# Automation: Script vs. Agent

By Dan Seyer, Founder, Build Marketing · Published 2026-10-31

> I talked with someone recently who had written a script to automate a business process, and it worked fine. When I brought up agents, they pushed back.

> Script builder: "Why would I need an agent? I have a script that already handles the actions."
>
> Dan: "The script handles the actions, but it doesn't think about them. Add AI reasoning and the software can do things nobody wrote into the script."

When they built it, a script was the obvious choice. Now AI can write most of that script for you, you can put AI inside the workflow so it reasons about the data as it goes, or you can build an agent. Which would you pick?

Take a salesperson who hasn't followed up with a customer in 30 days. A script can spot the gap and create a reminder, and if a reminder is all you need, that's probably enough. But maybe the customer went quiet because an important support ticket is still open. A good account manager would notice that before sending another sales email. I wanted to know if software could do the same.

## AI Summary

_Summary by AI._

- There are three ways to automate a process: a traditional script that follows fixed rules, a predefined workflow that uses AI to interpret data at certain steps, and an AI agent that chooses its own steps and tools.
- Using AI to write the code does not make the result an agent. The distinction is in what happens at run time. The article recommends the simplest option that meets the business need.
- The running example is a customer with no sales contact in 30 days. A script creates a reminder. An agent can read across CRM, support and ERP data and recommend fixing an open ticket before outreach.
- McKinsey (August 2026): nearly nine in ten respondents use AI regularly, about two in ten have scaled AI agents across the organization, 37 percent report a positive EBIT contribution, and 32 percent chose to build software in-house instead of buying it.
- Trust in an agent is built in stages: observe, recommend, measure, then delegate bounded decisions. Hard limits belong in permissions and fixed rules, and higher-risk actions keep human approval.
- A governed improvement loop can use human corrections and oversight findings to propose new prompts, tools and code, with review and testing before anything is deployed.

## The Choice We Didn't Have Before

An AI coding tool can write a traditional script, and that script runs without any AI at all. So who wrote the code doesn't tell you much. What counts is what happens when the automation runs.

A script follows the logic its author wrote. If you add AI to a workflow, the AI might read an email or classify an account, but the workflow still decides the next step. An agent decides that for itself, including what to look up and which tools to use.

People call almost any background automation an agent these days. I don't mind that in conversation, but it blurs a real difference in how the system is built. Anthropic's engineering team draws the line well: "Workflows are systems where LLMs and tools are orchestrated through predefined code paths." In an agent, the model directs its own process and picks its tools.

Flexibility costs money and adds risk. A fixed process is easier to test, while an agent can handle cases you didn't plan for. I don't think either one wins by default.

![Three ways to handle a quiet customer with no sales contact in 30 days. A script follows the rules you wrote: it sees 30 days with no contact and creates a reminder. An AI workflow adds reasoning inside fixed steps: it sees an open support ticket and flags it before creating the reminder. An agent decides its own next step: it looks at the ticket, recent orders and the renewal, and recommends fixing the issue before reaching out. Pick the simplest one that does the job.](https://www.buildmarketing.ai/images/script-vs-agent-three-options.svg)

## Connecting the Dots

Back to the quiet customer. The CRM knows there's been no sales activity. The support system knows about the open problem, and the ERP may know a big renewal is coming up. Someone who knows the customer would put those pieces together.

If AI has permission to read those systems, it can start doing the same. An AI workflow might check for open support issues before it creates a sales reminder. An agent could read the ticket, look at recent orders and suggest fixing the issue before anyone reaches out. None of that works on bad data, though, and you still need a way to check whether its conclusions are right.

Marketing has the same pattern. A script sends one campaign email to a thousand accounts, and an AI workflow can personalize it by segment. An agent could look at what each account has bought and what's open with support, then decide that some accounts should get a different message and some should get nothing.

You don't need an agent to personalize an email. You need one when the process has to work out what matters and change course when it finds something unexpected.

## The Numbers

McKinsey's State of AI survey from August 2026 found that nearly nine in ten respondents say their organizations use AI regularly in at least one function. Only about two in ten say they've reached the scaling phase with AI agents across their organization, and 37 percent say AI has added to their EBIT.

The number that stood out to me was 32 percent: the share who say their organization decided not to buy at least one software product or feature because it could build the same thing in-house with agentic coding tools. Software is getting easier to build. I wouldn't take that as a reason to make every process autonomous.

## Building Trust in an Agent

Say our CRM agent recommends holding the follow-up because of the support issue. It might be right. Would you let it email customers on its own? I wouldn't, at least not at first.

I'd let it watch for a while, analyzing accounts and showing me what it sees. After that it can recommend actions with its reasons and evidence, and a person approves or rejects each one. Meanwhile I'd keep score. How often is it useful? When it's wrong, what does that cost? Does it do as well on small accounts as on big ones?

If it holds up on a defined task, I'd give it a few limited decisions. It could create an internal task, for example, but not email a customer. I'd widen its permissions only when the results back it up, and only for that job. An agent that's good at creating follow-up tasks hasn't earned the right to change a price or post a journal entry.

## Who Watches the Agent?

A second agent can review the first one's work. Did it use the right facts? Does the message sound like us? Is there a legal or compliance problem, and would the action actually help?

The reviewer can be wrong too. Some limits belong in software permissions and fixed rules rather than another model's opinion, and higher-risk actions should still go to a person.

## Code Writing Code

Now say the agent sends an internal recommendation and the salesperson rejects it. The oversight system records why. Maybe the support case had already been closed and the agent was working from stale data.

A continuous improvement loop can keep that lesson. It collects the agent's decisions, the corrections people made, what happened in the business afterward and what the reviewer flagged, then proposes changes to the prompts, the data the agent pulls, its tools or the code itself.

This is the part I find most interesting. A coding agent can write a new integration or a better exception check, for the main agent or for the one watching it. A script, by comparison, repeats the logic we gave it until someone changes the code.

Deploying that new code safely is a separate job. Every change should be reviewed, tested against cases where you already know the right answer, and released the way you'd release any other software. Skip that and the loop can spread its own mistakes faster than anyone notices.

## Script, Workflow or Agent?

If you have a process in mind, paste this into ChatGPT, Claude or whatever you use. It interviews you about the process and recommends the simplest option that will do the job.

**Prompt: Do you need a script, a workflow or an agent?** (for your favorite AI tool (ChatGPT, Claude, Gemini, Copilot))

Interviews you about one business process and recommends the simplest way to automate it, including where a person should approve and how to improve it safely.

```text
Act as an automation architect. Interview me one question at a time about a business process I want to automate. Ask about its inputs, outputs, fixed rules, exceptions, data sources, the cost of mistakes, required approvals, and whether the next step changes based on what the system finds.

Then recommend one of three options:
1. A traditional script, written with AI's help.
2. A predefined workflow with AI reasoning at certain steps.
3. An AI agent that chooses its own steps and tools.

Explain why the simplest option that does the job is best. Show a concrete example of what each option would do for my process. Tell me which decisions should require a person's approval and which could earn autonomy after testing. Propose metrics, a way to oversee the system, and a safe loop for improving it over time.

Do not assume an agent is best. If a script is enough, say so.
```

More prompts: [Prompt Library](https://www.buildmarketing.ai/prompts.md)

If you're new to agents, I covered the basics in [What Is an AI Agent?](https://www.buildmarketing.ai/what-is-an-ai-agent) and how to scope one in [How to Build an AI Agent](https://www.buildmarketing.ai/how-to-build-an-ai-agent).

## What I'd Build

So, why would you need an agent if your script already handles the actions? You might not, and I wouldn't replace a working script just because agents are available.

I'd look for the spots where people still have to connect the dots or make a judgment call, and decide whether that reasoning fits inside a fixed workflow or really needs an agent. What interests me most is a system that can notice where the process itself could be better and propose the change.

An agent can only reason about what it can read. [Run MachineReady on your website](https://www.buildmarketing.ai/machineready/) to see how AI reads it today, or [get in touch](https://www.buildmarketing.ai/#Get-In-Touch) if you want help deciding whether a process needs a script, a workflow or an agent. Rather have it built for you? [We build AI agents for clients](https://www.buildmarketing.ai/ai-agents/).

*Sources: McKinsey & Company, "The State of AI in 2026: On the Road to ROI," August 25, 2026. Anthropic Engineering, "Building Effective Agents," December 19, 2024.*