AI in development has stopped being marketing: neural networks genuinely write code, tests and copy. But between “we have ChatGPT” and “this makes your project cheaper” there is a gulf. Some tasks AI does several times faster than a human; behind others there still needs to be an engineer who answers for the result. Let's look at where a neural network genuinely cuts the estimate, where the saving turns into a loss, and what a client should ask a contractor about AI.
Where AI genuinely cuts the budget
The gains appear where the work is routine and easy to verify:
- Routine code. CRUD endpoints, forms with validation, responsive markup against a finished design system, integrations following documentation — a neural network writes these faster than a human, and a reviewer checks the result
- Tests. A routine but critical part of a project: AI is good at generating tests for already-written code. Tests stop being the line item “cut to save budget”
- Migrations and routine refactoring. Upgrading a library, replacing deprecated APIs, bringing code to a single style — hours of mechanical work turn into minutes
- Copy drafts. First versions of service descriptions, FAQ, meta tags — editing a neural network's draft is faster than writing from scratch
- Repetitive edits. “Fix it in all sections”, “rename the field across the whole project” — the kind of task that used to be billed by the hour
In our practice, AI agents save a noticeable share of hours on such tasks — the amount depends on the project type. Not “ten times”, but enough to be visible in the estimate.
Where a human is still needed
- Architecture. Stack choice, project structure, data schema — these decisions define the cost of ownership for years. A neural network will offer an average, workable option; whether it is optimal for your task is an engineer's call
- Security. Payments, personal data, access rights. Generated code is by default “plausible”, not “secure”: holes and leaks cost more than any saving
- Acceptance and quality. Someone has to verify that what was built works on real devices and in real scenarios, not just “runs”. A neural network cannot share responsibility to the client with anyone
- Product decisions. What to remove from a page, how to simplify a form, why this integration is needed — business questions; the neural network doesn't have your context
Why an “AI website in one evening” costs more
The promise “describe the website in words — get it built” works great until the first real check. Then begins what we regularly see in rework projects:
- A generated project without tests or review contains bugs that surface after launch — and fixing things after launch is always more expensive than during development
- Code without a considered architecture doesn't scale: adding payments or a second language to an “evening” project often means rewriting it entirely
- Nobody knows how the project works: no documentation, no author, prompts not saved — every next contractor starts with “archaeology” and puts it in the estimate
- Nobody checked security: unprotected forms, keys in the code, outdated dependencies
Count not the price of the first version but the cost of ownership: development plus rework plus maintenance. An “evening” website almost always loses on that metric too.
How we use AI
At PROWEB we treat AI agents as a tool, not a team replacement: routine code, tests, migrations and drafts are generated with AI; architecture, security and acceptance stay with people. A separate principle is control: generated code goes through the same review as hand-written code, and agents work in a constrained environment without access to extra systems. We have a separate article on how to give AI agents access to projects safely and why it is a discipline of its own.
What to ask a contractor about AI
In 2026 AI tools are the norm — hiding them makes no sense. The question is not “do you use AI” but “how”:
- Which stages of yours run with AI, and which are strictly manual?
- Who reviews generated code, and against which criteria?
- How does AI affect price and deadlines — and is it fixed in the estimate?
- What about confidentiality: which project data goes to external services?
- Who is responsible for bugs in AI-generated code, and is there a post-delivery warranty?
A good contractor answers all of this without a pause. A vague “we use neural networks everywhere, that's why it's cheap” is a reason to be wary: cheap usually lasts exactly until the first rework.
In short
AI genuinely reduces development costs on routine tasks: template code, tests, migrations, copy drafts, repetitive edits. Architecture, security, acceptance and product decisions stay with people — not as a relic, but as the reason the project works at all. An “AI website in one evening” is cheaper only for the first evening: without tests, architecture and review, owning it costs more. Ask a contractor how exactly they use AI, who checks the result and how it is reflected in the estimate — an honest answer takes a minute and saves months.