01 / Where it genuinely helps
Four things AI does well on a real project
Writing code that already has a shape. Converting a design into markup, wiring a form, refactoring something repetitive. A person still reads every line, but the typing is gone. This is the largest and least glamorous win.
Getting a first visual on the table. Tools like Claude Design put something in front of a client in minutes rather than days. It is never the final design. It moves the conversation from abstract to concrete, which is where the useful disagreements happen.
Drafting content you then fix. A first pass at forty product descriptions is a real saving. Publishing that first pass is how sites end up sounding like every other site.
The boring middle of a migration. Mapping old URLs to new ones, reshaping a spreadsheet of products, finding every page that mentions a discontinued service.
02 / Where it does not
Four things it is still bad at, and they matter
Knowing what your business actually does. It will write a confident paragraph about your differentiators without knowing any. The result reads well and says nothing, which is worse than clumsy writing that is true.
Deciding what to build. Scope is judgement about your customers and your money. A model has no stake in either.
Anything where being wrong is expensive and hard to check. Pricing rules, tax, access control, anything touching payments. Generated code looks right at exactly the moment it is not.
Taste. It produces the average of what exists. The average website is forgettable, and looking like everyone else is not a neutral outcome for a brand.
03 / What changes for you as the client
Faster, but not in the way people expect
The honest position: AI has compressed the build, not the project. Writing the code was never the long pole. Deciding what the pages say, agreeing the structure, getting the content written, waiting for approvals: none of that has moved, and it is where the weeks go.
So be sceptical of anyone quoting dramatically less because they use AI. Either they were padding before, or they are skipping the parts that actually take the time.
04 / The question to ask
Not “do you use AI”: ask what gets reviewed
Everyone uses it now. The useful questions are:
- Who reads the generated code before it ships? “Nobody, we test it” is the wrong answer.
- Is the copy written or generated? If generated, who checked the claims in it are true?
- What did you decide, and what did the tool decide?
05 / Honestly
The part nobody selling AI will tell you
The biggest risk to your project is not that AI writes bad code. It is that it makes it cheap to produce a great deal of plausible material nobody has thought about: pages that exist because they could, copy that says nothing, features that were easy to add rather than needed.
The constraint used to be effort, and effort forced choices. That constraint is gone, and judgement has to do the job instead.
Tell us what you are trying to build and we will tell you where we would use it, where we would not, and what a person will be reading before it goes live.