AI Workflow Automation, DIY: Sheets, Gmail and Power Automate

AI workflow automation means adding an AI step (classify, summarize, extract or draft) to a workflow that already moves data from one place to another. The rest of the workflow stays rule-based; the AI only handles the part that used to need a person to read. You can start with tools you already have: a Google Form feeding a Sheet with an AI column, Gemini in Gmail for drafts, or a Power Automate flow with a Run a prompt action. The DIY route works until the AI step has to fire on every new item, by itself, and act on the answer.

What is AI workflow automation, in plain terms

A classic workflow follows fixed rules: when a form arrives, add a row; when a row is added, send an email. It breaks the moment a step needs judgment. Is this message a complaint or a sales lead? What did the customer actually ask for? AI fills that gap. It reads free text and returns something a rule can use: a category, a short summary, extracted fields, or a draft reply.

So the method always has three parts. A trigger brings the item in. An AI step turns messy text into a clean answer. A rule acts on that answer by routing, logging or notifying. The AI does not run the business. It answers one narrow question, many times. Where a wrong answer costs money or reaches a customer, a person reviews the output first. Our guide to AI automation covers the wider picture; this one is about doing it yourself.

Step 1: Pick one job and write the prompt like a rule

Before opening any tool, choose a job that repeats daily and needs reading, not deciding. Good first jobs: tagging inbound inquiries, summarizing long feedback, pulling an order number out of an email, or drafting a first reply that a person edits. Poor first jobs: anything that commits money, approves a refund, or speaks for the company without review.

Then write the prompt the way you would write a rule for a new hire.

1. Name the input: the customer message in column C.
2. List the allowed outputs, and nothing else: Quote request, Complaint, Booking change, Other.
3. Say what to return when unsure: Other.
4. Ask for the label alone, with no explanation.
5. Keep ten real past examples, with the right answer next to each, to test against.

The closed list of labels is the most important line. Free-form answers look smart and break every rule downstream, because no filter can match this seems to be a complaint about timing. Test the prompt on your ten examples before it touches live data. If two of them come back wrong, tighten the label definitions first.

Add AI to a workflow in Google Sheets

The simplest Google pipeline is a form that writes into a sheet, plus the Sheets AI function. Google Forms can send every response to a linked spreadsheet, and Sheets automatically puts that data in a table ^[1].

1. Open your form in Google Forms and click Responses, then Summary.
2. Click More, then Select destination for responses, choose Create a new spreadsheet, and click Create.
3. Open the sheet. At the top right of the last column in the table, click Insert AI column right.
4. In the first row of the new column, enter your prompt with the cell it reads, for example =AI("Categorize the inquiry as Quote request, Complaint, Booking change or Other.", C2).
5. Autofill the column, select the cells, and click Generate and Insert.

The AI function needs an eligible Google Workspace or Google AI plan, or a seat in Google's Workspace Experiments program ^[2]. It can generate text, summarize, categorize (sentiment included) and pull current information from Google Search. You can type =Gemini() instead of =AI(); both work.

Read Google's limits before you build on it. Responses are text only. The function sees only the range you hand it, not the whole spreadsheet or your Drive. It cannot sit inside another formula, so =IF(AI(...)) is not supported. Only the first 350 selected cells generate at once. Hit a long-term limit and you wait 24 hours.

AI email automation with Gemini in Gmail

Email is where most small teams want AI first. Gmail has Gemini features built in, but availability depends on the account. Help me write drafts a new email or refines your text from a prompt. On a personal Gmail account without a paid plan it works in the US only, while eligible Google AI and Workspace plans get it worldwide, in the same eight languages ^[3]. On an eligible Workspace plan you can also ask Gemini in Gmail to summarize a thread, draft, and suggest responses.

These are assistants, not a workflow. Each one starts when you open a thread or type a prompt, and nothing flows into a sheet or a CRM afterward. That still has real value. A practical split: let Gmail summarize and draft while you read, and leave the sorting of new messages to a rule-based tool. If you only need help writing faster, stop here. If every inquiry must be tagged and logged without anyone opening it, read on.

Build an automatic AI step in Power Automate

Microsoft 365 teams have a built-in way to run an AI step inside a real flow. Power Automate's Run a prompt action (called Create text with GPT using a prompt before May 2025) runs an AI Builder prompt as one step. Microsoft says these prompts run on GPT models powered by Azure OpenAI Service, the capability is limited to some regions, and it may face usage limits or throttling ^[4].

1. Sign in to Power Automate and select Create, then Instant cloud flow.
2. Select Manually trigger a flow, then Create.
3. In the designer, select + after the trigger and choose the action Run a prompt.
4. In the Prompt field, pick an existing prompt or New custom prompt, then fill its inputs with content from earlier steps.
5. Add a step after it, such as Post message in a chat or channel, and place the prompt's Text output in the message body.
6. Save, select Test, enter sample input values, and select Run flow.

Microsoft's own walkthrough also adds a human check right after the prompt: select +, search for Approvals, and add an approval step. Use it. That puts a person between the AI's answer and your customer. The manual trigger is only for testing. Once the prompt behaves on your ten examples, the same action can follow whatever trigger your flow needs, with earlier steps feeding its inputs.

A worked example: one inbox, four labels

Picture a three-person landscaping company. Inquiries arrive through a web form, about thirty a week. The owner reads each one, decides what it is, and forwards it. Here is the DIY version.

The form writes into a linked sheet. An AI column labels each row as Quote request, Complaint, Booking change or Other, and a second AI column writes a one-sentence summary. Once a day, the office manager selects the new rows, clicks Generate and Insert, sorts by label, and handles complaints first. Replies start from a Help me write draft that she edits before sending.

The result is real. Nobody reads every message twice, and complaints stop hiding under quote requests. But notice the click. Every step still waits for a person to open the sheet.

Where DIY AI workflow automation stops

The built-in route breaks at three specific points.

It does not fire on its own. The Sheets AI function generates when someone clicks, in batches of up to 350 cells, and a change in a concatenated input does not prompt a refresh. Gmail's AI runs per thread or per prompt. If inquiries must be labeled within minutes, at any hour, a person has become the trigger.

It cannot act on the answer. Because the AI function cannot sit inside another formula, the label cannot drive a branch in the same cell. Sending an alert, creating a CRM record or assigning an owner needs another tool.

It lives inside one suite. Power Automate does run on its own and has an approval step, which makes it the strongest DIY option. It fits a team that already works in Microsoft 365, in a supported region. A team on Gmail, Slack and a separate CRM needs something that connects across all three.

So the threshold is concrete: dozens of items a day, a reply time measured in minutes, or two or more apps after the AI step. Past that point, an automation platform runs the trigger, the AI step and the routing as one flow. Zapier's Copilot drafts that flow from a plain-English description. Make (Make.com) puts an AI module in any scenario step and branches on its answer. n8n can hold an AI agent's tools for a person's approval. Our comparison of AI automation tools covers the full set.

The bottom line

AI workflow automation is a trigger, one narrow AI question and a rule that acts on the answer. Start with a closed list of labels and ten test examples, then build it with what you already have: a Google Form and a Sheets AI column, Gemini in Gmail for drafts, or a Run a prompt step with an approval in Power Automate. The DIY version is worth doing even if you outgrow it, because the prompt and the labels carry over. Move to a platform when a person has become the trigger, or when the answer has to reach more than one app.