Should AI Do This on Its Own? Score It Before You Automate

"Can't AI do this for me?" is a fair question to ask of any job you do every week.

Most of us answer it with a mood. When you're tired on a Friday, the answer is yes to everything, and after one bad experience, it's no. Neither mood has looked at what the task costs when it goes wrong.

Before I build any automation, for myself or for clients, I run a short check. Automation just means setting AI up to do a job by itself, without you doing each step. The check takes two scores, four questions and one of three verdicts. The method is adapted from Laura L.'s AI Workflow Audits.

Write the task in one sentence

Say exactly what you want handed over, the way you'd tell a friend. For example, "Every Monday, add last week's receipts to the spending spreadsheet." If the task won't fit in one sentence, it's two tasks, and each one gets its own check.

Give the task two scores, from one to five

Impact is how much changes if the automation works. A one means you'd barely notice, while a five gives you back a real piece of your week.

Risk is what it costs when the AI gets it wrong. A one is a minute to put right, and a five is money lost, a relationship damaged or something you can't take back. Score the risk for an ordinary bad day, and write one line of reasoning beside each number.

Let four questions overrule your gut

Your risk score is a guess, and guesses run low when you want the thing built. So check your score against four questions, and answer each one with a plain yes or no.

1. Does the task commit you to something? That could be a payment, a booking, a price or a promise.
2. Does the task reach another person directly? That means anything sent, posted or shared before you've seen it.
3. Does something else rely on the task being right? Think of a budget, a tax return or next month's decision.
4. Would a mistake cost you someone's trust? That could be a client's, an employer's or a friend's.

One yes means risk can't be lower than three, and two or more yeses mean risk can't be lower than four. The questions beat the feeling, mine included.

Read off the verdict

• Let it run when risk is one or two and impact is three or more. The AI does the whole job and you check in now and then. This is rare, whatever the adverts say.
• Let it draft, then sign it off. This suits most real work. The AI does the early stages and then stops. Nothing goes out until a named person has checked it. Write that name down, because "Someone will check it" usually means nobody does.
• Keep it when risk is higher than impact. The AI can still gather information, summarise and lay out options, but you do the doing. That means the check has worked, because it stopped a mistake before it happened.

Take the receipts, where impact is a four.

Risk felt like a two, but your tax return will lean on that spreadsheet, so question three forces risk up to three. The verdict is that the AI adds the rows each Monday, and you check the totals against your bank statement before the month closes.

Automatic replies to new enquiries that quote your prices go further. The reply commits you to a price, reaches a person directly and could cost you their trust, so three yeses push risk to four.

Four against four means draft and sign off, which is how Let AI Answer Your DMs sets it up.

Make your AI run the check with you

Nobody remembers four questions in the excitement of a new build. So give the check to the AI as a standing instruction, which is a note the AI reads every time. In Claude, that can be a Skill or a Project, and in ChatGPT, it can be a Project or custom instructions. Project or Skill? explains the difference. Whichever you use, paste this in.

Before you help me automate or hand over any task, run this check with me. Ask one thing at a time.
1. Restate the task in one sentence and ask me to confirm it.
2. Ask me to score Impact from 1 to 5 (how much changes if this works) and Risk from 1 to 5 (what it costs when it goes wrong). Give your own scores with one line of reasoning each, but use mine unless I change them.
3. Ask me these four questions, one at a time. Does it commit me to a payment, booking, price or promise? Does it reach another person before I've seen it? Does anything else rely on it being right? Would a mistake cost me someone's trust? One yes raises Risk to at least 3. Two or more raise it to at least 4. Tell me when that happens.
4. Give a verdict. Let it run if Risk is 1 or 2 and Impact is 3 or more. Keep it if Risk is higher than Impact. Otherwise, let it draft and I sign off, and ask me who signs off.
5. Say in one line what you would and wouldn't touch under that verdict.
Don't design or build anything in the same reply. Wait for me to say go. The decision is mine.

The AI builds nothing until you say go, and it keeps your scores unless you change them.

Score the task again after a month

Verdicts drift. A month of drafts you checked and found right may earn a task "let it run", while a task that has changed shape needs a fresh score.

If something does pass, start the way The First Thing to Let AI Do While You're Not Watching suggests, with a report rather than an action.

Two honest people will score the same task differently, and that's fine, as long as the four questions get straight answers. What the check can't tell you is which jobs you'd miss.

For those, keep a list of what you refuse to automate. For work you share with the AI and don't fully hand over, there are the four boxes.

Before Friday, take the one task you've been itching to automate and score it on paper, before you open any tool.

Credit to Laura L. The Impact and Risk scores, the four-question risk check and the three verdicts are adapted from her AI Workflow Audits, which she shared in the Clief Notes community. The plain names and the examples are mine.