AI Agents vs Automation: How I’m Learning When to Use Each

This week I’ve been trying to get a much better understanding of AI agents. How do you build them? What resources do I already have available to me? What does each LLM offer? ChatGPT, Claude, Copilot…? For someone like me who’s addicted to learning, we really are living through an amazing time. Luckily, my workplace is very pro-AI adoption, which means I also have a lot of talented colleagues to learn from, collaborate with and share real-world projects with. But the biggest thing I’ve learned so far?

Not everything needs an agent.

In fact, I’m starting to realise there may be just as many cases for not using an agent as there are for using one. And naturally, my creative brain has started exploring the possibilities. The more I’ve learned about agents, the more I’ve found myself looking again at automation and where the dividing line between the two really sits.

AI agent or automation?

The simplest question I’m starting to ask is:

Does anything actually need to make a decision here?

If the steps are predictable and the logic is already known, automation is probably the better answer.
If the task needs judgement, context and a decision about what to do next, that’s where an agent starts to become more interesting. That sounds obvious written down. It wasn’t obvious to me when I started looking into this.
I think part of the problem is that AI agents are currently being talked about as if they are the solution to every repetitive task. Once I started thinking about actual problems rather than the technology, some of those “agent” opportunities turned out to be fairly normal automation problems.

Google Ads is a useful example

Negative keyword management is one of those jobs in Google Ads that can become incredibly monotonous. You review the search terms coming through your campaigns, identify irrelevant searches and add the appropriate negative keywords. Initially, that felt like a perfect example of something I would want an AI agent to do.
But there are actually two different jobs hiding inside that process. Adding a known negative keyword to a campaign is predictable. Once the decision has been made, the action itself can already be automated using things like Google Ads Scripts.

The more interesting part is making the decision. An agent could potentially look at a search term, compare it against the objective of the campaign, understand the products or services being advertised and assess whether that search is genuinely irrelevant. It might then produce a weekly report saying:

“These are the search terms I think should be excluded, these are the campaigns affected, and this is why.”

Initially, I would still want to approve those decisions myself. Once I trusted the process, there might eventually be circumstances where I allowed the agent to make certain changes automatically within clearly defined rules. That feels very different from simply writing a script that says:

“If X happens, do Y.”

• The script follows instructions.
• The agent has to make an assessment.

That is the dividing line that has started to click for me.

Power BI made the distinction even clearer

At the same time, I’ve spent a lot of this week working on a Power BI reporting project.

The goal is to bring together Google Ads, GA4 and WooCommerce order data so campaign performance can be measured against what actually happened commercially. The goal sounds simple. The reality has meant working through campaign data, attribution gaps, order statuses, customer history, new versus returning customers, and plenty of Power Query errors along the way. But once the connections and logic are working, the report itself does not necessarily need an AI agent.

• The data sources can be connected.
• The transformations can be defined.
• The relationships can be built.
• The report can then update using those predefined processes.
• That is automation doing exactly what automation is good at.

The interesting part comes afterwards. Instead of me opening the finished report and manually looking for changes, an agent could potentially analyse the latest data and ask questions such as:

• What has changed?
• What looks unusual?
• Which campaigns deserve attention?
• Where is there a gap between attributed Google Ads revenue and actual completed WooCommerce revenue?
• Has the proportion of new versus returning customers changed?
• Is there something happening that warrants investigation?

That feels like the exciting bit.

The reporting can be automated.

The interpretation can become agentic.

The real value is in understanding what each tool is good at

This week has made me look again at tools and processes I already use.

Power BI is still an excellent tool for collating multiple data sources, applying repeatable logic and building useful reporting. There is no reason to replace that with an agent simply because agents are new and interesting. Where an agent could add value is on top of that reporting layer. It could interpret the data, identify changes, surface anomalies and potentially recommend what I should look at next.

That is a much more interesting use case than simply using AI for the sake of using AI.

It has also made me revisit older automation techniques. Google Ads Scripts are a good example. They have been around for years, but looking at them again through the lens of newer AI models raises a different question:

What happens when traditional automation is combined with the reasoning capabilities of modern LLMs?

The AI does not necessarily need to perform every action.

It can potentially make or recommend the judgement, while an established and predictable automation performs the action. That hybrid approach is becoming increasingly interesting to me. It also feels safer. I don’t particularly want an autonomous agent making unrestricted changes to advertising campaigns because it has decided it knows best. An agent that reviews information, makes recommendations and hands them back to me for approval is much easier to trust.

Over time, specific low-risk decisions could potentially become automated as confidence in the system increases.

Agents don’t have to replace automation

This has probably been my biggest takeaway from the week. I had been approaching AI agents as another technology I needed to learn. Now I’m starting to see them as another layer that can sit alongside tools and processes that already work. Sometimes a basic automation is all that is required. Sometimes an existing tool such as Power BI is already extremely good at solving the problem. Sometimes AI can improve one step inside an otherwise predictable workflow. And sometimes there is genuinely a job that requires enough judgement and adaptation to justify giving an agent some autonomy.

The important part is recognising which one you are looking at before you start building anything.
Otherwise, it feels very easy to build something complicated simply because the technology is new and exciting.

A simple question I’m starting to ask

When I find a repetitive or frustrating task now, rather than immediately asking:

“Could an AI agent do this?”

I’m trying to ask:

“Does anything actually need to make a decision here?”

If the answer is no, I should probably investigate automation first.
If the answer is yes, the next question becomes whether that decision can be expressed reliably using normal rules.

Only when the answer starts becoming “it depends” does an AI agent begin to look genuinely useful. I’m still very early in learning all of this, so I fully expect my understanding to evolve. But ironically, trying to learn how to build AI agents has already taught me something useful before I’ve properly built one:

Not everything needs an agent.

And understanding what doesn’t need one might be one of the most useful places to start.