Use a Script for the Repeat Job. Use AI for the Messy One.

With AI everywhere, a person in the loop and predictable, repeatable automation matter more than ever. AI makes it easy to act fast at scale. The steps that must be right every time belong in a script. The decisions that matter belong with a person.

Human in the loop means a person still decides before an important action goes out. How to limit a tool that can act is in AI Agents Can Act. A Person Should Still Approve the Work.

Business owners hear two pitches. One says a script can run the busywork. The other says an AI agent can do the thinking. Both can be true. This is how to tell which job belongs to which.

A script does the same thing every time

Deterministic means the same input always leads to the same result. A deterministic script is a short program with fixed rules. Give it the same file, the same numbers, or the same form, and it takes the same steps. If the rules are right, Tuesday's run matches Monday's run.

That is why a script fits repeatable, rule-based work. It is cheap to run, fast, and easy to test. It is also easy to audit. Easy to audit means a person can read the steps and check the result against the rule.

Use a script when you can write the rule down:

• Backups that copy files on a schedule
• Invoices filled in from a template
• Data syncs that move records from one place to another
• Price checks against a list you already trust
• A web form that lands as a row in a spreadsheet
• Scheduled reports
• Security scans
• Reminders that go out on a date you chose

If a person would do those steps the same way every time, a script can do them. When a required field is missing, the script should stop and say so. It should not invent a number to fill the gap.

An LLM is for work that does not fit a rule

An LLM, short for large language model, is the kind of AI that reads and writes text. It is not a calculator, and it is not a checklist. An agent is software that can take a next step with that model: read a message, draft a reply, or change a file. It does not only suggest. It can act. Those actions need the guardrails linked above.

LLMs and agents are a good fit for messy, unstructured, or judgment work:

• Reading a free-form email and pulling out what the sender asked
• Summarizing a long thread or a pile of notes
• Drafting a page, a reply, or a first version of a procedure
• Research, when the question is open and the sources are not in one table
• Sorting items that do not fit the categories you already wrote down
• Deciding a reasonable next step when the rules run out

They are slower than a script. They cost more per run. Ask the same question twice and the wording can change. They can also be confidently wrong: the sentence sounds finished, and the fact is off. That is why an agent needs guardrails. A person approves anything that goes live, sends a message, or spends money.

A short way to choose

Ask these questions before you pick a tool.

• Does it follow fixed rules? Use a script.
• Is the result the same every time for the same input? Use a script.
• Does it need reading or judgment? Use an agent, and have a person approve the result.
• Would a mistake be costly or hard to undo? Keep a person in the loop.

You can answer yes to more than one. A nightly report can be a script. The paragraph that explains an odd number in that report can be a draft from an LLM, which a person edits before anyone sends it.

Use both, and turn repeats into scripts

The useful pattern splits the job. The agent handles the fuzzy step. Scripts handle everything that is already repeatable. Then watch the agent. When it does the same task the same way more than once, turn that task into a script. That cuts cost and keeps the result the same.

Picture a business that sends the same invoice shape every month. The customer name, the amount, and the due date already live in a spreadsheet. A script fills the template and saves the file. If the cover note should sound like a person wrote it, an LLM can draft that note. A person reads it and sends it.

Picture an inbox that does not follow a form. Customers write billing questions, new orders, and complaints in whatever words they have. An LLM can read one message and sort it. A person checks the sort before anyone replies. The reply that is always the same, such as "we received this and will answer by the next business day," can be a script. The reply that depends on the complaint stays a draft for a person to edit.

How D3 handles it

D3 follows an automate first rule. When a script or a built-in integration can do a job reliably, D3 uses it instead of having an AI do that job by hand each time. Reusable scripts stay on a shared list, so the same job is not rebuilt from scratch.

AI agents help with research, drafting, and code. JD approves anything that goes live. When an agent does the same task again, D3 turns that task into a script. D3 is always looking for ways to improve.

With AI everywhere, a person in the loop and predictable, repeatable automation matter more than ever. The steps that must be right every time stay in scripts. The decisions that matter stay with a person.

If you are not sure which your business needs, start with the job, not the tool. Use the contact page.