Convert a Zapier Agent Without a Task Bill Spike

You have one Zapier agent, and it already does a job. The migration question is whether the new AI step will bill like the old one or quietly multiply your usage.

Zapier announced on 6 October 2026 that Zapier Agents was becoming AI by Zapier. Zapier charges AI by Zapier steps using its task system, and an AI step can cost 1, 3, or 5 tasks depending on the model tier.

For paid plans, a newly created AI by Zapier step defaults to the Premium model tier and bills at a 5x task multiplier.

The work is a few minutes of setup before the first run, not hours of cleanup after the bill.

Set the model tier before the first bill

Start by asking whether the step needs Premium at all. A simple prompt can use a lower tier; test it. More reasoning earns the higher cost before the run, not after the bill.

Write the prompt like a brief to a careful employee: state the job, the allowed tools, and the expected output. Keep a cheaper tier when it can do the job. Note the reason when it cannot, so the next upgrade is a decision, not a guess.

Do not upgrade the tier to make the prompt feel safer. A clear job description is usually cheaper than a bigger model.

Cap the run and gate the risky tools

AI by Zapier steps have a default per-run task limit of 75 tasks, adjustable up to 500 tasks; a live run that reaches the limit pauses for human approval.

Set the limit to the amount of work you would accept in each run. A low limit is a control; raise it only after you understand why the step needs more work.

Treat the pause as the decision point: is this run doing the work you meant?

AI by Zapier lets you require approval before running on a per-tool basis, instead of using an all-or-nothing approval setting.

Use that for the tools that can create, send, delete, or charge. Let read-only tools run. The margin is in stopping the few actions that matter, not in approving every step.

A good approval rule is boring: catch the actions that are hard to undo, not the ones that are merely frequent.

Check the sources, then shut off the old agent

AI by Zapier does not yet work with HubSpot, Asana, Zendesk, Zoom, or Jira knowledge sources, with support planned for Q4 2026.

When your agent depends on any of them, do not assume the new step will see the same context. Wait, move that context into a supported source, or keep that part of the workflow manual.

The quiet failure is a confident answer built from missing context. Check the sources the old agent used, then compare them with what the new step can actually read. Missing context still lets the step run, but it runs with less context.

If the original agent is not turned off after publishing the migrated Zap, both the Zap and the agent can run at the same time and may cause duplicate actions.

Turn off the old agent only after the new workflow is live and you have watched it handle a real trigger. Do not leave it on as a backup. A backup that can act is a duplicate path: it can create the same email, item, or record while you watch the new step work.

Run the five checks in risk order

The known-cost AI step pre-flight follows risk order: cost, runaway, side effects, missing context, duplication.

• Choose the model tier before the step runs.
• Set the run cap to a number you can defend.
• Turn on approval for tools that can touch customers or money.
• Confirm every knowledge source the agent needs is available.
• Turn off the original agent after the migrated workflow is live.

Run the list in order. The model and limit checks come early because they shape the bill. The approval and source checks come next because they stop side effects and catch missing context. The shutdown comes last because it prevents duplicate paths.

Keep the list next to the workflow while you test it. The checks take little time and save cleanup work. If a check fails, stop and fix that layer before moving to the next. When the new step has handled a real trigger without a duplicate, a runaway, or a missing source, the migration is done.