What a marketing experiment taught us about becoming AI-native.
At Causeway Solutions, we've been exploring what it means to become an AI-native company. For us, that means integrating AI into how we work, make decisions, and get things done.
We've already seen how AI can help us move faster, work through complex information, and take on projects that would otherwise require more time or resources. But as we keep experimenting, we're also learning that just because AI can do something doesn't mean it's the most efficient way to do it.
Our marketing team recently got a firsthand lesson in that distinction.
We tested whether an AI agent could help manage our digital advertising. The assignment was straightforward: review our existing Google and LinkedIn campaigns, identify opportunities for improvement, and make the necessary changes.
In some ways, the experiment was a success. In others, it raised questions about what we actually gain when we hand work over to AI.
Where AI proved its value
The first part of the experiment went surprisingly well.
The agent reviewed our advertising accounts, identified problems with conversion tracking, uncovered inefficient spend, and helped build a more focused Google Search campaign. To do that, it worked across Google Ads, Google Analytics, Tag Manager, HubSpot, and our website to understand how everything connected.
That cross-system view mattered. Marketing platforms don't operate independently, and a problem in one system can quietly distort how performance is measured in another.
For example, the agent found that a campaign built for Google Search had been spending most of its budget outside search results entirely. That wasn't obvious from any single dashboard.
The agent also helped with the strategy behind the new campaign, from keyword research and targeting to ad messaging and conversion goals. Instead of starting from scratch, our team had a more informed foundation for every decision.
If the experiment had ended with the audit and recommendations, we would have called it a clear efficiency win.
But we wanted to see whether AI could go beyond identifying the work and actually complete it.
When automation creates more work
Moving from analysis to execution is where the experience changed.
What seemed like a straightforward process became a series of browser interactions, permissions, approvals, troubleshooting, and verification. The agent would navigate to a setting, make a change, check that it worked, and report back before moving to the next step.
Sometimes that worked exactly as intended. Other times, the browser stopped responding, a permission needed to be granted, or a task hit its usage limit before it was finished. Each interruption meant another round of instructions, troubleshooting, or waiting.
To be fair, some of the work genuinely required that level of care. Fixing conversion tracking across connected platforms isn't something we'd want to rush, and verification matters when changes affect performance and reporting. The agent made a few confident errors along the way, including flagging a tracking tag as duplicated when it wasn't. It caught and corrected each one, but only because every change was verified.
Other tasks were much simpler. Updating campaign settings, reviewing ads, and making routine edits inside familiar platforms were things our team could have done directly, often in less time. Adding a single line of copy to our landing page took the agent two attempts, both of which had to be discarded, before a member of our team made the change by hand in about a minute.
We realized that while AI was doing the work, we were spending considerable time managing the process. Instead of eliminating manual effort, we had introduced a different kind of effort: supervising the automation.
That raises a question worth considering beyond this experiment. If AI can complete a task but requires more time and oversight than doing it manually, is it actually making us more efficient?
The costs we don't always measure
When businesses talk about AI efficiency, the conversation usually focuses on how much work a tool can accomplish. But completing a task is only one part of the equation.
There's also the time spent writing instructions, monitoring progress, reviewing outputs, troubleshooting issues, and correcting mistakes. For agents that interact with software, those steps can become a significant share of the workflow.
Then there's the cost of the tools themselves.
During our experiment, we hit a usage limit that interrupted the work before everything was finished. We can't attribute that limit to this project alone, but it was a reminder that AI execution consumes resources, particularly when tasks involve repeated browser interactions across multiple systems.
None of this means AI isn't worth the investment. The issues the agent uncovered had real value. But judging that value requires looking beyond whether a task was technically completed.
A useful measure of AI efficiency weighs three things: the quality of the outcome, the total time required to achieve it, and the resources consumed along the way.
A process that produces a better result may be worth the extra effort. A process that delivers the same result with more time and supervision deserves a closer look.
What we'd do differently
Looking back, we wouldn't remove AI from this project. We'd change where we used it.
The most valuable parts involved analysis, problem-solving, and connecting information across systems. AI helped us find tracking issues, evaluate performance, research keywords, and develop recommendations that gave our team a much stronger starting point.
Those are capabilities we'd absolutely use again.
Where we'd change our approach is execution. For straightforward changes inside platforms our team already knows well, doing the work manually is often faster and more practical. For complex technical work, especially changes that span multiple systems or unfamiliar configurations, AI assistance may still be worth the extra oversight.
The distinction isn't between AI and human work. It's about deciding which parts of a process benefit most from each.
AI might analyze a campaign, identify where money is being wasted, recommend a better structure, and draft the messaging. A marketer can then review those recommendations and make the changes directly. The result is still an AI-supported workflow, even if AI isn't responsible for every click.
In many cases, that combination delivers more value than automating the entire process end to end.
Becoming AI-native means knowing when to use AI
As we've shared throughout our Building What's Next series a few months back, Causeway's approach to AI is about expanding what's possible for our team: working through more information, making better decisions, and building capabilities that would be hard to develop through traditional processes alone.
That hasn't changed.
But becoming AI-native also means being willing to question how we use the technology. It's easy to assume that the more we automate, the more efficient we become. In practice, it isn't always that simple.
Some workflows benefit enormously from automation. Others require so much supervision that the expected time savings disappear. And many fall somewhere in between, where AI provides the analysis and direction while people handle the execution.
We don't think the answer is to stop experimenting. Experiences like this are exactly why experimentation matters. They move us past what AI can do in theory and show us how it performs in a real business.
Not every experiment needs to end in full automation to be worthwhile. Sometimes the most useful outcome is learning where automation belongs in the first place.
Becoming AI-native isn't about using AI for every task. It's about learning which tasks are worth giving it.
For our marketing team, that may mean letting AI do more of the thinking and a little less of the clicking. And if that gets us to a better result in less time, we'd call that progress.