With great power comes great responsibility. More than a memorable movie line, it's an allegory for modern marketing technology. We have more data available than ever before, and thanks to AI developments, more powerful ways to examine, visualize, analyze, and activate it. You can surface the most granular information and have insights and recommendations at your fingertips, anytime, anywhere.
AI gives you superpowers, but like our CEO Tom Zawacki pointed out recently, "Most superhero stories end the same way: not with the superpowers saving the day, but with the person with them making the right call at the right moment."
As marketers, rising to meet the moment starts with understanding AI's weak spots and cultivating the judgment skills that allow us to spot and work around them to advance our work (rather than replace it).
Why do AI marketing recommendations fall short?
There's a lot to love about AI. It offers instantaneous data insights that you would've spent days or weeks digging for. But those same capabilities can create more complexity for teams. With so many AI features available in media platforms and dashboarding tools, it takes a trained eye to filter what's important, what's not, and what action to take. You need human expertise and discernment to cut through the noise and get the critical, must-have insights you actually need.
We've seen it firsthand:
-
In one case, we had a client using an MCP server to chat with an AI agent for data-driven recommendations. The tool had the right idea, but it recommended creating an FAQ schema for the client's website without ever crawling the site or specifying which pages actually needed it.
-
Another time, our client’s AI dashboard gave us a frequency optimization recommendation for Meta that couldn’t be implemented given the platform’s thresholds.
Directionally, both recommendations were spot on, but the AI missed the finer details that made them non-starters. It was humans, not tech, that caught the issues.
And these are just a few of the countless examples of industry best practices and nuances that AI tools don't, or can't, account for without human direction and validation. Left to its own devices, or in the hands of someone without the keen eye for inconsistencies, your campaigns can go off the rails in favor of the promise of efficiency.
How to humanize and improve AI-driven analysis
To make AI work for you, you need to give it the right context, drawing on your own expertise and experience, and guide it from there. That starts with giving your AI tools the right prompts to ensure they're delivering the specific information you need to make the most informed and impactful optimizations.
The term "prompt engineering" is tossed around a lot, but when employing AI, it's a critical component. AI can deliver countless insights in a snap, but it's the human guidance that makes them meaningful.
Let's take the example of building a dashboard for reporting and optimizations. An instinct might be to include every available dimension and metric your stakeholders might care about. Knowledge may be power, but quality trumps quantity. Here's how you can get to the insights you need using human-led AI power.
Define your questions
Like any marketing initiative, you need to start with your business goal. Say it's reducing your cost per acquisition. Instead of pulling every available data point, some of which may be irrelevant or even misleading, into your dashboard, you can ask AI to sift through it all and identify which dimensions and metrics you need to evaluate to reach that specific goal.
Keep in mind that it’s not enough to simply feed the AI tool your goal. Defining your question is one of the most important steps of the entire process. Give as much detail as possible when crafting your prompt to get a highly relevant answer. Include context about your strategy and how it’s evolved, along with other relevant background information. If you don’t, you risk a generic output.
Then review the reply and ask it how you should look at the data and use it. Starting from goal-aligned questions helps you zero in on the dashboard data and the follow-up questions that will drive the most impactful optimizations.
Refine the recommendations
Once you've nailed down your dashboard data, AI can deliver contextually relevant optimization recommendations. But here's where having a human in the loop becomes critical again. You need to decide what's actionable based on your organization's needs. Consider questions like:
- What will work best given your team's bandwidth?
- Do you have the budget to deliver on it?
- What might make sense as a quick win for eager stakeholders?
- Will it move the needle enough given internal pressures?
Your AI tools and agents have the power to deliver information and insights, but they need human initiation and guidance to move from data to insights, and finally, to action.
The AI-ready marketers' skillset
It's clear AI is changing marketing processes, and, in many cases, marketers' roles. To adapt and prove your human value, it's important to become adept at two in-demand skills: prompt engineering and error detection.
Prompt engineering and error detection as core analyst skills
If prompt engineering and error detection sound like the two ends of the same conversation, that's because they are. One is how you talk to AI. The other is how you talk back.
What does a strong marketing analysis prompt include?
Your AI output is only as good as the question and the context you give it. Vague prompts get generic answers. Specific prompts that spell out your business goal and constraints, paired with detailed markdown files, get you something you can act on.
So while a beginner prompt might look like "How do we lower CPA?", an expert prompt is much more robust and contains:
- A role you assign the AI
- Context about your business and what you're trying to achieve
- Your relevant campaign data
- Your program, platform or campaign constraints
- What you need your AI to deliver
- The format you want the answer to be in
Here's what that could look like:
Role
You're a senior performance marketing analyst specializing in paid search efficiency and Search Ads 360. You understand attribution methodology differences across platforms and will flag any assumptions you make about the data before drawing conclusions.
Business context
We're a B2B SaaS company selling a project management tool with a 30-day free trial. Our target customer is a mid-market operations manager. Average contract value is $8,400/year. Current blended CPA target is $420. We're 22% over target this quarter.
Campaign data
Attached: campaign_performance_Q3.csv
- Platforms: Google Search (native), Search Ads 360 (bidding on Google plus Microsoft Advertising inventory), Meta
- Date range: July 1 to August 15, 2026
- Columns: platform, campaign, engine, spend, clicks, conversions, CPA, conv_window, attribution_model, bidding_strategy
Known measurement constraints: treat these as hard rules
- Google Search (native) is reporting via Google Ads with a 30-day click plus 1-day view attribution window.
- Search Ads 360 is reporting via Floodlight tags with a 30-day click window; view-through isn't enabled, so don't assume parity with Google Ads conversion counts for the same campaigns.
- Meta is using a 7-day click plus 1-day view attribution window.
- Search Ads 360 Smart Bidding is optimizing to Floodlight conversions; Google Ads campaigns in the same account are optimizing to Google Ads conversions. These are not the same signal.
- Conversion events are NOT deduplicated across platforms, so cross-channel overlap likely exists.
- Microsoft Advertising traffic within Search Ads 360 should be segmented separately from Google traffic before any efficiency comparison.
What I need
- Identify the top three drivers of CPA inefficiency, controlling for the attribution and bidding signal differences called out above. Do not compare platform or engine CPAs at face value.
- For each driver, recommend one specific, testable lever we can pull in the next 30 days. Search Ads 360 bid strategy settings, audience layering, and budget allocation within Search Ads 360 are all in scope.
- Flag any part of the data where you cannot make a reliable recommendation without additional information, and tell me exactly what information would resolve it.
- Don't recommend budget reallocation across platforms until the attribution inconsistency is resolved. Focus recommendations within each platform and engine.
Output format
Return your analysis in three sections: Findings, Recommendations, and Data Gaps. Keep findings to three bullet points max per driver. Each recommendation should include an estimated effort level (low, medium, or high) and a hypothesis we can use to measure success.
While creating a highly specific prompt like this requires more up-front work, it gets you a more thorough and relevant reply without unnecessary back-and-forth with the tool.
How to catch errors in AI output
Error detection is less about knowing the right process and more about knowing when something feels off. It takes the kind of pattern recognition that comes from experience. When looking at conversion data, AI has no way of knowing if one platform is counting view-through conversions while another isn't unless you tell it, so it'll compare the two anyway. The output will look perfectly confident while being quietly wrong. We’ve seen it countless times.
In one instance, a client's AI-powered dashboard showed that people were seeing a single Meta ad 11 times in a week before converting. AI took the number as fact. Our team knew it didn't add up, and dug deeper into the measurement practice behind it.
Know your data
Before leveraging AI for data analysis, ensure you understand crucial platform-specific nuances, including variations in reporting settings, lookback windows, and other technical limitations that differ between platforms. You can't develop a sharp question or spot a potentially incorrect answer if you don't understand what the model knows about your technical platform setup and your business.
All of these skills are best coached, and that coaching can start before a client or team member ever gets in a room with you. We recently sat down with a client for a Q&A session, and they came in with 15 to 20 questions, most of them foundational, like "how do you read this data card?" and "what does this metric actually mean?"
We encouraged them to use a chatbot to answer these questions first, then check whether the answers are accurate. This process either answers the foundational question outright, or it surfaces two or three sharper follow-up questions the person wouldn't have known to ask otherwise. Used this way, the chatbot helps you get to a better question, faster, allowing the collaborative session that follows to move past the basics.
What's next for AI-ready marketers
With the right skills in place, you can move to creating more AI value for your stakeholders. It could be implementing pre-built prompts and one-click executive summaries for them so they aren't left guessing what to ask. Or building custom agents around specific, repeatable skills, set up once and put to work every time that specific question comes up.
As AI evolves, the tools are getting more tailored, as well as more powerful. But as the routine, foundational analysis gets faster and more automated, that doesn't mean expert work disappears. Instead, the value marketers bring shifts as they spend less time pulling and formatting data and more time on the judgment calls only a human can make, like what's actionable, what's worth the effort, and what nuances the data story isn't showing.
Which brings us back to where we started. AI can hand you an insight in seconds: the anomaly, the frequency card, and the recommendation that looks airtight on the screen. But with that power comes the responsibility of what you do with it. Knowing how to ask the tool the right questions and catch it when it's wrong helps you make the right calls at the right moments.
We’re happy to help you hone your skills or apply ours to help you make a true human-led AI-powered impact. If you’re interested in learning how we can help, please reach out.
