Workbench and Zapier
Workbench or Zapier?
Zapier connects thousands of apps, and now lets an AI step choose which of them to act on. Workbench is where your team designs that AI step, tests it with the people who know the work, and ships a version your product can count on.
New claim → triage → Slack
Zap · On · Version 7
When a new email arrives in Claims, AI by Zapier sorts it into one of four kinds.
Paths send each kind to its own Slack channel.
Prompt last edited Tuesday. No record of why.
Claim triage
- Classifier with four kindsLive
- Tried on 40 recent claims92%
- Claims lead reacted by link3 changes kept
- Called by the ZapLatest published
Every edit says what changed and what the tests said
Where the two part ways.
- 1
The AI step gets its own editor
Workbench opens an agent, a classifier, a judge or a structured extractor in an editor built for that job, with the instructions, the model and real examples side by side. In Zapier the AI step is one step in a Zap.
- 2
Tested before it ships
Caliper runs every version against your examples and rubrics, and shows the score when you publish. Zapier doesn't advertise test sets or scoring for its AI steps.
- 3
The people who know the work shape it
Send the claims lead a link to the draft, and their thumbs and comments become edits. Each edit reruns your tests and is undone if the score drops.
- 4
Built to be called from your product
Your code calls a pinned, published version over the API and gets back the answer, the cost and each step's timeline. Edits to the draft never reach it.
- 5
Processes that take days
Write a Task as a sentence, and Workbench turns it into steps, asks named people to approve or edit, escalates when they don't answer, and records the outcome in Ledger.
When Workbench is the right choice
- The AI's judgment is the hard part, and a wrong answer costs you.
- You need to prove a change made it better before it goes live.
- Your product calls the AI and needs a version that stays put.
- The process runs for days and needs named people to sign off.
Job by job
Building the AI step
Zapier: An AI by Zapier step with a prompt, a model tier and tools.
Workbench: An editor for each kind of job, with examples to try beside it.
Testing
Zapier: Zapier doesn't advertise test sets or scoring for AI steps.
Workbench: Caliper tests every version, and the score shows when you publish.
Versions
Zapier: Drafts, a new version on every publish, and rollback on some plans.
Workbench: Numbered versions your product pins, with the draft kept apart.
Approvals
Zapier: Per-tool approvals and a Request Approval step that pauses a Zap.
Workbench: Tasks with named approvers, escalation, and a record in Ledger.
Starting a run
Zapier: An event in any of 9,000+ apps, a schedule or a webhook.
Workbench: The API, a webhook link, a schedule, Slack, text, or a Napkin button.
Tables and pages
Zapier: Zapier Tables and Interfaces.
Workbench: Napkin tables with grid, board and form views, and pages people can click.
Keep the Zap, and give it a better AI step.
Zapier keeps listening to your apps, and hands the judgment to a Workbench flow that your experts shaped and Caliper tested.
- Add a Webhooks by Zapier step that posts to the flow's webhook link, and use the answer in later steps.
- Connect Zapier's MCP server to your workspace, and a Workbench agent can use the actions it offers.
- Publish a new version in Workbench when the tests improve, and the Zap uses it without an edit.
Questions
- Can a Zap call a Workbench flow?
- Yes. Post JSON to the flow's webhook link, or call the API with a key.
- How does a Workbench agent reach other apps?
- An agent uses the MCP servers your workspace connects, plus HTTP requests.
- Who builds a flow?
- Your team builds in the editors and on the canvas. Your engineers write code when your own product calls a flow.
Try Workbench with your domain.
Bring the Zap whose AI step you trust least, and start there.
No website? Describe what you do
If you'd like help with the first changes, talk to us.