The Checklist Big Companies Use Before They Trust AI, Shrunk Down for a Small Business

In healthcare and finance, an AI system doesn’t go live because the demo looked good. It goes live after someone answers a short list of unglamorous questions. What goes in? What does good look like? Who checks it? How will we know when it breaks?

Most small business automations skip all of them. That’s usually why they quietly stop working three weeks in, and nobody notices until a customer does.

You don’t need a compliance team to borrow the list. Here’s the version I’d use for a one-person business. It takes about ten minutes per automation.

1. Write Down What Goes In and What Comes Out

One sentence each. “In: a customer email. Out: a label and a draft reply.” If you can’t write those two sentences, it isn’t ready to build. Big teams call this a spec. Yours can be a note on your phone.

2. Save Ten Real Examples, Including Ugly Ones

Demos always use the clean example. Real inputs have typos, forwarded threads, three questions in one email, and the customer who writes in all caps. Pull ten from your actual inbox or files, and make at least three of them messy.

Larger teams call this a test set. If I could keep only one step on this list, it would be this one.

3. Decide What “Good” Means Before You Look

Write it down first, or you’ll talk yourself into whatever the AI produced. For a reply draft, “good” might mean: uses only facts from the price list, under 80 words, never promises a date.

Run your ten examples and score them. If more than one fails, fix the prompt and run all ten again. Not just the one that failed.

4. Put a Person Where a Mistake Costs You

AI drafts, you approve. That’s the default for anything that reaches a customer or moves money. Enterprise teams call it “human in the loop.” For you, it’s just reading before you hit send. Loosen it later, after weeks of outputs you didn’t need to change.

5. Make It Tell You When It Fails

The worst automations fail silently. A login expires or an app updates, and the whole thing just stops. Zapier and Make can both notify you when a step fails. Turn that on before anything else.

6. Read Five Outputs Every Friday

“Monitoring” sounds like a big-company word. Here it means five minutes on Friday, reading five things the automation made that week. AI models get updated, and the same prompt can start behaving differently. This is how you catch it early.

7. Know How to Turn It Off

Write down how to pause it, and who does the work if you do. “Me, by hand, the old way” is a fine answer. Just know it before you need it.

Copy This for Every Automation

Automation:
In:
Out:
Good means:
Ten test examples saved in:
Passed:   / 10
Who approves before it goes out:
How I'll hear if it fails:
Friday check: yes / no
How to pause it: