Zapier Agents Is Moving Into Zaps. Migrate the Controls Too
Zapier now converts Agents into Zaps. This migration plan shows teams how to preserve tools, set approvals, prevent duplicate runs, and control task usage.
Zapier's automatic conversion preserves an agent's prompt, tools, and trigger, but it does not preserve your operating judgment. Treat the move into AI by Zapier as a controlled cutover: test the whole Zap, require approval on sensitive actions, publish, then turn off the original agent.
The move is useful, but the cutover is still yours
The move is worth making for most internal workflows because it puts agent decisions inside the same place as normal Zap triggers, actions, filters, branches, and history. A product manager can keep deterministic routing fixed, let an AI step handle the ambiguous part, and inspect both in one run record.
AI by Zapier with looping tool calls became generally available on July 15, 2026. Zapier is converting the standalone Agents product into AI by Zapier steps inside Zaps. For Enterprise trial customers, the documented migration window runs from July 15 to August 15, 2026. On August 15, those trials expire and the accounts convert to a free Agents account.
That deadline is specifically for Enterprise trial customers. The broader decision is still urgent because a live agent should not be moved during a customer-facing incident or a busy operating window. Pick a quiet cutover period, name one owner, and define what a passing test looks like before anyone clicks Publish.
The conversion copies configuration, not risk decisions
The converted Zap is a starting configuration, not an approved production workflow. Zapier carries over the original prompt and instructions, connected tools, and trigger, then places them in a preconfigured AI by Zapier step.
The dangerous assumption is that a copied tool is also a reviewed tool. It is not. Require approval before running is off by default for new tools. When enabled, that setting pauses the run before the selected tool executes. Zapier recommends using it for actions that create, update, or delete data, while leaving low-risk reads uninterrupted.
Take a sales-operations agent that reads a CRM record, drafts a renewal email, and changes the account stage. The CRM lookup can run without a pause because it is read-only. The email send and stage update should require approval during the migration period. After you have reviewed enough run history to understand the failure modes, you can decide whether either write is safe to release.
Do not bury the boundary inside the prompt. A sentence such as "ask before changing a customer record" is an instruction to the model. The per-tool approval toggle is an execution control that pauses the action regardless of how the prompt is worded.
Cut over in a sequence that prevents duplicate actions
The safe cutover is inspect, test, publish, stop old. Zapier lets the original Agent and migrated Zap run independently until the Agent is disabled, so an untidy handover can send the same email, update the same record, or create the same task twice.
Inspect the copied configuration
Open the migrated AI by Zapier step and compare its prompt, trigger, connected accounts, tools, input fields, and output fields with the live Agent. For every tool, label the action as read-only or state-changing, then set the approval boundary.
Test the AI step and the whole Zap
Use Preview to inspect the structured output, the tools called, what each tool did, and the task count. Then run the complete Zap end to end with a reversible test record. Full-Zap tests currently consume tasks the same way production runs do, so include testing in the migration budget.
Publish the migrated Zap
Publish only after the downstream fields receive the expected values and every state-changing tool has the intended approval setting. Keep the first live trigger controlled if the workflow can send messages or edit customer data.
Stop the original Agent
Turn off the old Agent as soon as the migrated Zap passes its live check. Confirm the Agent is unpublished, then inspect Zap history for the new run rather than leaving both versions active as an informal fallback.
The fallback should be a deliberate rollback, not a permanent dual run. Record the old Agent's state before cutover, but reactivate it only after the migrated Zap is stopped and the duplicate-action risk is contained.
Recalculate the workflow before you publish it
Recalculate task use before migration because AI by Zapier charges by both model tier and successful tool calls. The model rate applies once to the AI step and again to every tool call it makes.
The official formula is:
Tasks used per run = (1 × model rate) + (number of tool calls × model rate)
An illustrative support-triage step on Advanced that looks up an account, reads an entitlement, and updates a queue uses 12 tasks per run: (1 × 3) + (3 × 3) = 12. That is the cost of the AI step plus three successful tool calls.
The right response is not automatically to choose the cheapest model. First remove tools the agent does not need, use fixed Zap steps for predictable transformations, and reserve the AI step for the decision that genuinely requires inference. Then use the Zapier task-cost calculation to forecast the workflow at expected volume.
Zapier also pauses an AI by Zapier step when it reaches 75 tasks in one run and asks for approval before continuing. Treat that as a last guardrail, not a target. A normal run that gets close to the limit is a signal to narrow the goal or split deterministic work out of the AI step.
Measure the tool count in Preview before projecting volume. The same prompt can consume a different number of tasks when the AI chooses a different tool path, so run history is the source for the operating estimate.
Three migration gaps can change the answer
Pause the migration only when the workflow depends on a behavior AI by Zapier does not carry over cleanly. Three gaps deserve a design decision before cutover.
Agent-to-agent calls need rewiring. A migrated child Agent must use a Sub-Zap by Zapier trigger, and every parent workflow that called it must point to the new Sub-Zap. Treat this as an interface change: verify the expected inputs, returned fields, error path, and connected account before publishing the parent.
Chat-triggered Agents need a new entry point. On-demand Agents started through a chat do not yet have a direct equivalent trigger in the Zap editor. Zapier documents a webhook trigger as the workaround. That is reasonable when Slack, a web app, or another interface can send a clean request. It is not a substitute for a product conversation layer that must authenticate a user, retain durable state, and recover gracefully.
Some legacy knowledge connections are not ready. HubSpot, Asana, Zendesk, Zoom, and Jira knowledge sources that were available in Agents are listed as coming to AI by Zapier in Q3 2026. AI by Zapier supports other knowledge sources, so the gap is specific rather than universal. If the migrated workflow depends on one of those sources, contact the account representative and do not replace live retrieval with a pasted snapshot that will go stale.
Migrate the workflow unless it is part of the product promise
Migrate internal, reversible operations; scope an owned build when the workflow is part of the product promise. The dividing line is not whether Zapier can connect the apps. It is who carries the consequence when the automation makes the wrong decision.
An internal lead-enrichment flow is a good migration candidate. A fixed trigger can collect the lead, deterministic steps can normalize fields, the AI step can classify intent, and a human can approve the CRM write. A failed run is visible, rerunnable, and operationally contained.
A customer onboarding agent is different when it controls entitlements, payments, account state, or a promised turnaround. That workflow needs product-level authentication, permission checks, durable state, recovery behavior, auditability, and an interface designed around the user's next action. Zapier may still handle a supporting notification, but it should not become the hidden foundation of the product simply because the first version was convenient.
Use this decision rule:
- Migrate when the workflow is internal, the actions are reversible, an operator owns exceptions, and Zap history gives enough evidence to diagnose a failure.
- Wait when a documented migration gap blocks a required source or trigger and the current Agent can stay controlled during the gap.
- Build when the workflow is customer-facing, differentiating, stateful, or too consequential for per-tool settings and task history to be the entire control model.
The move into Zaps is a better operating surface for automation. It is not a reason to outsource a core product boundary.
Questions teams are asking about the move
Do I have to rebuild my agent from scratch?
No. Zapier automatically converts the prompt, connected tools, and trigger into a Zap with an AI by Zapier step. You still need to inspect the configuration, run the preview and full test, publish the Zap, and turn off the original Agent.
What if my agent does not work the same way after migration?
Do not turn off the original Agent until the migrated Zap passes. Compare its output fields, tool calls, connected accounts, approval pauses, and task use in Zap history. If a copied capability is unsupported, use the documented workaround or contact Zapier support rather than silently changing the workflow's promise.
Can both my agent and my new Zap run at the same time?
Yes. They run independently until the original Agent is turned off. Use that overlap only for a tightly controlled live check because both versions can take duplicate actions.
What happens to my Enterprise trial if I do not migrate by August 15, 2026?
Zapier says the Enterprise trial expires and the account converts to a free Agents account. Migrate each required Agent before the deadline or contact the account representative if a documented feature gap blocks the move.
Scope Your MVP Sprint
Turn a product-critical agent workflow into a scoped, owned MVP with the controls its users and operators need.
Jul 25, 2026
