Telling leadership their system is broken

I took an organizational behavior assignment for an NYU class earlier this year. The client was a midsize creative agency, Global Growers, that wanted to fix retention by launching a mentoring program. Leadership had already decided what they wanted; the assignment was to recommend how to build it.

Before recommending anything, I gathered feedback from three groups: leadership, current employees, and former employees. The data didn't agree with leadership's assumption.

The diagnosis

Leadership thought a mentoring program would solve retention. The data pointed somewhere else.

Current employees said they didn't have career-development conversations with managers. They felt siloed across teams. Conference attendance was supported on paper but funded by the employees themselves.

Former employees were more direct. They'd left because they "did not find opportunities for growth." Managers "prioritized completing tasks instead of developing talent." Weekend work was widespread; burnout was the actual driver.

A new mentoring program wouldn't, by itself, fix any of that. The deeper problem was a weak development system, not a lack of employee ambition.

The challenge

The recommendation was easy to write in my head: the mentoring program will not succeed if leadership treats it as a stand-alone solution. Three structural changes were needed alongside it: formalize the mentoring program, make managers actually have career conversations, address the burnout culture.

The hard part wasn't the analysis. It was the delivery. Leadership had a clear belief about what was wrong and a clear preference about what to do. Telling them their broader development system was broken, without making them hear "you failed," was the actual project.

This is where I started using Gemini.

What I asked the AI to do

The first prompt was honest about the audience: I'm preparing a recommendation for leadership that may not be open to criticism. Help me improve the messaging for an audience that could be defensive.

Gemini's first revision was over-softened. It identified four "red flags" in my draft and rewrote them using what it called "Partnership Language." The output was diplomatic and polished. But it had drained the recommendation of its content. "We have an opportunity to integrate mentoring into a more cohesive development framework" reads like a consulting deck, not a recommendation. The hard message was gone.

What I had to do

I pushed back. The second prompt was more specific: revise so it sounds diplomatic and leadership-friendly, but still specific and evidence-based. Keep the core concerns about inconsistent development, siloed work, weak internal advancement, and burnout. Do not make the message vague.

The second pass was the version that worked: "the proposed mentoring program could improve retention, but its impact will be limited unless it is supported by a more coordinated employee development strategy."

The shift was small in word count, large in effect. The first version said leadership's approach won't work. The second acknowledges their intent while redirecting. Same evidence, same conclusion, different door to walk through.

The takeaway

AI over-softens by default. Gemini's first instinct was to eliminate friction entirely. Useful for a first pass; the wrong endpoint for a real recommendation. The human's job is to push back when the AI optimizes for politeness over substance.

Two prompts beat one. Iterative prompting, telling the AI exactly what the first pass got wrong, produced a significantly better result than any single prompt would have.

The original analysis was mine. The argument, the evidence, the framework citations, the diagnosis. All human work. Gemini's role was to pressure-test the tone for a specific audience. That's where AI adds the most value: refining human work, not replacing it.

The deeper lesson is about the parallelism. The AI's job mirrored the organizational problem itself. Both required navigating the gap between what needed to be said and what people were ready to hear. The skill, for the consultant, for the AI, for anyone using AI in advisory work, is knowing how to thread that gap without losing what you came to say. I wrote earlier about AI as scaffold, not authority. This case is the same lesson from a different angle: AI as revision partner, not author.