What an experienced developer does when AI writes the code
When AI writes the code, three jobs stay with a person: knowing what to ask for, spotting a wrong answer, and deciding what not to build.

Ask an AI coding assistant for a checkout page and you get one in minutes. It loads, it takes a card, it says thank you. So business owners ask us a fair question: when AI writes the code, what is an experienced developer still for?
Our answer, after 15+ years of building web and mobile products: for the three jobs AI does not do for you. Those jobs are worth more now, not less.
When AI writes the code, three jobs stay with a person
The short version: knowing what to ask for, recognising a wrong answer, and deciding what not to build.
- Knowing what to ask for. “Let customers pay online” is a wish. A brief says what happens when the card is declined, or when the customer presses Pay twice. AI builds what it is asked for. It rarely asks about what you forgot to mention.
- Recognising a wrong answer. AI answers in the same confident tone whether it is right or wrong. Telling the difference takes someone who has seen that kind of mistake before.
- Deciding what not to build. Building is cheap now, but every feature has to be secured, updated and supported for years. The cheapest feature is the one you were right to skip.
“It works” is not “it is right”
A demo proves that the easy path works: one customer, a good connection, nobody trying to cheat. A business runs on everything else. Five places where the gap shows most:
| Area | “It works” | What an experienced developer asks |
|---|---|---|
| Security | The login page accepts the right password | What does it do with a wrong one, a thousand times a minute? |
| Payments | A test card is charged | What if the payment confirmation arrives twice, or never? |
| Data | The order is saved | Is there a backup, and has anyone tried to restore it? |
| Performance | The page is fast with 20 products | And with 20,000 products, on a phone, on the day of the campaign? |
| App store rules | The app runs on your phone | Will the store’s review accept it? |
App store rules are a good example of knowledge that is not in the code at all. Apple, for instance, requires that an app that lets people create an account also lets them delete it inside the app. An app can work perfectly and still be rejected. These rules change, so read the current text first.
In 2025 Veracode tested more than 100 AI models on 80 coding tasks and found security vulnerabilities in 45 percent of the cases. Newer models got better at writing code that runs, and no better at writing code that is safe.
Why experienced developers trust AI the least
In Stack Overflow’s 2025 developer survey, 84% of developers use AI tools or plan to, yet more of them distrust the accuracy of the answers (46%) than trust it (33%). Their most common frustration, named by 66%: answers that are “almost right, but not quite”. The survey’s own summary is that experienced developers are the most cautious.
Then there is the feeling of speed. In 2025 the research group METR timed 16 experienced developers on 246 real tasks. They expected AI to cut their time by 24%. The tasks took 19% longer, and afterwards the developers still believed they had been faster. METR itself now says that result is out of date and that newer tools probably do speed developers up. What lasts is a different lesson: how fast the work feels is not a measurement. Somebody has to look at the result.
What 15+ years change about how we use AI
At yalla ya! AI is how we build now. Our experience moved: less of it goes into typing code, more into the brief before and the review after. Our own apps have passed 2 million downloads, and years of store reviews, payments and updates that must not break what people already use taught us what a demo never shows. That is what we bring to mobile app development and to the way we watch security.
Judgement should not be applied once at kickoff and then forgotten. Our engineers work inside Butler, the workspace we set up for each client, where the business numbers, the decisions and the code sit together (we wrote separately about what general AI tools for business leave unconnected). We run about 20 projects this way, so an experienced person follows each of them and can see when a change that “worked” moved a number the wrong way.
We get things wrong too. Experience does not remove mistakes. It shortens the time until someone notices, and it means there is a person you can call.
Where the person came in
In our projects AI writes and tests the software. A person decides what gets built and in what order, reads what came back before it reaches your customers, and says no to the feature that is not worth its upkeep.
What to ask before you accept AI-built work
If someone hands you software written with AI, whether that is an agency, an employee, or you with a chat assistant such as ChatGPT, don’t ask whether it works. Ask who checked it, and what they checked: security, payments, data, performance, store rules. If the answer is “the AI did”, nobody has checked it yet.
If you would like a second pair of eyes, book the free 30-minute consultation and project review. We look at what you have before the call.
Keep reading
Expert talentServer security updates: one week across dozens of servers
Four published flaws, dozens of servers, one week: what automatic updates fixed on their own, and what needed a reboot and a person.
Field notesHow we automate bookkeeping so no invoice gets skipped
A field note from our own books: three AI routines chase tax invoices and check month-end gaps, a tile flags unpaid invoices. Real counts.
Human in the loopAutomate customer support with AI: what waits for a person
Two AI agents read app reviews and guest messages every day. What they answered by themselves in November, what waited for a person and what is still open.