Zapier Agents: Now an AI Step in Zaps, Priced

Zapier Agents are AI teammates you configure with instructions, tools and a trigger so they act across Zapier's 9,000+ apps. Zapier now says the capability lives inside one AI step in the Zap editor, called AI by Zapier, billed in tasks. The older model gave Free 400 and Pro 1,500 activities a month. Checked 31 October 2026.

At a glance

What changed

Zapier's own blog post is titled "Zapier Agents is now AI by Zapier." The key statements:

• Agent capabilities "now lives inside a single AI step in the Zap editor." Agents used to sit outside the main platform. As a step, they can use Zapier triggers, filters, paths and run history.
• Existing agents are recreated as AI by Zapier steps with the same goal and tools. Eligible agents show a banner linking to a step-by-step migration guide.
• The parts map across: instructions go in the prompt field, tools go in the tools picker, and the agent's trigger becomes the Zap trigger. A chat or on-demand agent can be mimicked by triggering from a webhook, since some capabilities are still rolling out.
• Approval is now set per tool with "Require approval before running", where agents once had all-or-nothing approval.
• You can pick a model tier (Standard, Advanced or Premium) or name a specific model.
• Run history moves to Zap History.

The post does not give a date for the change. We could not confirm whether the standalone agents product is being retired.

Billing changes with it. The post says agent cost moves to tasks, "a universal task billing system" replacing the activity approach. Default models are free, and other tiers cost 1, 3 or 5 tasks per AI step. Zap History shows a per-run task breakdown. You can cap tasks per run, and a run crossing 75 tasks pauses for a human. Knowledge sources can still be used, except HubSpot, Asana, Zendesk, Zoom and Jira, which are not supported yet.

So there are two descriptions of one product at once: the agents page and the usage help article describe activities, and the new blog post describes tasks. Which applies to you depends on whether your account has migrated. Read the banner in your own account.

What it does

Zapier's agents page describes agents as AI "teammates" that take on work, using company knowledge and acting across its integrations "on command and while you sleep." The lifecycle it lists is build, monitor activity, chat when needed and work on the web, and a Chrome extension is mentioned.

From the help centre, an agent has these parts:

• Instructions. A prompt that says what the agent is for.
• Tools. Actions you add under "Tools this agent can use". These are Zapier actions in your connected apps.
• Knowledge. Sources such as FAQs, documents and public links, and live sources such as Google Drive, Notion or Asana that the agent can search and summarise.
• A trigger. On Demand (run it or chat with it), Schedule by Zapier, a trigger from a Zap or an MCP server, or an app event such as a new email. Publishing activates the trigger; until then the agent runs only during testing.
• Agent-to-agent calls. An agent can be called from another with a "Call an agent" action; the help centre advises fewer than 5 calls per main agent for performance.

Template examples on the page include Meeting Prep, Enterprise Lead Qualification, a Zendesk ticket responder, Lead Enrichment, an IT helpdesk Slack responder, Support Email, Candidate Ranking, a GitHub PR Slack notifier, a Sales Email Writer, an SEO Blog Writer and an Expense Classifier. The page does not explain how agents differ from Zaps beyond a "Next Gen Zaps" early-access banner, and it does not mention MCP in the text we read, though the help centre does mention MCP calls as a trigger.

Plans and prices

Zapier's pricing page, 31 October 2026, in US dollars per month:

Professional tiers run from 750 tasks up to 2,000,000 tasks ($3,389 a month annual, $5,099 monthly); Team tiers from 2,000 up to 2,000,000 ($3,999 annual, $5,999 monthly). Volumes above two million go through sales.

For agents specifically, the pricing page says:

• Free for everyone; the Free plan includes 400 activities a month. Pro allowance is 1,500 activities.
• Pro starts at about $33.33 a month on an annual basis for Agents; the page does not list a Pro monthly price for Agents.
• Per-run cap: 10 activities on Free, 40 on paid plans; at the cap the agent asks before continuing.
• Agent activities are a separate unit and do not affect your task usage.
• On Team, activities are shared across the whole account. Enterprise has a custom allowance.

The help article adds detail. What counts as an activity: a trigger firing, a knowledge source lookup, an action, a web browse or scrape, a web search and a Chrome extension message. Testing counts on Free and is free on paid plans. When you hit the limit, agents stop acting, but you can still chat. Emails come at 80 percent and at the limit, and the limit resets on your billing date. Free users cannot share agents.

Under AI by Zapier these activity numbers are replaced by task costs, so check your account's usage page for the unit that applies to you. Zapier's pricing page still lists 400 and 1,500 activities a month for its agents tiers, and we could not confirm whether those allowances apply after migration.

Data and privacy

The pages we read do not state whether Zapier trains models on your data, nor how long it keeps agent activity. Check Zapier's legal pages for its data-use, training and retention terms before you rely on it. An agent holds connections to your apps, so review what each tool can read and write before you publish. Use the per-tool approval setting for anything that sends messages or changes records, and give the agent's connected accounts the narrowest permissions your apps allow.

Designing a safe first agent

Take a common job: triage incoming support emails. A sensible first version, using the parts above:

1. Trigger. An app event: a new email in the support inbox.
2. Instructions. Two short paragraphs: classify the email as billing, bug or question; draft a reply; never promise refunds.
3. Tools. Read access to the help centre knowledge source, a tool to add a note to the ticket, and a tool to create a draft reply. Leave out any tool that sends email directly.
4. Approval. Require approval before running the draft tool at first, so a person sees every draft.
5. Cap. Set a low task or activity cap per run, so a confused agent stops early.
6. Test with ten past emails, including a rude one and one in another language.
7. Publish, then read the run history daily for a week.

The principle is to start with the agent suggesting and a person acting, then loosen one control at a time as the history earns trust. The same pattern appears in our guide to building a support triage agent and the AI customer support overview.

Watch out for loops. An agent that replies to an inbox it also reads can answer its own messages. Use filters so the trigger ignores your own address, and keep the cap in place. Add a rule that sensitive categories, such as legal threats or account closures, always go to a person.

What it does well

• Reach. The largest app library of the platforms we cover.
• Triggers. Schedule, event, webhook and call-from-another-agent.
• A real free tier, with 400 activities and 100 tasks.
• Per-tool approval, so risky tools wait for a person.
• One place. In the new model the agent is a step in a Zap with history, filters and paths.

Limits and cautions

• Two billing units in flight. Activities and tasks both appear in Zapier's own pages.
• Cloud only. No self-hosting, unlike n8n.
• Caps per run. 10 activities on Free, 40 on paid, and a 75-task pause under the new model, so long jobs need design.
• Cost scales with volume. The Professional ladder rises quickly past 20,000 tasks.
• Unsupported knowledge sources for HubSpot, Asana, Zendesk, Zoom and Jira in the new model, per the blog.
• Black-box reasoning. An agent chooses tools itself; check run history often.

Zapier versus the others

The wider decision is in AI agent platforms, and Zapier MCP covers using Zapier's actions from a coding agent.

Who should pick it

Pick Zapier Agents (AI by Zapier) if your work lives in SaaS apps, your team already uses Zapier, and you want an agent that reads a ticket, looks something up and writes back, with a human approval where it matters. It is a weaker fit if you need self-hosting, strict cost predictability at high volume, or an agent that runs for hours inside a codebase.

Questions to ask before you migrate an agent

1. Which of my agents show the migration banner, and which do not?
2. What do my agents cost per run today in activities, and what will the same run cost in tasks?
3. Which knowledge sources do they use? Check the unsupported list (HubSpot, Asana, Zendesk, Zoom, Jira).
4. Do any rely on on-demand chat? That mode is being mimicked through a webhook for now.
5. Where will run history live afterwards, and do I need to export the old one?

Migrate one low-risk agent first and compare outputs for a week before moving the rest.

Keep a log of the first month: the agent, its task, cost in the unit your account uses, and any bad outputs. It is the data you need to decide whether to widen the agent's permissions.

Our take

This is an editorial reading of Zapier's pages, not a hands-on test.

The change to AI by Zapier is sensible. An agent was always one step in a larger process, and putting it in the Zap editor gives it triggers, filters, history and per-tool approval that the standalone product lacked. The cost is a period of mixed messaging, with activities in some pages and tasks in others.

If you are starting, build inside the Zap editor with an AI step, set per-tool approval on anything that writes, cap tasks per run, and watch the first week of Zap History for the task count per run. If you are on the older agents product, expect to migrate. Either way, keep the prompt and tool list in a document outside Zapier so a billing or product change costs you wiring, not thinking.

Checklist

• Check whether your account shows the AI by Zapier migration banner
• Write the goal, tools and trigger in one paragraph before building
• Turn on "Require approval before running" for tools that send or change data
• Set a task cap per run
• Review Zap History after the first ten runs
• Confirm data terms before connecting sensitive apps
• Re-check pricing on zapier.com before upgrading