The First AI Workflow to Automate at a 10-Person Company

The highest-value first AI workflow at 10 people is almost always inbound email triage and draft-reply generation. It requires no in-house AI expertise, runs on tools most companies already pay for, and returns measurable hours within the first week. This guide shows you how to pick the right candidate workflow and ship it using Zapier or Make plus a frontier AI API.

Quick answer:

Target a recurring text-in, text-out task that your team does at least five times a week
Build the trigger → classify → draft loop in Zapier or Make using Claude or OpenAI
Run it in draft mode for one week before routing anything automatically

Verified against Zapier automation docs, Make scenario builder, Anthropic Claude API, and OpenAI API documentation on 2026-07-10.

Prerequisites

A Zapier or Make account (free tier is enough to start)
API access to Claude (Anthropic Console) or GPT-4 (OpenAI Platform)
A shared Gmail or Outlook inbox that receives the recurring messages you want to triage
30–60 minutes to build and test the first run

No engineering background required. Zapier and Make are no-code environments. The AI API keys take about five minutes to generate from each provider's console.

Step 1: Pick Your First Workflow

The right first AI workflow has three properties:

High repetition. The same type of input shows up at least five times a week. Inbound support questions, lead inquiry emails, meeting recap requests, or first-draft generation from a standard template all qualify.
Predictable inputs. The AI needs enough context to act. Open-ended requests with no structure are harder starting points than "here is a customer question, draft a reply using our standard tone."
Text-in, text-out. Avoid workflows that require reading images, navigating UIs, or making decisions across multiple systems as a first project. Start simple.

The most common winner at 10 people: the shared inbox where customer questions, vendor requests, or inbound leads arrive. Most of those messages need the same four or five response types. An AI can classify, draft, and route them in seconds.

Step 2: Map Your Inputs and Outputs

Before opening Zapier, write down exactly what the AI will receive and exactly what you want it to produce.

For email triage, that looks like:

Input: Email subject + body
Classification output: One of three or four categories (support question, pricing inquiry, partnership request, unrelated)
Draft output: A 3–5 sentence reply in your company's tone, with a placeholder for any information the AI cannot know (pricing, specific account details)

Write a sample input and a sample ideal output for each category. These become your test cases in Step 5 and your prompt examples in Step 3.

Step 3: Write the System Prompt

Open the AI provider's playground (Anthropic Console Workbench or OpenAI Playground) and draft the system prompt. A working starting template:

You are an assistant for [Company Name]. When given an inbound email, do two things: (1) classify it as one of: support_question, pricing_inquiry, partnership, or other; (2) draft a reply of 3–5 sentences in a professional, direct tone. Use [PLACEHOLDER] anywhere a specific fact is needed that you don't have. Output format: {"category": "...", "draft": "..."}.

Test it in the playground with three to five real example emails before wiring it into the automation. Adjust until the output is consistently usable. Do not move to Step 4 until the playground is producing drafts you would send with light editing.

Step 4: Build the Automation in Zapier or Make

In Zapier:

Create a new Zap. Set the trigger to Gmail → New Email Matching Search (or Outlook → New Email). Filter to the shared inbox.
Add an action: Code by Zapier (JavaScript or Python) or Zapier's native AI action if your plan includes it. Pass the email subject and body as inputs.
If using Code: make a fetch call to the Anthropic or OpenAI API with your system prompt and the email content. Parse the JSON response.
Add a Gmail action: Create Draft. Populate the To field from the original email's From field. Populate the body with the draft from Step 3.
Optional: add a Slack action to post the category and draft link to a channel for review.

In Make:

Create a new Scenario. Set the trigger module to Gmail → Watch Emails or Microsoft 365 → Watch Emails.
Add an HTTP module set to POST. URL: https://api.anthropic.com/v1/messages. Headers: x-api-key, anthropic-version: 2023-06-01, Content-Type: application/json. Body: construct the messages array with your system prompt and the email content as the user message.
Add a JSON → Parse JSON module to extract category and draft.
Add a Gmail → Create a Draft module. Route the draft to the correct inbox thread.
Add a filter if needed to skip emails from known senders (internal team addresses, newsletters).

Step 5: Test in Draft Mode

Run the automation for one week without sending anything automatically. Every output lands as a Gmail or Outlook draft, not a sent message.

Review each draft daily. Track:

How often the classification is correct
How often the draft is usable as-is or with minor edits
Which input types the AI handles well and which it does not

If the classification accuracy is below 80%, revisit the system prompt and add more specific examples for the failing categories. If draft quality is low for a specific email type, add a dedicated prompt section or example for that type.

Verify

The automation is working when:

Drafts arrive in the inbox within 60 seconds of the original email
Classification is correct for at least 8 out of 10 messages
At least half of the drafts are usable with edits of five words or fewer
The shared inbox is being cleared faster than before

Run a simple before/after comparison: log the average time to respond per email for one week before the automation and one week after. The difference is your baseline for measuring whether to expand to additional workflows.

Troubleshooting

Drafts not appearing: Check that the Zapier or Make scenario is turned on and the Gmail trigger is pointing to the correct inbox. Confirm the API key is active in your Anthropic or OpenAI console.

API authentication errors: Make sure the Authorization header uses Bearer format for OpenAI and x-api-key: for Anthropic. The header format differs between providers.

Classification keeps returning "other": Your system prompt category labels may not match the actual email types arriving in the inbox. Pull 10 recent emails, list the real categories you want, and update the prompt to match them.

Draft output is not JSON: Add explicit instructions in the system prompt: "Respond only with valid JSON. Do not include any text before or after the JSON object." Test again in the playground before re-deploying.

Getting Your First AI Workflow Live This Week

Building the first automation is where most teams stall. Not because the tools are hard, but because picking the right starting point and writing a prompt that actually works takes focused time that never seems to appear. ScaleIt helps startups and SMBs identify the right first workflow, write the prompts, and build the automation in Zapier or Make. Book a free call to get the first workflow shipped this week instead of next quarter.

Verified against Zapier automation documentation, Make scenario builder documentation, Anthropic Claude API reference (Messages endpoint), and OpenAI API reference (Chat Completions endpoint) on 2026-07-10.