Make and Zapier AI: features, agents, limits, and pricing

Make and Zapier let you add AI to an automation to classify, extract, or draft text, then trigger actions in connected applications. They also offer agents that can choose which tools to use, introducing more variability and risk. Start with a fixed workflow and a narrowly defined AI step, then measure output quality and usage before giving it more autonomy.

Traditional automation, an AI step, and an agent: three different uses

Traditional automation follows explicit rules: when a form arrives, it creates a record and sends a notification. You define the steps and conditions. If the incoming data is already organized into usable fields, adding a language model may simply create an extra expense.

A workflow with AI keeps that structure but assigns a specific task to the model. The model might summarize a message or suggest a category; your rules still determine what happens next. This separation makes troubleshooting easier because you can distinguish an extraction error from a routing error.

An agent receives a goal and a set of tools, then decides which tools to call based on what it finds. It might look up a record, search for missing information, and prepare an action. Its path can vary, making costs harder to estimate and every possible outcome harder to check.

Criterion	Traditional automation	Workflow with an AI step	Agent
How the path is chosen	Defined rules	Defined rules around the model	Model decisions within an authorized scope
Suitable data	Structured fields	Text or documents requiring interpretation	Requests that need different handling
Expected output	Predictable transformation	AI response requiring checks	Actions and response requiring checks
Budget	Based on steps executed	Steps and AI usage	Also depends on tools called
Suggested starting point	Simple synchronization	Classification or drafting	Research requiring several decisions

For a recurring task, first describe how it would work without an agent. Reserve autonomy for decisions that you cannot reasonably express through conditions.

In Make: AI modules and agents within a scenario

Make calls its automations scenarios. For a narrowly defined language task, Make AI Toolkit includes functions such as categorization, extraction, summarization, translation, and text standardization. According to the Make AI Toolkit documentation, reviewed on October 24, 2026, Make’s AI provider is available on all plans; connections to your own provider account require a paid plan.

The first practical decision is therefore which function you need. If you want to extract a reference number and a date, a dedicated step avoids giving the model tools it does not need. The scenario can then validate those fields before saving them.

For workflows with variable paths, the current application is listed as Make AI Agents (New). The official application reference, reviewed on the same date, still describes it as an open beta: features and pricing may change. It lets you add instructions, knowledge, and tools in the form of modules, scenarios, or MCP tools.

An agent could, for example, choose between checking a case file and querying a business system. Give each tool a narrow purpose and an explicit description. A tool that searches for an order should be separate from a tool that changes it.

The monitoring panel helps you inspect the calls the agent made. It does not establish whether the business decision was correct. You still need to verify that yourself.

In Zapier: AI by Zapier becomes the reference point

With Zapier, distinguish older product descriptions from the current integrated offering. The documentation for migrating Agents to AI by Zapier, updated on October 1, 2026, announces that agent capabilities are moving into the Zap editor. It does not yet set a shutdown date for Zapier Agents.

AI by Zapier brings together generation or classification steps and steps that can call tools. For a new project, review this setup before following a tutorial built around the older, separate service.

You still decide the architecture. You can ask the model to produce a category, then use a conventional step to route the request. You can also give it several tools and let it choose. In the latter case, inspect the available actions, connected accounts, and fields the model can populate.

Migration still requires a review. An automation transferred with its original instructions may retain excessive permissions or assumptions that no longer hold.

For a comparison beyond AI features, see the Make and Zapier comparison. The first check remains practical: does your software offer the exact action you need, with the right fields and permissions?

Choose based on the actual workflow and your team’s constraints

Connector counts alone will not tell you which platform fits. An integration may be able to create a record without being able to search for the record you need. Test the connector with the business records you use, your account type, and your actual permissions.

Also consider who will maintain the automation. That person should be able to understand an input, locate an error, and pause processing without depending on whoever built it.

Criterion	What to check in Make	What to check in Zapier
Building	How clearly the scenario and branches are presented	How clearly the Zap’s steps and conditions are presented
Targeted AI	Suitable module and connection type	AI by Zapier step and model tier
Agent	Available tools and call limits	Available tools and task usage
Troubleshooting	Scenario inputs, outputs, and calls	Zap history and step details
Budget	Fixed or variable credits	Tasks and AI multipliers
Operation	Connection ownership and recovery	Connection ownership and recovery

Check these points before building a complete workflow. Discovering a missing action late in the process may force you to add a technical workaround, changing the maintenance requirements you expected.

If your project calls for considering other architectures, the n8n and AI guide covers another approach. Include hosting and day-to-day operation in that comparison, alongside the visual editor. Our AI tool selection places these platforms in the broader software landscape.

Make pricing: understand credits before choosing a subscription

In Make, credits are the billing unit; operations describe execution. The official credits documentation, reviewed on October 24, 2026, distinguishes fixed usage for certain functions from variable usage associated particularly with AI.

According to that documentation, an operation in a conventional application without AI corresponds to a credit. When you use Make’s AI provider, usage can include both operations and model tokens. With a connection to your own provider account, you pay Make for execution credits and pay the provider directly for tokens.

That changes how you should read a budget estimate. An automation receiving long texts will not necessarily consume the same resources as one handling short messages, even if their diagrams look identical. Branches, tool calls, and the items being processed also affect execution.

The Make pricing page remains the place to check the price for your volume and subscription commitment. No numerical starting price is given here: the pricing record supplied for this page does not include one, and the review used for this guide does not provide a reliable basis for establishing it.

Before subscribing, record the plan, included credits, overage terms, and selected AI connection. Keep that configuration alongside your estimate. Without those details, a budget quickly becomes difficult to reproduce.

Zapier pricing: an AI step can consume several tasks

The Zapier pricing page, reviewed on October 24, 2026, lists Professional starting at approximately $19.99 per month. That starting price depends on your purchasing options, including volume and billing frequency. Check the order summary before paying.

For AI by Zapier, counting visible steps is not enough to calculate costs. The official model tier pricing table, reviewed on the same date, gives the following rules:

AI configuration	Tasks for the step	Tasks per successful tool call
Standard	1	Tools not supported
Advanced	3	3
Premium	5	5
Your own AI account	1	1, with a compatible model

Under this October 2026 pricing table, a Premium step that calls two tools therefore consumes 15 tasks. If you connect your own AI account, the provider’s bill comes on top of Zapier tasks.

The same documentation limits publishing AI by Zapier to the applicable paid plans. It also provides for pausing when a step reaches its per-run task limit. Check that setting before activation.

To estimate spending, account separately for the AI step, its tool calls, and the Zap’s other actions. Do not use the old Zapier Agents activity allowances to price an automation that has been migrated.

Build a first workflow: classify a request, then prepare a draft

Consider a teaching example, with no assumed performance results: a small business wants to classify incoming requests and prepare replies. Processing must stop before a message is sent.

The prerequisites are an authorized platform account, a test source, a place to save drafts, and a person responsible for approval. Use fictional or redacted data while building the workflow. Define the accepted categories and what should happen when none of them fits.

Define the trigger. Choose a dedicated folder or source instead of connecting the entire inbox.
Select the relevant data. Pass along the content needed for classification, without automatically including the entire message history.
Add a narrowly defined AI step. Ask for an allowed category, a summary, and a draft.
Validate the outputs. Reject an unknown category or a result missing a required field.
Prepare for review. Save the draft with a link to the original request.
Test recovery. Replay an event to check that it does not create a duplicate.

Here is an example instruction to adapt, with no guarantee of reliability:

Classez la demande parmi : commercial, assistance, administratif,
à examiner.

Utilisez « à examiner » si les informations sont insuffisantes.
Résumez uniquement les faits présents dans la demande.
Préparez un brouillon sans inventer de prix, de délai ni d'engagement.
N'envoyez aucun message.

The article on automating your inbox with an AI agent explores this use case further. The same method also applies to support tickets and forms.

Validate data before allowing an action

A polished response can contain a hallucination. Structured output makes a response easier to process, but does not prove that a customer exists or that an amount is correct. Validate the format and the content separately.

For extraction tasks, define required fields, permitted values, and information that must come from a business system. An order reference can be checked in that system before any changes are made. A recipient’s address should come from a verified field rather than a value freely generated by the model.

Provide a pending state for incomplete requests. Forcing the model to produce a usable answer in every situation can turn missing information into a silent error.

Human approval should apply to a specific action. Show the source, draft, recipient, and proposed changes. If the content changes after approval, it needs to be reviewed again.

For sensitive operations, place the final write action in a controlled step with fixed parameters wherever possible. AI can suggest a response; a rule should determine whether it is acceptable.

Finally, associate each run with the source event’s identifier. This helps prevent recovery after an incident from creating the same record or change repeatedly.

Protect connections and treat external content as data

An external message can contain instructions designed to redirect the model. This prompt injection becomes more dangerous when AI can read internal data and then communicate externally. A sentence in the instructions is not sufficient protection.

Reduce the capabilities available to the agent. An agent classifying requests does not need to delete records or export an entire database. Set permitted destinations and restrict connections to the accounts it needs.

MCP tools require the same attention. They provide an interface for accessing actions; they do not make those actions reliable on their own. The MCP server directory helps identify integrations, but you need to check the scope of each server you select.

Also list everywhere the data travels: the source application, automation platform, model provider, tools called, and logs. Review contractual terms and settings for your specific plan before processing sensitive information.

Logs may contain the text sent to the model and the responses it returned. Restrict access and avoid copying secrets into them. The guide to securing an AI agent covers permissions, secret management, and the approval checks to put in place.

Measure quality and cost before expanding

A successful test on a straightforward request is not enough. Build a representative sample that includes incomplete messages, unusable attachments, duplicates, ambiguous requests, and hostile instructions. Define the expected response before looking at the model’s output.

Measure the business outcome: correct category, accurate fields, acceptable draft, and no unauthorized action. A run can finish without a technical error and still classify a request incorrectly.

For budgeting, record usage for each path through the workflow. Distinguish a routine case, a case requiring additional research, and a failed case. Include review and correction time: an inexpensive subscription can become costly if requests need extensive rework.

You can use the following management formula, while keeping it separate from the platform’s billing rules:

Coût par dossier accepté =
(abonnement attribué + dépassements + facture IA + temps de contrôle)
divisé par le nombre de dossiers acceptés

Keep the instructions, configuration, and reference results. Run the sample again after changing the model, connector, or business rules.

Then start with a limited scope, a named owner, and a way to pause processing. Expand only when the observed errors and usage remain compatible with your requirements.

Common problems and checks before activation

The output does not work with the next step. Check the field being passed, the permitted values, and how missing information is handled. A longer instruction will not fix an incorrect connection between steps.

Spending exceeds the estimate. Examine tool calls, retries, and input length. In Make, check the AI connection type; in Zapier, check the model tier and its multiplier.

Retrying creates a duplicate. Look for the record associated with the event before repeating the action. After an incident, check whether the action succeeded in the connected application even if the automation platform received an error.

The agent uses an unexpected tool. Reduce the number of tools, clarify their descriptions, and fix sensitive parameters. If the path should always be identical, replace the agent’s choice with a conventional step.

Before going live:

The scope and responsible owner are identified.
Connected accounts have only the permissions they need.
Outputs are validated before any sensitive write action.
Incomplete requests and errors are routed for review.
Retries do not duplicate actions.
Actual usage and overage terms are understood.
Human approval and a way to pause processing are operational.

The goal is a process whose results you can explain, whose actions you can trace, and whose spending you can control. An agent becomes useful when its freedom to choose addresses an observed need and stays within that controlled scope.

Frequently asked questions

Can you use AI for free in Make and Zapier?

Make lets you use its AI provider on all plans, with usage charged in credits. For AI by Zapier, the free plan lets you preview a step in an unpublished Zap; publishing it requires a paid plan, according to documentation reviewed in October 2026. Free access to the platform therefore does not guarantee free AI automation in production.

Do you need an API key to automate tasks with AI?

Not always: built-in AI connections can handle access to the model. Connecting your own provider account lets you pay the provider directly, but adds another bill and a credential to manage. Also check whether your subscription supports that connection.

Does Zapier Agents still exist?

Official documentation reviewed in October 2026 announces its migration to AI by Zapier within the Zap editor. It does not yet specify a shutdown date for the older service. For a new automation, assess the integrated offering rather than relying on the old activity-based pricing.

Can Make or Zapier automate my entire inbox?

The scope depends on the actions available through the connector and your account permissions. Start with a test inbox or a dedicated folder, then check attachments, duplicates, and replies in existing threads. Evaluate drafting a message and sending it as separate actions.

How should you compare Make and Zapier costs?

Run the same sample through both platforms, with the same expected outcome. Record the credits or tasks consumed, any model charges, and the time spent correcting results. Comparing entry-level subscriptions alone hides the cost of actual workflows.

AI automation in Make and Zapier: features, agents, limits, and pricing