A human being can approve an AI recommendation without adding much to the marketing decision behind it. The facts may have been checked, the presentation may be clear, and the suggested next step may sound sensible. But we still don't know how carefully the person considered whether it was the right decision.
In my latest edition of The Real ROI, I explored what we expect human judgment to contribute when AI is involved. I used our LEGO campaign experiment to examine the question. The largest behavioral multiplier pointed to an audience already enthusiastic about LEGO, while another finding helped me see an opportunity to reach parents who might not have considered it.
Evaluating human judgment in AI-assisted marketing decisions requires understanding how someone arrived at their recommendation. The assumptions they made, the alternatives they considered, and the evidence they found convincing all matter. Without that context, it's difficult to determine whether human review improved the decision.
How to Evaluate Human Review of AI Decisions
A September study in Human Systems Management looked at trust in AI decision support among 400 employees. Employees who reported greater trust in AI decision support also tended to report putting less independent effort into evaluating recommendations. The relationship was weaker among employees who reported higher levels of accountability pressure.
There are a few important caveats to consider. This was a survey taken at one point in time, measuring people’s perceptions of their own effort. The research doesn't establish whether AI weakens human judgment or whether accountability leads to better decisions. But it does raise the possibility that people evaluate recommendations more carefully when they're responsible for the outcome.
Simply attaching someone’s name to an approval or presenting a decision with confidence tells us very little about how carefully the recommendation was evaluated. Review becomes more useful when colleagues can see the assumptions being made and the alternatives being considered, including the evidence that would lead the team somewhere else.
For a marketing decision, that might mean explaining why an audience’s interest in a category creates an opportunity for this brand. Or it might mean acknowledging that the research describes existing customers well while leaving prospective customers less understood. The distinction changes what a campaign is being asked to accomplish.
Give People a Chance to Reconsider AI Recommendations
One practical option is to ask people for an initial assessment before they see the AI recommendation. A September article in AI Magazine examines this approach through exploratory navigation research. The samples were small, but the proposed sequence creates an opportunity to compare independent thinking with outside guidance.
The process should give people room to change their minds without feeling the need to defend their original assessment. Otherwise, asking for an early recommendation could simply create another position to protect.
Our October RADAR exercise gave us a chance to see how three marketers approached the same decision. Each of us made an initial recommendation before receiving the same audience intelligence. One narrowed her audience. Another found support for gifting while still needing more cultural context. I kept my original audience and developed the creative idea further. Looking at how we arrived at our decisions was more revealing than simply counting how many of us accepted or rejected the research.
Teams also need to be able to look back at a decision and understand why they made it. If a campaign disappoints, knowing that an executive approved it offers very little direction. Knowing which assumption the team was testing gives the next discussion somewhere to begin. The result may challenge the audience choice, the creative interpretation, or the way the test was designed.
My newsletter asks us to be more specific when we talk about judgment. For marketing teams using AI decision support, that means giving people room to explain their interpretations, question recommendations, and acknowledge uncertainty. The quality of that conversation is something a team can examine. An approval checkbox tells us only that someone reached the end of the form.
