It’s never been easier to outsource your marketing work to automation. Advertising platforms will manage bidding, targeting and creative. CRMs will score leads, trigger workflows, and suggest next action. You can transmit performance data to an AI assistant and get optimization recommendations before the coffee is finished.
None of this is exaggeration. These systems actually work towards the goals they are assigned, autonomously and continuously, and the suggestions become increasingly useful.
But notice what hasn’t changed in that sentence: the goals you give him.
Automation doesn’t set a vague goal. A vague goal assigned to a person usually produces messy results and unclear direction.
Your customers search everywhere. Make sure your branding introduces himself.
The SEO toolkit you know, plus the AI visibility data you need.
Start your free trial
Start with
When automation does exactly what you asked
Give the AI a vague goal and you’re likely to get overly confident results and direction. The system will find the most efficient route in the wrong direction.
Demand a higher ROAS and automated bidding will happily lean on branded search, warmer prospects and retargeting. People who would have bought anyway. ROAS goes up.
Request more signups and your campaigns may fill the funnel with low-intent volume that never fires. Enrollments are going up.
Request a lower CAC and the system will silently reduce your reach to the easiest audience you have. CAC drops.
In any case, the metric improves, but the company may not. The automation did exactly what you said. It just wasn’t doing what the company needed.
Stop giving direction to automation. Give him a field.
“We need higher ROAS” is not a goal. It’s a direction. An optimizer will follow one direction forever, even past the point at which it stops helping the business.
What automation actually needs is a playing field. Clear lateral lines on both sides. What counts as a win and, just as explicitly, what counts as a loss, even if the metrics look good.
Take a brand running paid campaigns with 8x ROAS. Leadership wants growth, so the real goal is not to protect the 8x. It is acquiring more and more new customers. Put correctly, the goal sounds like this: We will accept ROAS to drop from 8x to 5x if the volume of new customers grows with it. Below 5x we stop and reevaluate. That’s the floor.
Now AI has room to move. It can expand the audience, chase incremental customers, and spend in less efficient territories, because someone decided in advance how much efficiency the company is willing to trade and what the limit is. Without that plan, you get one of two failure modes. Either the team strangles campaigns by protecting a ROAS number that no one actually needs, or the system continues to spend without any agreed upon stopping point.
The skill isn’t just in choosing the metric. It is to define both sidelines before the start of the match.
Decide what to disable before activating it
The same thinking applies to new AI capabilities within the platforms themselves, and this is where I see teams skipping tasks.
Consider an advertiser in a regulated industry, such as insurance, who turns on Google’s AI Max. The default posture is to enable everything and let the system optimize. But for that advertiser, the loss conditions include things that no dashboard will ever report. The AI rewrote the carefully vetted ad copy into something that was never approved by compliance. Brand terms are placed in expanded correspondence where they don’t belong.
So the right move is to decide what to disable before enabling anything. Text personalization disabled. Brand excluded. Then let the AI work hard on what’s left.
This is not distrust of technology. It’s the opposite. It’s giving the system a field it can run on. Guardrails are what make autonomy safe enough to actually use.
Automating a hypothesis is still a hypothesis
Another version of this, from the CRM side, because it’s easy to assume that the problem lies only in paid media.
It’s very easy to create elaborate automated workflows. Activate this email, assign this task, push the customer towards this action. The question that is rarely asked is whether there is data to show that customers who take this action actually stay better. Many workflows are engineering efforts overlaid on an assumption. Automation works. The hypothesis has never been tested.
If you wouldn’t let a person do the task manually because you can’t tell what makes them better, automating it doesn’t make them smarter. It simply causes the guess to run on a schedule.
Where humans go
None of this is an argument for less automation. The tools are good and getting better, and refusing to use them is a risk at this point.
It’s a discussion about where human judgment now gains ground. Don’t approve every offer change or review every suggested action. Human work is owning the definition of the field: the limits, the exclusions, the compromises the company will accept and the advantages that don’t matter. So pay attention to the moments when the system wins the metric but loses the game.
Automation will hit whatever target you give it. This is exactly why the goal deserves more thought than it receives.
So here’s a question worth dwelling on: If every automated system you manage hit its number this quarter, how many of those successes would move the company forward?
