Why AI dies on the deck (and how to stop it)

Most AI pilots look flawless in the boardroom and fail in week one of field use. We've seen the same pattern on vessel decks, plant floors, and remote sites across Indonesia — and the fix is almost always the same.

The pattern

It starts with a great demo. The AI reads documents perfectly, triages defects instantly, and generates beautiful job cards. Leadership is convinced. The pilot begins.

Then reality arrives. The crew on the deck has five seconds between tasks, a phone with a cracked screen, and no patience for a form with twelve fields. The connectivity is patchy. The promised "effortless capture" requires a two-minute workflow that nobody has time for. Within two weeks, the app sits unused. Within a month, the pilot is quietly declared "successful" — because the dashboard said so — while the field never adopted it.

The AI didn't fail. The interface between the AI and the human did.

Why it keeps happening

AI projects are usually designed top-down: start with the model, work down to the user. But the model doesn't capture data — people do. When capture is the bottleneck, everything downstream is fiction.

We see three recurring design sins:

• Forms pretending to be "AI capture." If the crew still fills in fields, you've just added an AI report at the end of the same old admin.
• Online-only workflows. Designed in a Jakarta office, deployed where the internet is a rumor.
• No human in the loop. The system acts invisibly, so nobody trusts it — and trust is the entire adoption model.

The fix: design from the deck upward

Start with the person doing the work, then build the AI around them. Three principles we build by:

• Capture must be a photo, not a form. The crew's job is to photograph and move on. OCR, classification, and record-building happen after — invisibly.
• Offline-first. Everything works with no signal. Sync happens when the phone finds one. The field should never notice the network.
• Humans approve everything. Every AI output gets reviewed, corrected, and owned by a person. The system learns from corrections — and earns trust one approval at a time.

The honest test

Before you buy any AI pilot, ask one question: what does the field user do differently in their first five minutes? If the answer involves more work, the pilot will fail — no matter how good the model is.

We build for the deck, not the demo. It's slower to demo and faster to adopt — and adoption is where the value actually lives.