Clients increasingly ask us, sometimes nervously, whether we use AI to write their code. The answer is yes, in parts of our workflow, with rules. We'd rather be open about how, because there's a lot of hype and a lot of fear around this topic, and the reality is somewhere in between.
Where AI tools genuinely help
Boilerplate and repetitive code. Database migrations, form validation rules, CRUD screens for an admin panel, API client code. Work that follows a known pattern and is tedious to type. AI tools produce a solid first draft in seconds.
Exploring unfamiliar code. When we take over an old project, asking an AI assistant to explain what a confusing function does, or to trace where a value comes from, is a big time saver. We still verify, but it gets us oriented quickly.
Tests. Writing unit tests for existing code is something developers tend to put off. AI tools are good at generating a first set of test cases, including edge cases we might have missed.
Converting and migrating. Updating syntax for a new PHP version, converting old CodeIgniter 3 code to CodeIgniter 4 patterns, rewriting jQuery into plain JavaScript. Mechanical work with clear rules.
Debugging help. Pasting an error message and the relevant code often surfaces the likely cause faster than searching forums.
Where they cost time
Subtle business logic. Pricing rules, tax calculations, permission systems. AI-generated code can look right, pass a quick glance, and be wrong in ways that only show up with real data. In these areas we write the code ourselves and use AI only for review.
Confident nonsense. Tools sometimes invent functions that don't exist in a library, or use outdated APIs. Experienced developers catch it quickly. Less experienced ones can lose an afternoon.
Large, sweeping changes. Letting an agent make changes across dozens of files is tempting. Reviewing those changes properly takes almost as long as making them, and skipping the review is where bugs creep in.
Our rules
- A human owns every line. Whoever commits the code is responsible for it, regardless of who or what typed it.
- Everything gets reviewed. AI-assisted code goes through the same review as hand-written code.
- No client secrets in prompts. No passwords, API keys, or customer personal data pasted into AI tools.
- Security-sensitive code gets extra scrutiny. Authentication, payments, file uploads and anything handling user input.
- We check licences. We're careful about code that looks copied from somewhere else.
Does it make projects cheaper?
Somewhat. On typical projects, we estimate the time saved at maybe 15 to 30 percent of the pure coding work. But coding is only part of a project. Planning, design, content, client feedback, testing and deployment take just as long as before. So overall timelines are a bit shorter, not dramatically shorter.
What has improved noticeably is quality in the boring areas. More tests, better documentation, more consistent code, because the effort to produce them dropped.
The bigger picture
AI tools are a genuine step forward for developers, the way good IDEs and frameworks were before them. They don't replace understanding what the client needs, making sensible trade-offs, or taking responsibility for the result. That's still the job.
