Minutes 15 to 20: test 20 recent inputs on Standard. If quality holds, switch the tier and note the date.
The habit matters more than the single fix. Every time you add an AI step, ask two questions: does this need a model at all, and does it need the expensive one? We understand AI. We understand you better. With UD by your side, AI doesn't feel cold.
Reviewed by the UD AI team. Task rules and prices reflect Zapier's documentation and pricing as reported by DevToolLab on 6 and 7 October 2026, plus Zapier's guide to minimising AI spend; all figures are list prices in US dollars and may change. Related reading: how to build your first n8n agent.
Want automations that run without watching the meter?
Tuning Zaps saves money, but some teams are better served by an AI employee that handles the whole workflow. UD's AI Staff Solution covers customer service, admin and marketing roles for Hong Kong businesses, and we'll walk you through every step, from tool setup to workflow design and deployment.
