In Part 1, I wrote about how we redesigned Consulting in Hospitality & Tourism at Nova SBE around a simple assumption: AI already exists, students are already using it, so pretending otherwise is the least productive option.
That was the professor's point of view.
This is the other side.
At the end of the course, I asked all ~40 students two things: to reflect, in their final exam, on how their group used AI throughout the consulting project — and separately, in a quick feedback form, to just list it. Bullets. No essay. No performance.
Reading the two side by side was the most useful thing I did all trimester.
Because the exam answers were thoughtful and a little rehearsed — written for a grade. The feedback form was raw. And the raw version is where the truth lives.
Here is what I found.
1. They Stopped Describing AI as a Tool That Gives Answers
The most striking pattern was linguistic.
Nobody — almost literally nobody — described AI as something that did the work for them.
Instead, completely unprompted, they kept reaching for the same words: sparring partner. Co-pilot. Personal assistant. Accelerator. Force multiplier. Intermediary. Sounding board. Devil's advocate.
One student put the whole philosophy in a single sentence: "AI was a multiplier rather than a substitute for our own work."
Another: "It helped us work faster and communicate more clearly, but it didn't tell us what to think."
This matters. Nobody instructed them to keep AI in its lane. They arrived there by doing the work. The boundary between human judgment and machine assistance was discovered, not imposed.
That is a very different outcome from the one most people fear when they imagine students with unlimited access to AI.
2. The Killer App Was Not Analysis. It Was Packaging.
If you ask people what students use AI for, they assume: writing essays, generating answers, doing the thinking.
The data says something more boring and more interesting.
The single most common use, mentioned by almost everyone, was slides. Structure, layout, action titles, and visual consistency across a seven-person team working in parallel.
After that: rephrasing and polishing. Summarizing the research they had already gathered. Translating sources across Spanish, Portuguese, Italian, and German. Proofreading. Checking consistency across five deliverables.
In other words, AI absorbed the time-consuming, low-judgment layer of consulting work — the formatting, the synthesis, the cleanup — and gave students their hours back to spend on strategy.
One student described it almost exactly as a consultant would: AI is "pretty risk-free" on tasks where you can validate the output instantly, like summarizing and polishing. It is dangerous precisely where you cannot — sourcing facts.
The judgment was already in the room. AI just removed the friction around it.
3. A New Workflow Appeared on Its Own: Using AI to Prompt AI
This one I did not expect.
A recursive habit showed up across multiple groups, in nearly identical language:
"Asking ChatGPT to write me prompts to ask Gemini to generate images."
"ChatGPT as an intermediary, so I could then ask Claude, Gemini, NotebookLM what I really wanted."
One student went further and used a formal prompting structure (that I taught in the first class) — context, role, action, format, tone — and "trained" her AI to remember her preferences across the whole term so she could get to good outputs faster.
Students are no longer just prompting. They are building prompt pipelines. One model to structure the request, another to execute it. A year ago, that vocabulary did not exist in a classroom. This year, it spread across groups without anyone teaching it.
This is exactly what I meant in Part 1 when I said prompt engineering is becoming a form of structured thinking. You can now watch it happening between students, not just between a student and a model.
4. They Built Things They Had No Training to Build
This is where it got genuinely exciting.
One student programmed an AI-built web scraper to collect every Google review for every spa Vila Galé owns — then used AI to run a correlation analysis in Excel he did not know how to do manually, which surfaced what actually drives a spa's performance.
Several teams built functional, two-sided app prototypes with no one on the team who could code. As one wrote, "none of us had programming skills," but they wanted the client to see the recommendation, not just hear it.
Another student built a reusable "artefact" to hold the project's context across all deliverables and brief the client's CEO — updating it with feedback instead of starting from scratch each week.
The ceiling on what a motivated student can produce has moved. Not a little. A lot. The constraint is no longer a technical skill. It is knowing what is worth building.
5. The Better the Student, The More Honest They Were About the Limits
Here is the pattern that should reassure every professor.
The strongest students were not the ones most impressed by AI. They were the most critical of it.
The limitation named most often, by far, was the need to fact-check everything. Right behind it: hallucination — fabricated sources, invented data, confident nonsense.
And the best catches were specific. Two students, in different groups, independently caught the same hallucination: the model claimed there was room to build a new hotel on a particular Cádiz beach. Both checked Google Maps and Booking themselves. Both found the same thing — the segment was already overcrowded, no room to expand. "It sounded a bit too good to me, so I went and checked."
One student named the deepest problem of all — the one the whole industry is quietly struggling with. When you ask AI to challenge your thinking, it tends to agree with you instead. The "yes-man" effect. A tool sold as a critical sparring partner that has a structural bias toward flattering you.
The students who notice that are precisely the ones who will be good consultants.
This is the same thing I described in Part 1 from the teaching side: AI does not eliminate rigor. It exposes where the rigor was fake. The exam answers proved it from the student side. The ones who deeply understood their projects used AI with skepticism. The ones who did not use it with relief.
6. They Also Discovered the Limits of AI as a Team Sport
The most forward-looking observation in the entire cohort came from one student, almost as an aside.
His group's biggest AI problem was not the model. It was that each member used their own separate chat. The outputs diverged. Context did not carry. Token limits and mismatched subscriptions blocked a shared workspace. His attempt to build one shared artefact for the team simply failed.
His conclusion, written on a feedback form: he will use a shared setup next time.
I am highlighting this because it is a year ahead of where most organizations are. We have spent two years asking whether individuals can use AI well. The next question — the one this student already hit — is how teams coordinate around it. When AI goes multiplayer, the bottleneck stops being the model and becomes the humans organizing themselves around it.
That is not a student problem. That is the next problem for every company reading this.
7. The Honest Minority: Over-Reliance Is Real
I will not pretend it was all sophisticated.
A couple of students reported finding essentially no limitations to AI. Set against the overwhelming majority who fixated on fact-checking, that reads less like confidence and more like thinner engagement.
And the sharpest line in the whole set was a warning, written by a student about his own peers: there were cases where people "relied too much on AI, instead of thinking certain things through by themselves." His verdict: AI worked best when it amplified thinking, and worst when it replaced it.
That is the exact risk I flagged in Part 1. AI can amplify thinking. It can also amplify intellectual passivity. The difference becomes visible the moment a student has to stand up, without their slides, and defend the logic.
What This Changed in How I Think About Teaching
Three things stayed with me.
First, the skill has moved up the stack. When generating a draft, a slide, a translation, or even a working prototype is essentially free, the differentiator is no longer production. It is judgment — knowing what to ask, what to trust, what to verify, and what to keep stubbornly human. Our grades tracked that almost perfectly.
Second, students will tell you the truth about AI if you ask them the right way. The exam gave me their polished reasoning. The two-line feedback form gave me their real workflow — emojis, frustrations, and a thank-you for a suggestion that did not even work. Both were necessary. Neither alone would have been enough.
Third, and most importantly: this generation is not trying to cheat with AI. They are trying to work with it, and they are quietly building the norms for how to do that as they go — augmentation over automation, verify before trusting, keep the strategy human. Nobody handed them those rules. They wrote them themselves, one deliverable at a time.
Our job is not to protect students from AI.
It is to put them in situations — live presentations, real clients, hard questions — where the difference between using AI and thinking becomes impossible to hide.
That is not a threat to education.
It might be the best thing that has happened to it in a long time.
← Back to Part 1 How we redesigned the course around the reality that AI already exists.Originally shared on my Substack →