An operations manager spends a weekend with an AI app builder and puts together the request-tracking system his team had been asking for over a year. It works, twenty people use it, everyone is happy. Three months later someone asks where the customers' ID numbers and addresses are stored, who has access to that database, and what happens if the manager leaves the company. Nobody can answer.
Software development with artificial intelligence stopped being an experiment a while ago. Almost every serious team uses it, and for some cases you really can build software with AI in days instead of months. What gets discussed less is where the tool's contribution ends and where you need someone experienced looking over it. That is what this piece is about, and at the end there's a list of questions to ask anyone proposing a project built this way.
What the latest data shows
Google's DORA 2025 report, published in September 2025, found that 90% of the technology professionals surveyed use AI at work. Stack Overflow's 2026 Developer Survey, released this month with more than 30,000 responses, confirms that coding assistants and agents are now part of everyday work.
Trust is a different story. In that same survey, only 6.6% say they trust AI with important work decisions. About 48% trust it when they can easily check what it produces. More than 70% find it useful for writing and debugging code, but that drops to 49% for design and 35% for communication.
Productivity is where things get odd. A controlled study by METR in July 2025 found that experienced developers working on their own projects took 19% longer when using AI. They believed they had been 20% faster. In a February 2026 update, METR says the speedup today is probably larger, while admitting its new data isn't strong enough to say by how much.
DORA has a line we think sums it up best: AI amplifies what is already there. An organized team gets more out of it; a messy one watches its problems grow. In the report, higher AI adoption was linked to faster delivery and, at the same time, less stable delivery.
Where AI really speeds things up
In an agile software development project, the difference shows up mostly in these tasks:
- Repetitive code and base structure, like admin screens, forms, connections to well-known APIs or simple migrations.
- Unit tests generated from existing code, including edge cases nobody had bothered to write.
- Documentation for functions and endpoints, plus that README nobody keeps up to date.
- A first pass on every pull request, so the human reviewer doesn't waste time on obvious mistakes.
- Clickable prototypes in a few days, handy for validating an idea with users before committing budget.
- Explaining old, undocumented modules, which saves hours on legacy systems.
What these have in common is that you can check the result quickly. If the test fails or the screen doesn't load, you see it right away. We wrote more about how this changes the way teams work in AI agile software factory: how AI speeds up development.
What is still human work
When you can't check the result at a glance, AI can hand you something that looks right and is wrong. That's where a person has to make the call.
| Area | What AI contributes | What a person decides |
|---|---|---|
| Architecture | Suggests options and known patterns | Which one fits your volume, your budget and whoever will maintain it |
| Business rules | Implements what it's told | What to do with the exception nobody wrote down |
| Security | Catches some common mistakes | Who can access what, and what ends up exposed |
| Integrations | Writes the connector | What happens when the other system fails or changes |
| Personal data | Nothing on its own | Compliance with Chile's Law 21.719 and agreements with vendors |
| Production | Helps read logs | Running, monitoring and responding to incidents |
Developers draw that line themselves. According to Stack Overflow's summary, only 20% use AI to deploy, operate or troubleshoot production systems, which is exactly where mistakes get expensive.
A year earlier, in the 2025 survey, the most common complaint (66%) was AI answers that are "almost right, but not quite," and 45% said debugging AI-generated code takes them longer. That helps explain METR's result: the code comes out fast, but reviewing and fixing it eats into the savings.
The risk of apps built only with AI
The manager in the opening story didn't go wrong by using AI. He went wrong because nobody with experience reviewed what he built.
On security, the data is fairly concrete. Veracode's July 2025 report tested more than 100 language models and found that 45% of generated code samples introduced an OWASP Top 10 vulnerability. In Java, the failure rate reached 72%. Its March 2026 update showed the overall number barely moved. Models keep getting better at writing code that works, but that doesn't make it safer.
This is what usually turns up when someone reviews an app built only with AI:
- Credentials or API keys written into the code.
- Databases with no per-user access rules.
- Real customer or employee data in test environments, an issue we already looked at under Law 21.719.
- Outdated libraries, or ones that simply don't exist, suggested by the model.
- No tests, no documentation and nobody who knows how to maintain it.
None of this shows up in a demo. It shows up once the app has users and data.
What to ask before approving an AI-built project
If you're about to hire or approve a development project that leans on AI, these questions do a good job of separating a team that uses it with discipline from one that just moves fast:
- Which AI tools do you use, and what part of our code or data gets sent to external services?
- Who reviews generated code before it reaches production?
- Are automated tests and security scans part of the delivery pipeline?
- Who made the architecture decisions, and where are they documented?
- Who maintains the system six months from now, and whose name is on the repository?
If the answers are vague, the project probably will be too.
Frequently asked questions
Can AI replace a development team?
Not today. It speeds up tasks that are easy to verify, like repetitive code, tests and documentation. Architecture, security and business rules still need experienced people.
Is it safe to build software with AI?
It can be, as long as generated code goes through human review, testing and security analysis. Without that, studies show a significant share of it carries known vulnerabilities.
Is software development with artificial intelligence cheaper?
It usually lowers the cost of prototypes and iterations. Total savings depend on how much time you gain writing code versus the extra time it takes to review it.
Custom development, with AI and someone in charge
At Disrupsoft we build custom web and mobile applications for companies in Chile, connected to what you already use, whether that's Azure, Power Platform or an AS/400. We use AI where it actually saves time, and the team makes the architecture and security decisions. If you have an idea, or an app someone already put together with AI that you'd like reviewed before it grows, let's talk.
