ChatGPT Pages, Agents, and Dots: Three Ways to Automate Work

A shared document, a developer-built agent, and an always-on assistant can all look like "ChatGPT automation" from the outside. In Rob The AI Guy's 15-minute walkthrough, they do different jobs. He builds a sponsorship inbox in ChatGPT Pages, sketches a YouTube reporting agent on the API platform, and asks his Dot to create a Google Sheets analytics report through Zapier MCP. The useful lesson is choosing the smallest workflow that can produce a result you can inspect.

Watch Rob's ChatGPT Automation Walkthrough

Credit: Rob The AI Guy for the demonstration and examples, published 5 October 2026. The Zapier portion includes repeated sign-up calls to action; no sponsorship terms are established by the supplied transcript. Product availability and pricing below were checked against OpenAI and Zapier documentation on 11 October 2026.

Three Surfaces, Three Jobs

This distinction matters because a Page is not an autonomous worker, and an API agent is not included simply because you pay for a ChatGPT subscription. OpenAI describes Pages as editable documents in Space, the Agents API as developer infrastructure, and a Dot as an ongoing personal agent.

Pages: Give Repeated Work a Reviewable Home

Rob first shows a Page that collects AI news, source links, social-post ideas, drafts, and published results. A second Page summarizes his calendar and Gmail. He then creates a sponsorship-triage Page and asks ChatGPT to identify outstanding replies, follow-ups, and payments from connected mail. The result is a queue with contacts and next actions. His earlier attempt to save a news Page needed an extra "configure it" instruction, which is a reminder to inspect the result rather than treating the first generation as a successful setup.

OpenAI's Space guide confirms that Pages support writing, charts, comments, subpages, and sharing. It also says the separate Keep updated automatic-update feature was unavailable at launch. Rob demonstrates schedules associated with his work, but a reader should verify the actual task instructions, frequency, source permissions, and latest run. A Page does not maintain itself merely because it exists.

For sponsorship mail, keep the output to "draft and flag" until someone checks the sender, terms, due date, and amount. Sharing a Page can expose information copied into it, even when viewers cannot open the original email. Review inherited page and Space permissions before inviting collaborators.

Agents API: Build a Workflow You Own

At 9:16, Rob opens the developer platform, explores an invoice-review template, then generates instructions for an agent that would check YouTube analytics and write to Google Sheets. He adds tools, considers subagents, and opens a session to test it. The recording shows configuration, not a production-ready daily agent or a validated invoice-control system.

OpenAI's Agents API describes agents, environments, sessions, events, and MCP connections. It also states that model, tool, and hosted-sandbox usage are billed at API rates. If you build this for a client, define the data contract first: which account, which metrics, which date range, what a missing field means, and how a failed run appears. Test with a read-only connection before granting write access to a Sheet. The older visual Agent Builder is scheduled for deprecation; do not base a new long-lived delivery promise on that older product.

Dots: Delegate an Ongoing Task, Then Inspect It

At 12:16, Rob asks his Dot to pull 30 days of YouTube analytics through Zapier and create a Google Sheet. The Dot asks which of two connected channels to use, later sends progress messages, and eventually supplies a Sheet with daily data, top videos, charts, and takeaways. When he first asks for the link, the Sheet is not ready yet; he receives it only after another update. That makes the finished artifact stronger evidence than the intermediate "I pulled it" message.

Rob says he wants to approve the report and then schedule it daily. The transcript shows the report, not a confirmed saved daily schedule. Check the Dot's Scheduled view and task history before relying on repeat runs. OpenAI says a Dot can work in the background, use connected apps, and ask for input, but can make mistakes. Its initial personal Pro rollout excludes the European Economic Area, Switzerland, and the UK; Business Premium and administrator-enabled Enterprise beta have different rules. Availability is gradual.

Zapier MCP: More App Choices, Not Blanket Access

Rob's bridge is Zapier MCP. Zapier advertises connections to more than 9,000 apps. That is a catalog, not 9,000 automatically usable accounts: you select tools, connect each relevant account, and grant the actions you want the AI client to run. Rob demonstrates creating a custom MCP connection in ChatGPT and also finding Zapier on the developer platform. The exact menu and availability can differ by account and workspace.

1. Choose one workflow and one account. For a first test, read one analytics source and write to a test Sheet, or read mail and create a draft queue. Avoid an all-apps connection.
2. Configure only the actions required. Zapier's setup guide recommends enabling specific actions. Read and draft are safer first steps than send, delete, refund, or publish.
3. Connect and test in the intended client. Follow Zapier's current quickstart and your ChatGPT workspace's MCP permissions. Do not assume a plugin that works in ChatGPT will automatically work in every Page, Dot, or API session.
4. Review the run and quota. Check the resulting source links, records, and Zapier action history; then set a repeat schedule only if the first output is correct.

Cost reality: Zapier MCP is included on its plans, including Free, but one MCP tool call consumes two Zapier tasks. The currently listed Free plan includes 100 tasks per month, shared with other use. A workflow making five MCP calls per run would use ten tasks before any other metered work; daily use would exceed that free allowance. ChatGPT plan eligibility and any Agents API charges are separate. Treat Rob's "for free" as possible to start on a free tier, not unlimited operation.

A One-Week Pilot That Can Actually Be Judged

Pick one recurring job: sponsorship follow-up or a YouTube performance brief. For the first week, use a narrow read connection and a Page or Sheet where every row has a source, date, proposed action, owner, and status. Compare the AI queue with the original records for three runs. Count missing items, incorrect facts, time saved after review, and tool calls consumed. Only then consider a schedule. Keep outbound mail, account edits, payments, and publishing under explicit approval. If the assistant cannot reliably identify the right channel or source, fix the input contract before adding autonomy.

Turn This Into a Business Idea

The sellable part is not "an agent connected to everything." It is a reviewable queue or briefing that removes a recurring bottleneck for a named buyer. This prompt works in any capable AI assistant and has manual alternatives if the launch features are unavailable.

Video Chapters

Sources and Useful Links

• Creator demonstration: Rob The AI Guy: ChatGPT's New Automation Features Are INSANE! (video and supplied transcript).
• OpenAI product guides: Space and Pages, Dots, schedules, and controls, and Agents API and billing.
• Integration and pricing: Zapier MCP, quickstart, action-scoping guide, and current plan limits.