How we work
We use AI to go faster. We do not use it to decide.
Every developer has AI now, so having it proves nothing. What matters is knowing where it helps, where it quietly costs you, and who is accountable for the difference.
Every web developer has AI now. In Stack Overflow's 2025 survey of developers, 84% said they use it or plan to. So "we use AI" is not a reason to hire anybody. It is table stakes, and it tells you nothing about what you will receive.
The interesting question is what happens next, and the published research on that is not flattering. In the same survey, 46% of developers said they do not trust the accuracy of what AI produces, up from 31% a year earlier, and only 3% said they highly trust it. The most common complaint, from 66% of them, was answers that are almost right but not quite.
Almost right is the expensive kind of wrong. It compiles. It looks reasonable in review. It fails on the case nobody tested, six weeks after launch, on somebody's phone.
What the research says
Three findings worth knowing before you hire anyone.
We would rather show you the uncomfortable numbers than pretend this technology is free. Sources are listed at the bottom of the page.
19% slower
In a controlled trial, experienced developers using AI tools took 19% longer to finish real tasks on codebases they knew well. They predicted it would make them 24% faster, and afterwards still believed they had been 20% faster.
The gap between how fast AI feels and how fast it is, is the single most important thing to know about it.
METR, 2025
An amplifier
Google's DORA programme surveyed nearly 5,000 technology professionals and concluded that AI amplifies whatever a team already is. Teams with strong testing, version control and fast feedback got faster. Teams without them got less stable.
AI does not supply judgement, discipline or a deployment pipeline. It multiplies whatever you brought.
DORA, 2025
Duplication up
Analysis of millions of lines of committed code found duplicated blocks at the highest level on record and refactoring in sharp decline, with copy-pasted code overtaking moved code for the first time.
Code that is never refactored is code somebody pays for later. Usually you, when a change costs three times what it should.
GitClear, 2026
Our line
Where we let it help, and where we do not.
The distinction is simple. AI is good at work whose shape is already decided. It is bad at deciding the shape.
Where we use it
First drafts of the mechanical parts
Boilerplate, form scaffolding, type definitions, repetitive refactors across many files, migration scripts. Work where the shape is already decided and typing it out is the only cost.
Exploring options quickly
Three approaches to a layout or a data model in the time it used to take to write one. We still choose, and we choose on merit rather than on which one appeared first.
Test scaffolding
Generating the boring half of a test suite, including the edge cases that are easy to forget. Then reading every one, because a test that asserts the wrong thing is worse than no test.
Research and reading
Summarising documentation, comparing library trade-offs, checking what changed in a framework release. Verified against the primary source before it influences anything.
Content structure, not content
Outlines, headings and consistency passes. We do not let it invent facts about your business, and we do not publish text nobody has read.
Where we never do
Architecture and data models
The decisions that are cheap now and expensive in year two. These come from having watched earlier choices age, which is exactly the thing a model cannot have done.
Performance budgets
AI will happily add a library that solves your problem and costs 90KB. Someone has to be accountable for the number at the bottom, and it has to be a person.
Anything we cannot explain
If nobody on our side can explain a line to you, it does not ship. That single rule removes most of what goes wrong with generated code.
Facts about your business
No invented review counts, client numbers, certifications or founding dates. This is a rule about honesty, and models are confidently wrong in exactly this area.
The final read
Every line is reviewed by the person who will still be maintaining it in two years. That is the whole difference between using AI and depending on it.
The guardrails
What catches it when it is wrong.
DORA’s finding was that AI amplifies whatever a team already has. These are the things we already had, which is why going faster has not cost us stability.
- TypeScript across the whole project, so a rename breaks the build rather than a page
- Linting and formatting enforced automatically, not by memory
- A performance budget in the build, so added weight fails the check
- Verification against the built output rather than the source
- Lighthouse measured as a median of several runs, against production
- Every page's structured data parsed and asserted, not assumed
- Preview deployments so you approve the change, not a description of it
- Code review by a human before anything reaches your domain
Questions
What clients actually ask us about this.
Is my website written by AI?
Parts of the first draft, yes, in the same way parts of it are written by a code formatter. The architecture, the data model, the performance decisions and every line that ships are chosen and reviewed by a person. If nobody here can explain a line to you, it does not go in.
Does using AI make it cheaper?
It makes some stages faster, and where it saves us time that shows up in the estimate. It does not turn a twelve week application into a two week one, and any studio telling you it does is quoting on the demo rather than on the delivery.
Will the code be maintainable by someone else?
That is the specific risk with heavy AI use, and it is measurable: analysis of committed code across the industry shows duplication at record highs and refactoring in decline. We review for it, we refactor rather than paste, and we hand over code any competent developer can pick up.
Do you use AI to write the content on my site?
For structure and consistency, yes. For facts about your business, never. No invented review counts, client numbers, certifications or dates. Everything factual comes from you or it does not appear.
How do I know AI has not introduced a security problem?
Reviewed code, typed interfaces, dependency scanning, secrets kept server side, and a human read of anything touching authentication or payments. The guardrails on this page exist because generated code is confidently wrong in exactly these places.
Sources
Check them yourself.
Every figure on this page comes from published research, not from our own marketing. The links go to the originals.
Want the speed without the shortcuts?
That is the whole offer: modern tools used by people who have been doing this since 2014, and who will tell you when the tool is wrong.