In a CIO demo published April 1, 2026, Zapier’s Copilot turns a spoken instruction into a workflow that takes a Google Forms sales-interest submission, asks AI by Zapier to draft a personalized email, waits five minutes, and sends it through Gmail. The example shows how an AI prompt can become part of a connected business process—but it is a vendor executive’s demonstration, not independent proof that generated workflows are consistently reliable or can run without review.
What the Zapier demo builds
On CIO’s 24-minute DEMO episode, host Keith Shaw speaks with Philip Lakin, Zapier’s Director of AI Transformation. Lakin describes a workflow for following up with people who submit a sales-interest form. He says Zapier connects more than 8,000 apps; that is his statement in the 2026 interview, not an independently verified current count.
- Trigger: A new response arrives through Google Forms.
- AI step: The form data goes to AI by Zapier, using the ChatGPT 5 mini model named in the episode, to draft a personalized subject line and email.
- Delay: The workflow waits five minutes.
- Action: Gmail sends the message.
Lakin gives Copilot a spoken, natural-language instruction. Copilot creates a visible sequence of workflow steps and automatically maps fields. During review, Lakin spots and corrects a field mapping, tests with records, previews the draft, removes an AI-generated signature so Gmail can handle it, and tests delivery. The model name reflects what the episode demonstrated in April 2026; model availability can change.
How to turn an AI prompt into a workflow
The demonstration’s practical lesson is to describe a process with a clear starting event, the AI task, and a defined action—not just ask for an email in a chat window. In this example, the requested outcome is a follow-up message based on form data, sent through Gmail after a short delay.
#1 Best Overall
- State the trigger and outcome: Specify that a new sales-interest form response should start the process and lead to a follow-up email.
- Define the AI’s bounded task: Ask it to draft a subject line and message personalized with the submitted information.
- Set the surrounding steps: Include the five-minute delay and the Gmail send action shown in the demo.
- Inspect the generated workflow: Check that each form field feeds the intended AI input and email field. Lakin’s correction of a mapping shows why automatic setup still needs review.
- Test before relying on it: Use test records, preview the message, check details such as signatures, and verify delivery before enabling the workflow for real submissions.
This sequence reflects the episode, not a guarantee that Copilot will interpret every instruction correctly. The demo provides no measured reliability, time-saving, or comparative performance result.
Is this an AI agent or an automation with an AI step?
Lakin explicitly calls the example “a deterministic workflow with an AI step, not a full AI agent.” The distinction is about how the process is controlled: the form trigger, delay, and email action are defined steps, while AI generates the message within that structure. Lakin contrasts this with agents that rely more on inference and offer less control. That is his explanation of the design, not a universal finding established by a head-to-head test.
| Concept | How work is directed | Where review fits | What the episode establishes |
|---|---|---|---|
| Structured workflow with an AI step | Actions such as the form trigger, delay, and email send are set as explicit steps; AI handles a bounded task such as drafting. | A person can inspect mappings and AI output before testing or relying on the process. | The demo illustrates this approach; it does not measure reliability. |
| More autonomous AI agent | Lakin describes agents as more inference-driven and less controlled by explicitly defined steps. | The episode offers no detailed review procedure or evaluated agent example. | This is Lakin’s distinction, not a comparative product test. |
Does a person review the generated email?
Yes, in the demonstration. Lakin checks test records, fixes a mapping, reviews the generated email, removes the generated signature, and tests sending. He sums up the approach: “This is human-in-the-loop. AI gives a strong first draft, but it’s still my job to review it.” The episode shows those review steps; it does not establish that every Zapier workflow requires the same process or that every AI-generated message will be correct.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.What problem does the workflow address?
The concrete alternative in the episode is to connect services into a process rather than wait for individual vendors to add a desired AI feature or handle the work manually. Here, the form, AI drafting step, delay, and email action are joined in one sequence. That illustrates a way to move AI output into an existing business process; the interview does not quantify how much work it saves or show that this approach suits every company.
Free tools Windows power users keep installed
One-click scans. No signup required.
Lakin also recommends splitting agent work into smaller tasks. He compares editing five pages with editing 100 and argues that focused responsibilities work better. Treat that as practitioner guidance from the episode, not a measured performance comparison.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
