top of page

Building Your Own Fleet Software With AI: Thirteen Questions to Ask First

1 hour ago
9 min read
Illustrative photo.

 

The argument comes up in almost every meeting, and it is a fair one: why pay for software when an AI can build you a bespoke tool in an afternoon? The demonstration is usually convincing. The tool exists, it speaks the company's own vocabulary, and it carries not a single feature nobody asked for.

 

A tool that did not work would be dropped within three weeks, with no consequence. This one works, so it will be adopted, and then written into the way people do their jobs. Adoption is what commits you, which is why thirteen questions are worth asking first. None of them is about the quality of the code.

 

The essentials

 

  • AI builds quickly and builds well. What it does not build is everything that surrounds a piece of software: permissions, support, training, updates, evidence.

  • A subscription costs a fixed sum, known in advance. A tool built on model calls costs in proportion to its use, so it costs more precisely when it succeeds.

  • Keeping software alive costs, on what we see across the deployments we support, 15% to 25% of what it cost to build, every year, in money and in staff time.

  • Who owns the code an AI produced, and who will still be able to touch it in two years: two questions rarely asked before the tool becomes indispensable.

  • The question is not whether it works, but who looks after it in eighteen months, and what it is worth the day somebody challenges it.

 

What AI genuinely does well

 

Let us be clear from the outset: we build with AI every day, and we would be poorly placed to advise anyone else against it.

 

It is excellent at turning an idea into a working mock-up, which beats twenty pages of specification. It handles in an hour everything that is disposable by nature: data migrations, file conversions, the clean-up that precedes an import. It is good at exploiting your own data, which gains value as everything else becomes commodity, since a twelve-year maintenance history exists in no model. And it automates very well on top of existing tools, which remains the most profitable route of all.

 

If your need falls into one of those boxes, build it. The questions below concern everything else.

 

The thirteen questions

 

What happens on board

 

1. What happens with no connection? An AI needs the internet to answer. In an engine room, in a hold, on a trawler thirty miles offshore, it does not answer. A logbook has to work offline, which is not a preference but a condition of use.

 

2. Who knows your trade? A general model will answer a straightforward question about minimum rest hours correctly. It will not know how your flag administration applies them in practice, nor how they interact with your collective agreement. That is not a shortage of computing power, it is an absence of context, and somebody has to write that context down.

 

3. Who trains the crew? An AI does not sit down beside a chief engineer to show him how to record an intervention, or to explain what the inspector expects from that register. The best-designed tool produces nothing if nobody uses it properly.

 

4. Who answers on a Saturday? Something jams on board on a Saturday in December, at sailing time. An AI does what it is told; it does not pick up the phone, does not call back, and does not own the problem until it is solved.

 

What becomes of your data

 

5. Who sees what? Several people working on the same data implies roles, permissions, and a record of who did what. A deckhand, a technical superintendent and an outside contractor have no business seeing the same information. That machinery is most of what makes professional software complex, and it is precisely what a prototype never has.

 

6. Where is your data? Which server holds it, under which jurisdiction, encrypted or not? Can you retrieve all of it at any moment, in a readable format? These questions look academic until the day you change supplier, or until the incident.

 

7. Who answers for the mistake? Models state false things with complete confidence, and that is a property of the technology rather than a passing defect. An AI guarantees nothing and is liable for nothing. If a maintenance plan omits a regulatory interval, the failure sits with the operator.

 

8. Who keeps it current? Attacks are automated now, and therefore continuous. A tool built in isolation becomes vulnerable the moment it stops being maintained, and that maintenance is a permanent line in the budget, not an installation task.

 

What matters once the tool becomes indispensable

 

9. Whose view of the work got built in? A tool built in an afternoon encodes how one person pictures the work: their vocabulary, their shortcuts, the way they sort things, their sense of what matters. It is often a good picture, because it comes from the field, but it is not the whole company's. The technical superintendent, the chief engineer, operations, payroll and purchasing do not carve the work up the same way or share the same priorities, and software settles those questions without saying so.

 

Speed has removed the stage where they used to be argued out. When you have to write a specification, compare offers and commit a budget, several departments speak up and the disagreements surface before go-live. When the tool is ready the next morning, it takes hold before anyone has had the chance to challenge the model it carries. The risk is not ending up with a bad tool, it is freezing one person's view of the business and noticing once the whole operation depends on it.

 

10. Who will still be able to touch it in two years? The obvious worry is that whoever built the tool leaves, but the more awkward case is the one where they stay. A single person knows how to change a system the operation depends on, and the company loses the freedom to manage them like anyone else. Moving them to another role, asking them to change how they work, or parting company becomes costly for reasons that have nothing to do with their ability. This is not a question of loyalty, it is a question of organisation: a tool nobody else can pick up creates a dependency nobody chose.

 

11. Who owns the code? A recent ruling has made the question very concrete. On 13 February 2026, the Munich court refused to recognise three logos generated by an AI from a user's prompts as protected works, for want of a preponderant human creative influence on the final form. The third party who had reproduced them without permission therefore walked away. The European Parliament took the same line on 10 March 2026, stating in a resolution that content entirely generated by AI should not be protected by copyright.

 

The AI itself holds nothing: only a person can own rights. The question is therefore not whether the code belongs to it, but whether it belongs to anyone at all. If nothing protects it, you are using it without owning anything exclusive, and nothing would stop a competitor doing the same. The courts expressly set aside the time spent, the care taken over the prompt, its length and the use of a paid tier: effort does not substitute for originality. What counts is the trace of human choices, which means keeping the prompts, the successive versions and the edits.

 

The absence of protection does not, however, deprive you of the right to use your tool, since what belongs to nobody may be used by everybody. The risk lies elsewhere: if the model reproduced code belonging to a third party, it is that third party who can act. Orange found this out, ordered in February 2024 by the Paris Court of Appeal to pay 800,000 euros for infringement after incorporating a library published under the GNU GPL into a platform delivered to the French state, at the end of proceedings begun thirteen years earlier. The exposure grows as soon as the tool leaves the company, whether delivered to a client, a partner or a subsidiary. Ownership also depends on who held the keyboard: in France, rights to software written by an employee in the course of their duties sit with the employer, whereas a freelance contractor's work requires a written assignment. We are not lawyers and none of this is advice; it simply flags questions better settled before the tool becomes indispensable.

 

12. Who has had it audited? A maintenance system is not a neutral tool in the eyes of class. A ship's planned maintenance scheme is approved by its classification society, which requires a computerised system with backups and controlled access, audits it every year, and can withdraw that approval if the records do not hold up. Some vendors also hold a type approval for their product, issued by a class society.

 

A tool built in a few hours does not necessarily meet those requirements, and more to the point, nobody is going to have it checked. Having software audited and certified by an external body takes time, a budget, and someone to carry the file for months. Whoever put the tool together one afternoon has neither the mandate nor the means to start that process, and usually does not think of it at all. Which makes the question a simple one: the day an inspector asks what your maintenance record rests on, what will you have to show him?

 

13. What does it really cost? A subscription carries a fixed cost, known in advance, that does not move with usage. A tool built on model calls costs in proportion to the work done, so it costs more as it succeeds. Then comes your teams' time: before, to frame the need; during, for the back and forth; after, to make what is in production reliable. Then routine upkeep, which on what we see across the deployments we support runs at 15% to 25% of the build cost every year, in money and in staff time.

 

A precedent worth remembering

 

We have watched a barrier of this kind fall before. Between 2004 and 2010, WordPress went from a blogging tool to a system that let anyone build a complete website without writing a line of code, and the promise was kept.

 

Twenty years on, the record is instructive. In 2025, 11,334 vulnerabilities were logged across that ecosystem, of which 91% came from the plugins added to meet a particular need, and only six from the software itself. Close to half were published with no fix available, their author having moved on. One detail is worth pausing on: an abandoned plugin keeps displaying as up to date in the site's admin panel. Its owner finds out on the day of the incident.

 

None of this is an argument against building your own. It is a reminder that the maintenance bill always arrives later than the satisfaction of the first day.

 

Build, buy, or combine

 

The question is not build or buy, but: how long does this tool have to live, and will it one day have to stand as evidence in front of a third party? A maintenance record is not an organisational convenience. It is a document you will produce in front of a port state control inspector, in front of class, and in front of your insurer after a claim.

 

Build it yourself when

 

  • the use is one-off or seasonal;

  • a single person uses it;

  • the data is not sensitive;

  • dropping it tomorrow would cost you nothing.

 

Buy when

 

  • several people work on the same data;

  • that data will serve as evidence to a third party;

  • you will need support and continuity;

  • the tool must outlive two years, and the team will change.

 

Combine when

 

  • the core is bought and the automation built around it;

  • the connection runs through an API into the existing system;

  • the operating rules stay inside the software;

  • the AI acts on top of it, rather than instead of it.

 

The first list is wider than people assume, and we genuinely encourage using it: plenty of tasks do not deserve software at all. It is the third route, however, that is growing fastest, and it is the one we build. The BoatOn Book holds the operating rules, the audit trail and the evidence, while AI sits on top of it to populate, analyse and alert, without standing in for the register.

 

What to take away

 

A tool built in an afternoon is not a bad tool, and whoever built it showed initiative that many organisations lack. It is simply a tool whose cost nobody has paid yet, and whose strength nobody has tested.

 

Two questions serve better than asking whether it works, since the answer is usually yes. Who looks after it in eighteen months, once its author has gone? And what is it worth the day somebody challenges it?

 

If both answers suit you, build it.

 

Further reading

 

 

Sources: « State of WordPress Security in 2026 » whitepaper, Patchstack, February 2026 (a vendor of security solutions); WordPress release history; IACS Unified Requirement Z20 on machinery planned maintenance schemes; « Quel droit d'auteur à l'ère de l'intelligence artificielle générative ? », Mission Appui au patrimoine immatériel de l'État, legal affairs directorate of the French Ministry of the Economy, 4 June 2026; Paris Court of Appeal, 14 February 2024, Entr'Ouvert v Orange, following the Cour de cassation, 5 October 2022. The 15% to 25% annual upkeep range is a BoatOn estimate, drawn from the deployments we support.

 
 
bottom of page