Put AI to Work

Level 3 · Put AI to Work

1. Concept: Pick the right job

   The best first job happens often, bores you silly and follows rules you could fit on an index card.

   1. Term: The three-question test

      Ask three things: does it happen often, is it boring, and does it follow rules?

      See also Jobs to keep

      1. In plain words: Three yeses

         Often means every day or every week, not twice a year. Boring means you could do it half asleep. Rule-ish means you could explain it to a new hire on one page. Three yeses make a good first job.

      2. How it works: Score it 1 to 3

         Give each question a score from 1 to 3 and add them up. Copying open-house sign-in sheets into a CRM scores a 9. Choosing your new logo scores about a 3. Start with your highest score.

      3. Try it: Score five chores

         Write down five chores from last week and score each one. Circle the winner. That's the one you'll build in class.

      4. Watch out: Rare jobs don't pay off

         A job you do twice a year isn't worth automating. By the time it comes around again, the apps may have changed and you'll have forgotten how the automation works.

   2. Term: Jobs to keep

      Anything that needs judgment, a soft touch, or a call about money or the law stays with you.

      1. In plain words: A hammer, not a carpenter

         Let AI do the pounding, and keep the calls that need judgment for yourself: pricing a special job, calming an upset customer, anything legal or medical.

      2. How it works: Split the job in two

         Most jobs have a boring half and a judgment half. Let AI pull the facts out of an angry email and draft a reply, then you decide what to actually say.

         See also The approval step

      3. Watch out: Money and promises

         Never let an automation spend money, give a discount or promise a date on its own. Those mistakes cost real dollars and real trust.

   3. Term: Start with one

      One trigger, one job, one place it lands; grow it after it earns your trust.

      1. In plain words: Small enough to say out loud

         A good first automation fits in one sentence: when a quote request comes in, AI pulls out the details, and I approve the new row.

      2. Try it: The one-breath test

         Say your automation out loud in one breath. If you run out of air, split it in two.

      3. Watch out: Don't automate a mess

         If the job is confusing by hand, automating it just makes the confusion faster. Clean up the steps first.

         See also The SOP prompt

2. Concept: Write it down like an SOP

   Write the job the way you'd train a new hire: numbered steps, one good example, and what to do when something looks off.

   1. Term: The SOP prompt

      Your instructions to the AI, written like a standard operating procedure.

      1. In plain words: The new-hire binder

         An SOP is the binder that shows a new hire how a job gets done. An SOP prompt is the same binder, written for an AI.

      2. How it works: Four parts

         Input: what arrives, like a quote request. Steps: numbered, one action each. Output: the exact fields you want back. Rules: what it must never do.

      3. Try it: Test it by hand first

         Paste your SOP and a fake example into two different AI chat apps, say ChatGPT and Gemini, or Claude and Copilot. If either one stumbles, fix the steps before you build anything.

         See also Build the small version

      4. Watch out: Vague in, vague out

         "Handle my leads" is a wish, not an instruction. "Copy the name, phone and email exactly as written" is an instruction.

         See also Show an example

   2. Term: Show an example

      One good example teaches the format faster than a page of rules.

      1. In plain words: Show, don't tell

         Paste one example of what comes in and exactly what should come out. The AI copies the pattern.

      2. How it works: Same fields, every time

         Ask for the same labeled fields every time, like Name, Phone, Email and Notes. The automation tool can only drop each answer into the right column if it comes back in the same shape.

      3. Watch out: Use made-up people

         Build examples with fake names and phone numbers like 555-0123, from the 555-0100 to 555-0199 block set aside for make-believe. Real customers don't belong in a test.

   3. Term: When in doubt, flag it

      Tell the AI to say when it's unsure, instead of guessing.

      1. In plain words: Permission to say "I don't know"

         AI would rather guess than leave a blank. Give it a way out: if a field is hard to read, write CHECK instead of guessing.

      2. How it works: A column you can sort

         Ask for a Needs review field, yes or no, with a short reason. Sort by it and look at the iffy ones first.

      3. Watch out: Guesses look tidy

         A misread phone number looks just as neat as a right one. That's why a person looks at every row before it goes anywhere.

         See also The approval step

3. Concept: Connect the tools you already use

   A connector tool watches for something to happen, hands it to AI, and puts the result where you already work.

   1. Term: Triggers and actions

      A trigger starts the automation, and actions are the steps that follow.

      See also Connector tools

      1. In plain words: When this, do that

         Every automation is a sentence: when this happens, do that. The trigger is the when: a new form entry, email or spreadsheet row. The actions are the do: read it, add a row, draft a reply.

      2. How it works: A bucket brigade

         The tool watches for the trigger, then runs each action in order and passes the details down the line. Zapier calls each link a step, Make calls it a module, and n8n calls it a node.

      3. Try it: Fill in the blanks

         Finish this sentence for your chore: when ___ happens, AI should ___, then I check it, then it goes to ___.

   2. Term: Connector tools

      Zapier, Make, n8n, Power Automate and Google's own tools all do the same basic job, so pick by the apps you already use.

      1. In plain words: Same job, different shops

         Zapier has the biggest app list, and Make draws your steps as a map. n8n can run on your own computer. Power Automate lives in Microsoft 365, and Workspace Studio lives in Google Workspace. Pick by the apps you already use, not by whose ad you saw last.

      2. What it costs: Free to start

         As of Oct 23, 2026: Zapier Free gives 100 tasks a month, and paid plans start at $19.99 a month billed yearly. Make Free gives 1,000 credits a month. n8n is free to run yourself, or from $20 a month billed yearly on its cloud. Power Automate's basics come with Microsoft 365 business plans.

      3. How it works: They count differently

         Zapier counts tasks (each action that runs), Make counts credits (most actions use one), and n8n counts whole runs. Multiply your steps by how many times a month it runs to guess the bill.

      4. Watch out: Free plans have catches

         Zapier's free plan stops at two steps, too few for an AI step plus an approval. Make's free plan allows two active automations, which it calls scenarios. Read the limits before you build on a free plan.

         See also Check your plan

   3. Term: The sign-in sheet example

      A real example, client name left out: an open-house sign-in sheet becomes CRM contacts with one tap.

      1. In plain words: Snap, check, tap

         A local real estate agent snaps a photo of the open-house sign-in sheet. AI reads each name, phone and email. The agent checks them on screen, fixes any smudges and taps send.

      2. How it works: One tap, four places

         The tap hands the contacts to Zapier, which adds them to the agent's CRM, email list and phone contacts, then calls a self-hosted n8n flow that adds a row to a shared spreadsheet. The app in daily use reads sheets with xAI's Grok, and the rebuild being tested uses Google's Gemini, so the AI step is swappable.

      3. Watch out: The check screen is the point

         Handwriting is messy, and AI misreads some of it. The check screen comes before the tap on purpose, so a smudged 7 never lands in four places at once.

         See also When in doubt, flag it

      4. Try it: Build the small version

         Fill out a sign-in sheet with three made-up guests, snap a photo, and ask an AI chat app for each guest's name, phone and email. Count the mistakes before you trust it with real guests.

   4. Term: Keep it on your own laptop

      n8n and a small AI model can both run on your own laptop, so the AI step never sends customer details to an AI company.

      See also Keep customer data safe

      1. In plain words: Your machine, your data

         Self-hosted means the software runs on a computer you control. When n8n and the AI model both run on your laptop, the AI step doesn't hand customer details to anyone.

      2. How it works: Two free programs

         n8n starts with one command, npx n8n, and opens in your browser at localhost:5678. Ollama runs a small model, like Llama 3.2 3B, on the laptop's own processor, and n8n hands it each job.

      3. What it costs: Free, but slower

         Both are free to run yourself. You pay in speed and smarts: on a class ThinkPad, a small model can take several seconds to a couple of minutes per item (our estimate, so we time each laptop before class), and it misreads more than the big cloud models.

      4. Watch out: The laptop has to be awake

         A flow on your laptop only runs while the laptop is on and n8n is open. For around-the-clock jobs, use a small always-on computer or a cloud plan.

4. Concept: Stay in charge

   Put an approval step before anything reaches a customer, send each tool only what it needs, and know where the off switch is.

   1. Term: The approval step

      A pause where the automation waits for a person to approve, fix or reject its work.

      See also Drafts, not sends

      1. In plain words: Nothing goes out without your OK

         Techies call it human in the loop. The automation does the work, then stops and asks you. You approve it, fix it or toss it.

      2. How it works: Built into most tools

         Zapier has a Human in the Loop step, n8n can email or Slack you and wait for a yes, and Power Automate sends approvals you answer from Outlook or Teams. The simplest version works anywhere: an OK to send? column in your spreadsheet that a second automation watches.

      3. What it costs: Check your plan

         As of Oct 23, 2026, Zapier's approval step needs a multi-step Zap, so a paid plan or the 14-day trial, and on its Professional plan it can only ask you, not your staff. Power Automate's Approvals come with Microsoft 365 business plans. In n8n, approval is a regular step, so it's free when you run n8n yourself.

      4. Watch out: Approve once, or every time?

         A template you wrote and approved yourself, like a welcome email, can go out on its own. Anything the AI writes fresh waits for a person, every time.

   2. Term: Drafts, not sends

      Have the automation leave replies in your Drafts folder, so sending is always your click.

      1. In plain words: Your finger on the button

         A draft is a finished email waiting in your Drafts folder. The automation does the writing; you read it and press Send.

      2. How it works: Every tool can do it

         Zapier, Make and n8n can all create Gmail drafts. Google's Workspace Studio has a Draft an email step, and Apps Script can do it for free with a little code an AI can write for you.

      3. Try it: Read it like the customer

         Have your automation draft a reply to a fake customer email. Open your Drafts folder and read it as if you were the customer. Would you be glad to get it?

      4. Watch out: Drafts pile up

         If nobody checks, drafts pile up and customers wait. Pick a time each day to clear the folder, and pause the automation when you're on vacation.

   3. Term: Keep customer data safe

      Send each tool only what the job needs, and use plans that don't train on your data.

      1. In plain words: Only what the job needs

         Every tool in the chain sees the data that passes through it, and most keep a copy in their run history. Send the name and email the job needs, not the whole customer file.

      2. How it works: Plans and settings decide

         Business plans from Anthropic, OpenAI, Google and Microsoft don't train on your data by default. Neither do the paid AI hookups (APIs) from Anthropic, OpenAI and Google that automation tools plug into. Free tiers can differ: Google says its free Gemini API tier is used to improve its products, and its paid tier is not (checked Oct 23, 2026).

         See also Keep it on your own laptop

      3. Try it: Find the training switch

         Open the settings of the AI tool your automation uses and find its data or privacy page. If it says your chats may train its models, switch that off or move to a business plan before real customers go in.

      4. Watch out: Some data never goes in

         Keep Social Security, driver's license, bank and card numbers out of automations entirely, and health details too. Under Iowa's breach law (Iowa Code chapter 715C), a leak of names with those numbers can mean notifying every customer it touched.

   4. Term: Run history and the off switch

      Every tool keeps a record of each run and has a pause button, so learn both before you need them.

      1. In plain words: Know where the brake is

         Before you turn an automation on, find the switch that turns it off. Write it on a sticky note, along with who else knows where it is.

      2. How it works: Read the run history

         Each run leaves a record: what came in, what each step did and where it stopped. When something looks wrong, the history shows which step to fix.

      3. Try it: Break it on purpose

         Feed your automation a blank form, a smudged photo and a joke entry. A good one flags all three instead of passing them along.

         See also When in doubt, flag it

      4. Watch out: Logins are keys

         Connecting an app hands the tool a key to that account. Turn on two-step sign-in, never paste passwords or API keys (the passwords apps use to talk to each other) into a chat, and remove connections when someone leaves.