Automate your workflow with AI: from saved prompts to simple agents

Start with the ladder

Automation with AI has five rungs. Most people jump from the first to the fourth and get burned. Climb one at a time, and only climb when the rung below is stable.

1. Saved prompts and templates. You still run the task, but you stop rewriting the instructions.
2. Custom assistants. The instructions and reference files live in the tool; you open it and paste the input.
3. AI inside your existing apps. Scheduled prompts, meeting recaps, and inbox features that run without you opening a chat.
4. Automation platforms with an AI step. Zapier, Make, or n8n watches for a trigger, calls a model, and does something with the answer. Nobody has to be at the keyboard.
5. Simple agents. An AI that decides which steps to take and uses tools to take them, with you approving the important ones.

The test for moving up a rung: you have run the task at the current rung at least ten times, the output was acceptable in nearly all of them, and you can write down what "acceptable" means.

Rung 1: templates and saved prompts

A prompt becomes a template the moment you replace the specifics with placeholders. Keep a plain document titled "Prompts" and give each one a name, the placeholders in ALL CAPS, and a one-line note on what to check afterward. A text expander or your notes app is enough; the value is in having the wording settled.

Two features make this faster. ChatGPT's custom instructions and memory let you set standing preferences (your role, your tone, formats you like) once. Claude Projects and Gemini Gems do the same for a specific job. Both mean the saved prompt gets shorter, because the context is already there.

Worked example: an operations manager writes the same weekly update every Friday. The saved prompt asks for a five-bullet summary from pasted notes, in the team's tone, with any missed deadline named first. The update that took forty minutes now takes ten, and the prompt has not changed in six months. That stability is the signal to move up a rung.

Rung 2: custom assistants

When the prompt and its reference files are stable, move them into a custom GPT, Claude Project, or Gem so that anyone on the team can use them without seeing the instructions. The full walk-through, including three example assistants, is in Custom GPTs vs Claude Projects vs Gemini Gems.

The rule at this rung: one assistant, one job. An assistant that formats notes, drafts emails, and answers policy questions will do all three worse than three narrow ones.

Rung 3: AI inside the apps you already use

Before building anything, check what your existing tools already automate. Zoom AI Companion and the meeting bots produce recaps and action items without a prompt. Gemini in Gmail and Copilot in Outlook summarize threads and draft replies in place. Notion AI answers questions across your workspace. ChatGPT and Gemini both offer scheduled prompts that run on a timer and deliver the result to you, which covers "every Monday, summarize what changed" jobs with no automation platform at all.

These are the cheapest automations you will ever set up, because they are already paid for and already inside your company's security perimeter.

Rung 4: Zapier, Make, or n8n with an AI step

This is where the work runs without you. The pattern is always the same: a trigger (something happened), an AI step (a model reads it and produces something), and one or more actions (send, save, notify).

Which platform

• Zapier if you want the biggest catalog of app connections and the easiest builder, and you are fine paying per task.
• Make if you want a visual canvas with branching, better handling of lists and data, and a lower cost per operation.
• n8n if you have someone technical and want to self-host so data never leaves your infrastructure, or you want to build more complex agent workflows.

All three have built-in steps for ChatGPT and Claude models, plus generic HTTP steps for anything else. The integrations use the vendors' API side, where inputs are not used for training under the standard API terms, which is one reason to prefer them over pasting into a consumer chat account. Confirm this with your own organization's policy before the first real record flows.

Worked example: form submission to AI summary to Slack and email

The scenario: your website's contact form gets fifteen to forty submissions a week. Half are vendors pitching you, some are support requests, and a few are real sales leads. Someone reads every one, which means nobody reads them until Thursday.

Here is the automation in Zapier. Make and n8n are structurally identical.

Step 1: Trigger. "New form response" from your form tool (Google Forms, Typeform, Jotform, HubSpot, or whatever you use). The trigger pulls in each field: name, email, company, message.

Step 2: AI step. Add a ChatGPT or Claude step. Map the form fields into the prompt below. Ask for a fixed output format so the next steps can read it reliably.

You are the intake assistant for [COMPANY], a [ONE-LINE DESCRIPTION OF THE BUSINESS]. A new message just arrived through our website contact form. Classify it and summarize it for the team.

Form fields:
Name: [NAME FIELD]
Email: [EMAIL FIELD]
Company: [COMPANY FIELD]
Message: [MESSAGE FIELD]

Respond with exactly these five lines and nothing else:
Category: one of SALES LEAD, SUPPORT, VENDOR PITCH, PARTNERSHIP, SPAM, OTHER
Priority: HIGH, MEDIUM, or LOW. HIGH only if the sender names a budget, a deadline within 30 days, or a problem with an existing order or account.
Summary: one sentence, under 30 words, in plain English.
Suggested owner: SALES, SUPPORT, or FOUNDER, based on the category.
Draft reply: two sentences we could send as a first response, or NONE if the category is SPAM or VENDOR PITCH.

Rules: use only what is in the form. Do not invent company details. If the message is empty or unreadable, set Category to OTHER and say so in the Summary.

Step 3: Parse. Use a formatter step to split the response into fields (Zapier's Formatter, Make's text parser, or an expression in n8n). You now have Category, Priority, Summary, Owner, and Draft as separate values.

Step 4: Filter and route. Add a filter: continue only if Category is not SPAM. Add a path: if Priority is HIGH, run both actions below; otherwise, Slack only.

Step 5: Actions. Post to a Slack channel: the summary, the category, the sender's name and company, and a link to the form entry. For HIGH priority, also send an email to the suggested owner with the draft reply included so they can edit and send it.

Step 6: Log. Add a row to a Google Sheet or Airtable base with every field plus the raw AI response. You will need this for the audit described in the testing section.

Cost is small: one model call per submission, priced per use by the model vendor, plus one platform task per step. What changes is that a lead now reaches sales in two minutes instead of two days, and the vendor pitches never reach anyone.

Note what the automation does not do: it does not reply to the sender. The draft goes to a person. That is deliberate, and the section on human-in-the-loop checks explains why.

Rung 5: simple agents

An agent differs from an automation in one way: it chooses the steps. You give it a goal and tools; it decides what to do. ChatGPT's agent mode can browse and complete multi-step tasks in a browser. Microsoft's Copilot Studio builds agents that act across Microsoft 365. Zapier Agents and Make's agents do the same across connected apps. Claude Code is an agent for software work.

Treat agents with the caution that "it chooses the steps" deserves. Give an agent read-only access first. Then let it draft but not send. Only after you have watched it succeed repeatedly should it act on its own, and even then, keep a short list of actions that always require a human: anything that spends money, deletes data, or messages someone outside the company.

Agents that read web pages or inbound documents face one more risk: text inside those sources can try to redirect them. Keep the agent's permissions narrow and its tools few, and treat any unexpected action as a reason to stop and read the log.

When not to automate

• The task is different every time. If you cannot write the rules, the model cannot follow them.
• The volume is low. Ten submissions a month do not justify an afternoon of building plus ongoing maintenance.
• The failure is silent. An automation that sends a wrong summary to a customer fails in a way nobody notices until it matters. Automate the internal draft, keep the human on the send.
• The inputs contain regulated data. Health information, privileged legal material, payroll, and customer financial data need compliance sign-off on the specific platform and model plan before a single record flows. Never route confidential data through a consumer AI account.
• You have not done it manually yet. Automating a process you do not understand produces a fast version of the wrong thing.

Testing and human-in-the-loop checks

Every automation with an AI step needs four things before it runs on real traffic.

A golden set. Collect ten to twenty real past inputs, including the awkward ones: the empty message, the one in Spanish, the angry customer, the pitch that looks like a lead. Write down the output you would expect for each. Run them through the AI step and compare.

A review gate. For the first two weeks, route every output to a person before any action fires. In the form example, that means the Slack post goes to a private channel and the email goes to you instead of the owner. Promote it to live only after you have watched it get a run of real cases right.

A log and a weekly sample. Keep the sheet from Step 6. Once a week, read ten random rows and grade them. When the model misclassifies, fix the prompt, add the case to your golden set, and re-run the whole set before you ship the change.

A kill switch and an alert. Know exactly how to turn the automation off, and set the platform to email you when a run fails. The most common failure is not a bad summary; it is a form field that was renamed and now maps to nothing.

Keep a human in the loop wherever the output leaves the company or commits anyone to anything. Inside the company, a wrong summary costs a minute. Outside, it costs trust.

Next steps

• Compare the platforms on their tool pages: Zapier, Make, and n8n.
• Build the assistant that feeds your automation: Custom GPTs vs Claude Projects vs Gemini Gems.
• Get the underlying prompt right with the prompt engineering guide and the prompt builder.
• Read the terms: workflow automation, AI agent, human-in-the-loop, and prompt injection.
• Estimate the payoff with the time savings calculator.
• See how this plays out for a role: AI for operations managers and AI for small business owners.