Justin Fanelli, the US Navy's chief technology officer, has spent the last few years making the Navy faster at getting new software into use. His hardest remaining problem is the opposite of speed: what to do with all the software that is already running.
Fast building moves the cost to the end
The Navy's own results show how much faster buying software has become. Its buying cycle for new technology fell from 12 to 18 months to a few weeks. That is the good news, and it has a side effect. When it takes weeks instead of a year and a half to bring something in, far more things get brought in.
Fanelli describes what follows. When 100 apps prove popular, users add more around them, and the count can swell into the thousands. On ships, which can't all be reached for updates at once, there are at times about 30 versions of the same software in use. Each tool was cheap to add. But each one has to keep working whenever something it connects to changes, and that cost does not show up on the day the tool is built. It shows up later, as a pile.
He is not trying to rebuild everything. His answer is to take out the software that can be retired, with AI agents helping to find it and carve it out, while the Navy surveys its users to learn which tools have become slow or no longer work well.
Spreadsheets already taught this lesson
This has happened before with a tool that anyone could build. Spreadsheets let people create their own calculators without asking the IT department, and many of them were never retired. After JPMorgan lost billions of dollars in 2012, its own review found that the model meant to measure the risk in those trades ran through a chain of Excel spreadsheets that people filled in by hand, copying numbers from one file to the next. The review said it should have been automated, and it never was. The spreadsheet had been easy to start, it grew with the business, and nobody was responsible for replacing it.
AI-built tools are the same thing with more reach. A spreadsheet only calculates. A tool made with an AI assistant can read your inbox, send emails and update records, so one that is forgotten can do more harm than one that is merely out of date.
What it costs a small business
Take a firm of twenty people where someone builds one small tool a week. After a year there are about 50. If each needs fifteen minutes a month of checking, fixing or answering "is this still used?", that is about 12 hours a month, more than a day and a half of someone's time, before anything actually breaks. In practice something will break, usually on a day when the person who built the tool is away.
The same AI that built the tools can help clean them up. A manager could ask an assistant to list every tool the business uses, what each one connects to, when it last ran and who used it. The list will show the three tools doing the same job and the one nobody has opened since spring. The manager still decides what goes, because only a person knows that the quarterly report depends on the odd-looking one.
So when you build a tool with AI, write down three things at the start: who owns it, what it reads, and the date you will look at it again. If nobody has used it by then, switch it off. Making software has become cheap, so the people who keep up will be the ones who are good at deciding what to stop running.