What to have ready before you automate a task with AI

Jerome Privott · · 7 min read

Five labelled boxes in a row reading Steps, Volume, Rules, Correct and Examples, with the caption: Five facts about the task. No AI knowledge needed.

The short answer

You don't need to know anything about AI to get ready. You need five facts about the task: the steps in order, how many times a month it happens, which decisions follow written rules, what a correct result looks like, and about 20 real past examples. Add the name of the person who approves anything that reaches a customer or money. With that on one page, a builder can tell you in a short call whether the task is worth automating and what it costs. Ours is a fixed $5,000 for one workflow.

What should I write down first?

Write the steps in order, in plain words, as if you were training a new hire. Start with what kicks it off ("a quote request lands in the sales inbox"), then each thing a person does, then where it ends ("the reply goes out and the CRM row is updated"). Name every tool it touches: the inbox, the spreadsheet, QuickBooks, your CRM.

Don't tidy it up. The messy bits are what I need to see. "If the drawing is missing, Sam emails them back" is a step. So is "Maria checks the price against last year's sheet, unless it's a repeat customer."

How many times a month does it happen?

Count it, don't guess. Pull last month's emails or invoices and count them. The free Workflow Test on our AI page treats 20 a month as the line where automating starts to pay back. Its sample report shows a task failing that check at "12/mo, needs 20."

Below 20 it can still be worth it, if each one takes an hour or a mistake is expensive. Write down roughly how long one takes too. Volume times minutes is the number that tells you whether this is worth paying for.

Which decisions follow rules, and which need a person?

Go through the steps and mark every decision. For each one, ask: could I write the rule on an index card? "Hold any invoice more than 2% off the purchase order" is a rule. "Decide whether this customer is serious" is judgment.

Rules are the cheap, reliable part. Judgment steps are where a model drafts and a person decides. If most of the task is judgment, the honest answer is an assistant that helps a person, not full automation. Which steps need a model at all is its own question, and I worked through it in AI agent vs Zapier or n8n.

What does a correct result look like?

Write three to five pass/fail statements about the output. Specific enough that two people would grade the same result the same way. The example on our AI page, for a quote-request workflow, is:

  1. Every quote request answered within 2 hours
  2. The price matches the rate sheet
  3. Anything over $5k goes to a person
  4. Every reply is logged in the CRM

This is the most important thing on the page. These statements become the test, and the test is how you know it works after the builder leaves. We sign them off before any code, and the second half of our fee is owed only if they pass. If you can't write them yet, that's fine. Working them out is the first thing we'd do together. What to ask a builder about their test is in Hiring an AI developer? Ask for the test.

Why do I need past examples?

Past examples become the test cases. I ask for about 20 real ones: the input as it arrived (the email, the PDF, the form) and the output a person produced for it (the quote, the keyed invoice, the reply). Twenty covers the normal cases and usually a few odd ones.

The odd ones matter most. The missing drawing, the customer who writes in all caps, the invoice in a currency you rarely see. If the past examples are all clean, the test will be too, and the first messy one will be a surprise in production.

Who approves before it reaches a customer or money?

Name the person. If a wrong result would reach a customer, move money, or touch a legal record, someone signs off before it goes out. The automation drafts, the person approves, and the approval is a step in the workflow, not a habit.

For internal work, like a weekly summary for yourself, you may not need one. Write down which kind your task is.

What should I not send yet?

No passwords, logins or customer records on the first call. A description and redacted samples are enough to scope it. Before you share systems, data or credentials, we sign your NDA, or you can use our short mutual one.

A worked example: the prep sheet for our own intake

The Workflow Test is a live AI workflow on byitl.com that I built for us. Someone describes a task in a form and gets a pass/fail report back by email. Here is its prep sheet, filled in the way I'd ask you to fill in yours:

Item Our answer
Steps Form submitted, Claude on Amazon Bedrock pulls out the facts, plain code checks each fact against the submission, seven fixed checks score it, I approve, the report is emailed
Input Free text, 200 to 5,000 characters
Decisions All seven checks are written rules in code. The judgment call is mine: approve or reject each report
Correct result Every fact backed by a quote that really appears in the submission. An unknown fact never counts as a pass. Same input, same report, every time
Past examples Test cases for every check. The review engine's suite reads "Tests 20 passed (20)"
Sign-off Me, on every report, before it reaches the visitor
Limits Past 50 submissions in a day, new ones are stored and held, not analyzed

Writing the "correct result" line first is what shaped the build. "Every fact backed by a quote" is why there is a separate check, in plain code, that throws out anything the model claims without a matching quote. If I had started with "the model reads the description and scores it," that check wouldn't exist. When I fed it planted errors, it caught the invented ones. One real quote attached to the wrong conclusion still got through, which is why I read every report. Both runs are written up in the Zapier post and in A stranger can steer our AI workflow.

What happens with the list?

You bring it to a 15-minute call, or you paste the steps into the free Workflow Test and get the same seven checks back. Then you get a fixed price. One workflow is $5,000 and takes 2 to 3 weeks. A production build across several workflows starts at $10,000 and takes 4 to 8 weeks. Both are half up front, with the second half owed only if the agreed criteria pass.

What moves the price is on the list you just wrote: how many steps, how many systems it touches, how costly a mistake is, and the volume. Market prices for comparison are on our AI cost page.

Questions people ask

Do I need to know which AI tool to use? No. Describe the task, not the tool. Picking the model and the platform is the builder's job, and often most of the work runs in tools you already pay for.

Can I automate a task that only happens a few times a month? Sometimes. If each one takes an hour, or a mistake costs real money, low volume can still pay back. If it's five minutes a few times a month, probably not, and a good builder should say so.

What if I can't say what a correct result is? Then nobody can tell you it works, including whoever builds it. Start with what you'd check by hand when you review someone else's work. Those checks are your first draft.

Does this work with Zapier, Make or n8n? Yes, and often the right answer is to use them. They connect apps well. What we add is the written criteria and the check that runs them.

I have a device idea, not a task. Is there a list for that? Yes: What to have ready before you talk to a prototype engineer.

Have a task that eats your week? Run it through the free Workflow Test or book a 15-minute call.

Related