Two things are true at once, and most commentary picks one and ignores the other.
The first: AI has genuinely changed how software gets written. Google's 2025 DORA report found that around 90% of software professionals now use AI at work, spending a median of about two hours a day with it, and more than 80% report increased productivity. Agentic tools read whole codebases, resolve dependencies and generate working features from a description. This is not a demo any more.
The second: the gap between "working code on a screen" and "a product in the App Store that real people trust with their data" is large, and AI closes very little of it. Nearly everything expensive about shipping an app lives in that gap.
What AI genuinely does well now
Being specific matters more than being enthusiastic.
- First drafts of well-understood code. Standard screens, forms, list views, API clients, data models. Things done a million times before, which is exactly what a model trained on public code is good at.
- Boilerplate and glue. Wiring, configuration, type definitions, conversions between shapes — real hours, low judgement.
- Tests, given clear intent. Often better than the tests a rushed human writes, because it does not get bored.
- Explaining unfamiliar code. Genuinely excellent, and underrated. It compresses the ramp-up on an inherited codebase dramatically.
- Prototypes for a conversation. A clickable thing in an afternoon to decide whether an idea is worth pursuing.
That is a real productivity gain. It is not a small one. It also is not the same as building your product.
Where it still falls down
Deciding what to build
AI will implement whatever you describe, including the wrong thing, quickly and confidently. It does not push back that your version one has three features too many, or that the flow you sketched will confuse the exact user you are targeting. The most valuable thing a good engineer does on a project is sometimes refusing to build something — and that judgement is precisely what does not come in the box.
Architecture that survives contact with growth
Generated code is locally sensible and globally inconsistent. Ask for twelve features and you can get twelve reasonable implementations that disagree about state management, error handling and data flow. It runs. Then at month four, a change in one place breaks something unrelated, and nobody can explain why, because no human ever held the whole design in their head.
The last 20%, which is most of the cost
Edge cases, offline behaviour, error states, race conditions, memory on older Android devices, accessibility, App Store review, the payment flow that fails only on one bank's 3-D Secure page. This is where app projects actually spend their time, and it is the part AI helps with least — because it is specific to your product and mostly absent from public code.
Anything with real consequences
Handling money, health data or personal data under a regime like India's DPDP Act means someone has to be accountable for correctness. "The model wrote it" is not a position you can hold with a regulator, an auditor or a user whose data leaked.
Debugging what you did not write
This is the failure mode we get called about most. A founder ships an AI-assisted build, it works, then it breaks in a way the model cannot fix — because fixing it needs a mental model of the system nobody has. The rescue costs more than doing it properly would have, and it arrives at the worst possible moment.
How we actually use it
Not using these tools would be malpractice at this point — and pretending they replace engineering would be a different kind. Concretely, on our own builds: AI drafts routine code and tests that a senior engineer reviews line by line; it explains unfamiliar third-party code; it accelerates prototypes when we are pressure-testing an idea. Architecture, data modelling, security decisions and anything touching money or personal data stay with people who can be accountable for them.
The net effect is real but unglamorous: we get to the same quality faster, and spend more of the saved time on the parts that actually decide whether a product works.
If you are considering building it yourself with AI
Sometimes that is genuinely the right call, and we will say so on a call. It tends to work when:
- You are validating an idea rather than launching a business on it.
- The app is largely standard — CRUD, forms, content — with few hard integrations.
- It handles no money, health data or sensitive personal data.
- You or a co-founder can read the code well enough to judge whether it is sane.
- You would be comfortable rewriting it if it takes off.
That last point is the real test. AI-assisted prototypes are excellent at answering "is this worth building?" and poor at becoming the thing you scale. Treating the prototype as disposable is the difference between a cheap experiment and an expensive dead end.