10/20/2026
My AI Knows When Not to Think
Should you use AI or a plain script?
Start with the script. If the steps are fixed ("deterministic") and you can check the answer yourself, a script does the job for close to nothing. Save AI for work where the input is messy and the answer takes judgment.
The cheapest AI calls are the ones that leverage deterministic code in the form of tools, utilities and libraries to make sure tokens don't do what scripts can do and the AI is aware of what's available. AI needs to know about these "paved" roads so they take advantage of them without going "off road".
Paying a model to do what a script does perfectly is like hiring a consultant to flip a light switch. I work with AI every day, so you'd expect me to reach for it every time. I don't, and most of what a business asks its systems to do lands on the script side.
| A plain script | An AI model | |
|---|---|---|
| Best for | Fixed steps and an answer you can check | Messy input and an answer that takes judgment |
| Same input, same answer? | Every time | Usually, but not guaranteed |
| Cost to run | Close to nothing | A charge or an allowance used up on every call |
| When it's wrong | Wrong the same way each time, so it's easy to find | Wrong in a new way each time, so it's harder to find |
| Example | Is every link on this page still alive? | What should I say to an unhappy customer? |
What this looks like in my own setup
Four jobs in my own setup that never need a model to think about them:
- Is anything down? A scheduled script checks every hour whether my services are answering. It either sees them or it doesn't.
- Is this draft ready to go live? A checklist script flags broken links, filler phrases and a search snippet that's too long. An AI wrote the rough first draft. A script checks it, and so do I.
- Which searches deserve attention? A report sorts my search data for pages that show up on the first page of results and still get no clicks. That's a sort and a filter, not a judgment call.
- Is it safe to switch the app back on? Before my content app starts, a gate runs a fast set of tests and refuses to start if they fail.
The same split saves real money inside the work itself. Statistical analysis and other hard number-crunching can run in a system, and so can much of the data manipulation, research and aggregation, all without a model. Each of those jobs is small, but they add up quickly in time and in tokens, so I only send a model what needs judgment.
Where AI earns its keep is the rest: the rough first draft of an article, a reply that needs the right tone, a messy question where nobody wrote the steps down in advance.
Let AI write the script once
The best arrangement I've found uses both. Have the AI help you build the script, then let the script do the job every day without a bill. You pay for the thinking once, and the repeating is free.
One caution, the same one I give about everything an AI builds: review what it wrote as thoroughly as your goals and time allow. A script that runs unattended every hour is worth reading before you trust it. How I decide what stays with me is in Which Decisions to Hand AI, and Which to Keep. This question comes first, because it asks whether you need an AI at all.
Three questions before you reach for AI
Could I write the steps down in advance?
If you can list every step ahead of time, a script can follow them.
Can I tell if the answer is right without asking anyone?
If a result is easy to check, a script can check it, and a wrong one is easy to catch.
Does it have to come out the same every time?
Invoices, backups and security checks should never improvise.
Three yeses means a script. Any no means AI, with you checking the result.
This is also one half of an idea I wrote about in Will We Even Need Admin Panels?: once a conversation can run your systems, the next question is which parts of that work should ever need a model to think. The cost side is easy to look up, because AI providers publish what they charge per use; Anthropic's pricing page shows it for Claude.
FAQ
Do I need a developer to use a script instead of AI?
Not to get started. An AI can help you write the script, and you can ask it to explain each step in plain language before you run anything.
What if I can't tell whether a task needs AI?
Try the three questions. If the answers are mixed, start with the AI and watch what it does. If you see it doing the same thing every time, that's a script waiting to be written.
Does this mean AI is too expensive?
No. It means AI is worth the cost on work that needs judgment, and a waste on work that doesn't.
Let my Claude help your Claude
Sorting your own work into script, AI and your call is exactly what my one-on-one AI guidance sessions are for: we do it with the systems you actually use, in your business.
Researched and drafted with AI, then reviewed and edited by the site's editors.