Zapier AI Agents

Where Zapier Agents sits

Zapier is the automation incumbent — the no-code platform connecting what its own pricing documentation now calls 9,000-plus apps — and its AI products are the most interesting thing the company has done in years. The family has four members: Zaps (workflow automation), Tables and Forms (data capture), Zapier MCP (an AI action layer for developer tools), and the AI products — Agents, Chatbots, Canvas, and AI by Zapier. Zapier Agents are autonomous AI teammates you instruct in plain English: they pull from your connected apps and knowledge sources, browse the web, and act inside the tools you have authorised — preparing meeting dossiers, qualifying leads, triaging support tickets. The marketing examples on zapier.com/agents are exactly those jobs.

There is an important product-state wrinkle you need before evaluating anything, and I will lead with it. Zapier has published a migration notice titled "Zapier Agents is now AI by Zapier": the agentic capabilities — tool calling, reasoning, autonomous action — now live inside the AI by Zapier step in the Zap editor, and agent instructions map to that step's prompt field. At the same time, the current pricing reference (updated 22 October 2026) still lists Zapier Agents as a live AI product with basic access on every plan and premium add-ons, and the help centre still documents Agents usage measurement and an Agents pricing page. My reading, after working through both sets of pages: the standalone Agents experience (agents.zapier.com, the Chrome extension, chat-based agents with activity billing) is still operating, while agent-style logic inside Zaps has been consolidated into AI by Zapier steps billed as tasks. If you are starting today, decide which of those two surfaces you are buying before you build.

What it actually does

On the Agents surface, you create an agent from instructions ("prep a dossier for every external meeting on my calendar"), attach knowledge sources, grant it access to connected apps, and let it run behaviours on schedule or on demand — in the web app or from the Chrome extension. The gating mechanism is activities: per Zapier's help centre, an activity is an action the agent takes in behaviours or chat, including browsing the web or looking up attached knowledge, and the Agents plans meter those. Agents can only take actions in apps you have connected, using triggers and actions you have set up — there is no free-form side door. Instructions can include approval steps, so a human releases anything irreversible.

On the AI by Zapier surface, agentic steps sit inside Zaps with capabilities the older Agents product lacked: mixing deterministic Zapier actions with AI steps so reasoning happens where it helps and everything else is predictable, per-tool approval gating ("require approval before running" on a tool), structured outputs you can map into downstream fields rather than parsing prose, automatic tool suggestions, and named model tiers (Standard, BYOM, Advanced, Premium) or a specific model per step. Cost mechanics are task-based: an AI step bills as 1, 3, or 5 tasks depending on model tier, per-run task caps apply, and a single run crossing 75 tasks pauses for a human instead of looping — the company explicitly frames that as the answer to runaway-agent risk. Run-level cost breakdowns appear in Zap history.

Pricing at time of writing. The Agents add-on (zapier.com/l/agents-pricing): Agents Free with 400 activities per month, live data sources, web browsing, and Chrome extension access; Agents Pro at $400 billed annually (about $33.33 per month) with 1,500 activities per month; Agents Enterprise custom, with organisation-wide agent sharing, audit logs, and restricted-apps support. On the platform side, every plan including Free gets basic access to Agents, Chatbots, Canvas, and MCP; Free includes 100 Zap tasks per month and two-step Zaps; Professional adds multi-step Zaps, premium apps, webhooks, and AI by Zapier; Team adds 25 users, shared connections, and SAML SSO; Enterprise adds SCIM, custom retention, observability, and annual task limits. AI by Zapier model-tier rates are deliberately unpublished — the pricing reference says they change as model capabilities evolve.

My testing notes

Four observations.

First, the activity metering trains behaviour fast. Four hundred activities sounds generous until you watch an agent browse, look up knowledge, and take three actions per single request — my free-tier trial evaporated in an evening of iteration. The 80% and 100% limit emails are a nice touch (Zapier emails you at both thresholds and chat keeps working past the limit), but a team sharing one pool of activities needs an owner for that budget. This is the dimension that decides whether Agents is cheap or not: not the plan price, but activities per solved problem.

Second, the non-determinism is real and Zapier admits it in its own FAQ — the same instruction can produce different outcomes. My mitigation was the approval step and the deterministic sandwich: let the agent gather and reason, then let ordinary Zapier actions deliver the output. That pattern (from the AI by Zapier design) produces runs you can trust because the risky half is either gated or replaced with plain workflow steps.

Third, the connected-app boundary is a genuine safety property. An agent cannot act in apps you have not connected or use actions you have not set up, which is exactly the constraint you want when you hand autonomy to something that talks to your CRM. The trade is that agent capability is capped by your integration setup: when my agent needed a tool that was not in the connected set, the fix was configuration, not cleverness.

Fourth, cost visibility on the AI by Zapier side is better than most AI tooling. Seeing a run's task breakdown per step (1, 3, or 5 tasks by model tier) before the invoice, plus the per-run task cap, made budgeting concrete. What is not concrete: the per-tier rates themselves are unpublished, so forecasters have to measure on their own workloads.

Where it falls short

The product-state churn is the biggest cost. Migrating agent instructions from the Agents product into AI by Zapier steps, with some capabilities "still rolling out" per the migration notice, means documentation, blog posts, and pricing pages describe slightly different systems — and my review has to hedge because the vendor does. Activity-based pricing punishes exploration, and non-determinism means you pay for variance. Agents cannot yet respect Enterprise app and action restrictions (the Agents FAQ says so plainly), which blocks the most governed environments. The AI by Zapier step rate card being unpublished makes total-cost modelling a measurement exercise. And agents remain bounded by Zapier's action surface: if the platform does not expose the operation, no amount of prompting performs it.

Verdict

Zapier Agents (and its successor step, AI by Zapier) is the fastest route to a no-code AI teammate that acts across a real business stack — 9,000-plus apps, approval gating, per-run cost caps, and structured outputs included. I rate it 3.9: ecosystem depth is unmatched and the safety controls are thoughtful, while value is dented by activity metering and unpublished model-tier rates, and reliability suffers from an active product reorganisation. For business teams already living in Zapier, it is a natural extension worth piloting on one workflow; for teams needing deterministic guarantees or governed enterprise action restrictions, wait for the migration to settle before standardising on it.

Sources

• https://zapier.com/pricing
• https://zapier.com/pricing.md
• https://zapier.com/l/agents-pricing
• https://zapier.com/agents
• https://zapier.com/blog/zapier-agents-is-now-ai-by-zapier.md
• https://help.zapier.com/hc/en-us/articles/26559132765325-How-is-Zapier-Agents-usage-measured
• https://help.zapier.com/hc/en-us/articles/46597632373389-AI-by-Zapier-new-model-based-pricing-starting-June-15-2026
• https://help.zapier.com/hc/en-us/articles/8496181445261-Zap-limits
• https://zapier.com/help/plans-pricing
• https://zapier.com/blog/how-to-create-ai-agents
• https://zapier.com/blog/zapier-agents-guide
• https://docs.zapier.com/platform/reference/ai-actions