Agentic Layer Check

Zapier

Zapier runs two meters at once. Zaps bill in tasks, and only when an action succeeds. Agents bill in activities, a separate allowance on a separate product, where every action, every web page and every knowledge lookup counts. One of those meters has an overflow valve. The other stops dead. Here's how that reads against the nine layers.

Agentic Layer Check · Public signal read · October 2026

What they sell

The category defining automation platform: Zaps connecting thousands of apps, plus a separate Agents product, Chatbots, Tables, Forms and an MCP endpoint for connecting AI assistants to the same app library.

Who they sell to

Operators and business teams self serve from Free through Professional and Team, plus larger organisations on Enterprise, with Agents and Chatbots sold as separate subscriptions alongside whichever plan the account is on.

How they price

Tasks for Zaps: Free (100/mo), Professional (from $19.99/mo), Team (from $69/mo) and Enterprise, with tiers running to 2M tasks. Activities for Agents, on their own plan: 400/mo free, 1,500/mo on Pro at around $400 a year. AI steps, code, MCP and the SDK all draw from the task pool.

What's working

Zapier's original meter is still one of the most value aligned units in software, and it is worth saying why.

Layer 6 · Metering & Entitlement

You are billed for work that worked

Zapier's own help centre defines a task as "any successful action that runs in Zapier," and the exclusions are as telling as the definition: triggers don't count, polling doesn't count, Filters and Paths don't count, and steps that error, halt or never run don't count. Formatter, Delay, Looping, Digest, Storage, Tables and Forms are all free. The customer pays when something was actually delivered into another system, which is the closest thing this category has to a value metric, and Zapier has held that line for over a decade while competitors metered runs, seats and connectors.

Layer 8 · Revenue Architecture

A real overflow valve on the main meter

Hit the task limit with pay per task billing on, and usage keeps running: extra tasks are billed separately rather than the account stopping, with warning emails at 80% and 100% of a ceiling set at three times the plan limit. That is the expansion mechanism most usage priced vendors never build, and it solves the exact failure this series keeps finding elsewhere, where the billing stack has no way to say yes at the moment a customer wants to spend more.

Where there's room to extend it

The agent product didn't inherit any of that. It got a second meter, with different rules.

Layer 3 · Pricing & Packaging · Opportunity

Two meters, and the customer has to forecast both

Agents run on activities, billed as an add on subscription, and Zapier's own documentation states flatly that "using activities does not affect your Zapier task usage." So the same customer automating the same process now holds two units, two allowances and two renewal conversations, with no exchange rate between them. It gets stranger inside the Zap: AI steps, code, the SDK and MCP all draw from the task pool instead, and a single successful MCP tool call costs two tasks. Agentic work is priced inside the old unit in one product and outside it in the other, which makes "what will this cost us" a question even Zapier's own pricing page cannot answer in one number.

Layer 6 · Metering & Entitlement · Opportunity

On the agent meter, the agent spends the budget

An activity is "any billable action that your agent takes," and per Zapier's own help centre that includes every action in a multi step run, every record in a batch, every knowledge source lookup and every web page visit or search. The customer sets the intent; the agent decides how many activities reaching it costs. Zapier has published a help article specifically on unexpected activity usage, listing batch processing, knowledge lookups, web browsing, reprocessing old records and agent to Zap loops as the causes, and has fitted a per run cap of 10 activities on Free and 40 on paid plans that pauses the agent for input. Those are sensible guardrails on a unit the customer cannot forecast, which is the tell: the meter moved from counting delivered outcomes to counting the agent's own effort.

How this moves through the model

The sharpest cost sits in what happens at zero. On tasks, running out is a commercial event: pay per task keeps the work flowing and bills for it. On activities, running out is a stop. Per Zapier's documentation, agents stop running, cannot act, browse, search or receive triggers, and upgrading mid cycle raises the limit without resetting usage, unlike a Zapier plan. So the valve exists on the mature meter and not on the new one, which lands squarely on Revenue Architecture (Layer 8): the product most likely to produce a step change in customer value is the one with no mechanism to monetise that step change mid cycle. It lands again on Adoption & Customer Success (Layer 7), because a customer whose agents go dark mid month learns to run them conservatively, and a conservatively used agent never produces the result that justifies the next tier. Zapier built the best overflow valve in the category and then shipped its agent product without one. That is the pattern this series keeps finding: the capability exists, it just hasn't reached the layer where the new product needs it.

Part of the Agentic Layer Check series: named companies, read against the nine layers. Want your own read? Request a teardown →

Methodology & sourcing

This is a public signal read: Zapier's own pricing page, its own help centre documentation on task usage, Agents activity measurement and unexpected activity usage, not a private briefing or company input. It reflects what's publicly documented as of the date above: pricing pages and usage terms change, and this piece will note it if Zapier's do. Zapier was not consulted on or paid for inclusion; see the disclosure approach this work is run under.

Sources: Zapier: pricing → · Zapier: how task usage is measured → · Zapier: how Agents usage is measured → · Zapier: unexpected activity usage →