Why Cruise Lines Are Overrating AI Models and Underrating Their Crew

The Model Isn't the Problem

Three months back, I got handed a project to build an AI review system for our cruise line's onboard incident reports. I'm not a coder—I came from guest services. Two months in, the system was live, the flow was manageable, and the results were surprisingly decent.

A colleague from another ship asked me the other day: "Which model are you using for that auto-review thing?"

I said, "Honestly, they're all about the same."

He blinked. "Then what did you do for two months?"

"Talked about operations, argued with people, and stared at logs with the AI."

He went quiet. I knew what he was thinking. In most people's heads, building an AI system looks like: pick the strongest model, write a magic prompt, fine-tune on high-quality data, ship it.

That's the biggest illusion of this AI wave. Especially in an industry like cruise lines, where the real work happens in the dining room, at the purser's desk, and in the engine room—not in a data center.

AI Is Already Ahead of Most Cruise Applications

McKinsey estimates that current top-tier models can already automate or significantly enhance 60–70% of white-collar tasks. Dario Amodei and Sam Altman have said similar things. Translating that to our world: the AI capabilities we have access to could technically cover 80% of the traditional back-office work on a cruise ship—things like booking adjustments, excursion scheduling, guest feedback routing, and compliance checks.

But look around. How many applications are actually replacing human tasks on a daily basis? Most of the stuff teams build—little websites, dashboard widgets, weekly report generators—barely scratches the surface. Some of it is solving problems that don't even exist.

During my project, I watched AI applications die in predictable ways:

• The demo works, but production falls apart. Everyone claps at the meeting, then the system spends three months fighting field changes.
• Perfect on three test cases, collapses in the real world. The edge cases show up faster than you can log them.
• Chasing models. Switching from GPT to Claude to Gemini, but the business process doesn't move an inch.
• Calling it AI when it's really just a glorified filter. The AI guesses, a human reviews, then a customer service rep replies. That's not AI; that's a more expensive workflow.
• Leadership kicks off with enthusiasm, but nobody owns it. The project dies from lack of ownership.

Notice a pattern? None of these failures had anything to do with model strength. The bottleneck wasn't the AI. It was the application design. And underneath that, it was the ability to articulate business rules clearly.

The Real Shortage: People Who Can Explain Their Business

My first version of the review system was intuitive: feed the incident log to an LLM, give it a detailed SOP, let it output a decision, and keep feeding it data to refine. Accuracy? 60%. The other 40% of cases all pointed to the same issue—the model wasn't short on intelligence. It was missing context.

Here's a real example. A complete authorization document listed Company A as the rights holder and Law Firm B as the agent. Format was fine, stamps were all there. But the complaint form had the parties reversed—the filer put Law Firm B as the rights holder.

The model saw the mismatch and rejected the case. Wrong call.

In real operations, people make mistakes like this all the time. A human reviewer who's seen fifty similar cases would catch it instantly. The model couldn't, because we never told it what the underlying chain of authorization looks like.

So I spent a week building a dictionary of all the major agencies, their common aliases, and their historical authorization patterns. When a new incident came in, the system checked that table first. If it matched, it got a green light. That one table—no fancy tech, just pure business knowledge—boosted accuracy by 12 percentage points.

I built three more tables like that. Total code: under 500 lines of Python and JSON. And honestly, I didn't write a single line of it myself. I explained the business rules to the AI, it wrote the code. Then I'd point out where it was wrong, and it would fix itself. We went from 60% to 90% that way.

The lesson: AI can write code. AI cannot write business rules. Those are the hard-won instincts you get from actually working in the field. They're not in the training data.

Maturity Means Knowing What the System Won't Do

Getting from 90% to 97.8% was about engineering stability. I stepped on every common-sense landmine you can imagine, and the AI kept pulling me out of the mud. I won't list all the mistakes here—you have to step on them yourself to really get it. My developer colleagues laughed at me more than once.

But the moment I knew the system was mature wasn't when accuracy crossed 97%. It was when I sat down and wrote a list of things the system would not do.

That list included:

• No AI review for international incidents. Low volume, high stakes—a single misjudgment could trigger a lawsuit or a diplomatic headache.
• If the AI says "reject," a human must review it. If it says "pass," it can go through automatically. Because a missed review is a recoverable oversight, but a wrongful rejection is a permanent damage to guest trust.

Writing that list changed how I think about AI systems. A mature system isn't just about what it can do. It's about being explicit about what it refuses to do. And that judgment call—deciding what not to automate—only comes from someone who's been on the ground, sweating through the edge cases.

A New Kind of Role in the Cruise Industry

Put the three lessons together and you get a new kind of professional: someone who's deep in the business, knows how to leverage AI, and respects its limits.

Business immersion lets you tell the difference between a model failure and a context gap. AI application skills mean you don't have to write code, but you know when to let the AI run and when to make it stop and call in a human. And boundary awareness lets you write that "do not do" list, keeping the system effective within its real capabilities.

This person doesn't fit the traditional org chart. They're not a product manager—they actually get their hands dirty pushing the system into production. They're not a software engineer—they don't write code. They're not a data scientist—they don't fine-tune models.

They're a hybrid: a business person with an AI copilot.

The old path was build a product, then bolt on AI. The new path is soak in the business first, then use AI to fill the technical gaps. That's the role with the most leverage in this wave of personal AI adoption.

Are You This New Species? Three Questions

Question 1: Can You Explain How a Task Is Done Without AI?

If you can't walk through the manual process step by step, you're not building an AI application. You're just riding the AI hype. If you can, you're already at the starting line.

Question 2: When You Hit an Engineering Wall, Do You Grind or Ask AI for Help?

If you're stubbornly debugging logs by yourself, you're wasting time pretending to be an engineer. If you paste the log into the AI and ask what's going on, you're using AI the right way—to cover your technical blind spots.

Question 3: Can You Write a List of What This AI Should Not Do?

If you can't list the boundaries, you haven't spent enough time in the business. If you can, you've crossed the threshold that most people never reach.

If you answered all three with "yes"—you can explain, you ask AI for help, and you've listed the don'ts—then congratulations. You're the new species. Update your resume and start looking for a role that values that mix.

The Model Was Never the Point

So back to the colleague's question: "Which model are you using?"

Today, I can answer more completely: the model was the least important part of the project. Before you've nailed down the business rules, closed the engineering loop, and agreed on boundaries, no model in the world will save you. Once you've done those three things, almost any model will work.

In the cruise industry, where guest satisfaction is everything and a bad call can ripple through the entire voyage, the real competitive edge isn't the AI. It's the people who understand the nuances of onboard operations, who are willing to sit with the AI at 2 a.m. debugging a weird log, and who take the time to write down what the system should never do.

Those people are the quiet ones, but they're the ones who'll sail furthest in this AI-driven era.