Your AI-Built Website Is 80% Done. Here’s What the Last 20% Actually Involves

AI tools have changed how quickly people can get a website off the ground.

A few years ago, building a functioning site meant choosing a platform, designing layouts, writing code, setting up hosting and then spending days or weeks getting the basic structure into place.

Now, someone can sit down with an AI website builder or coding tool and have something surprisingly convincing running in an afternoon.

That is useful.

It is also where a lot of projects become slightly misleading.

The website may technically exist. The pages load. The navigation works. There is a form. Maybe there is even authentication, a dashboard or some custom functionality.

But there is usually a point where the project stops getting dramatically better with every prompt.

That is the part I think of as the last 20%.

And in practice, that final 20% can make the difference between something that looks like an experiment and something you would actually feel comfortable putting in front of customers.

The first 80% is easier to see

AI is very good at visible progress.

You can ask it to:

• create a homepage

• add a services section

• build a contact form

• create a pricing table

• generate a dashboard

• add animations

• build a mobile menu

• connect a database

• write some copy

• change the colour scheme

Each request produces something obvious.

That makes the project feel like it is moving quickly.

The harder work starts when the individual pieces need to feel like one website rather than a collection of generated sections.

That work tends to be less dramatic, but more important.

Consistency is usually where things start to fall apart

One of the most common things I notice in AI-built websites is inconsistency.

A single section might look good.

Five individually good sections do not necessarily make a good page.

Spacing starts to vary. Headings change size without a clear reason. One button is rounded, another is not. Some cards have shadows, some have borders, some glow. The page starts to feel like each section came from a slightly different design system.

That is not because AI cannot produce attractive interfaces.

It is because good design relies heavily on decisions being repeated consistently.

Typography, spacing, layout rhythm, interaction, image treatment, buttons, cards and content hierarchy all need rules.

A lot of what makes a website feel finished has surprisingly little to do with adding more features. I wrote more about that in What Actually Makes a Website Feel Professional?, where I look at the smaller design decisions that make a site feel considered rather than simply polished.

The last 20% often involves removing variation rather than adding more design.

AI tends to add before it removes

Ask an AI tool to improve a page and it will often add something.

Another feature strip.

Another comparison section.

Another set of cards.

More icons.

More copy.

Another gradient.

Another animation.

That can work for a while, but eventually the page becomes heavier without becoming clearer.

One of the most useful things a designer can do at that stage is ask a different question:

What can be removed without making the website worse?

This is something I have increasingly noticed across my own projects too. Better Websites Don’t Always Need More goes into that idea in more detail: sometimes the biggest improvement comes from removing friction, repetition and unnecessary design rather than adding another section.

A lot of professional-looking websites are not impressive because they contain more. They feel better because fewer things are competing for attention.

That judgement is difficult to automate because there is no universal rule for how much is enough.

The copy often needs a second pass

AI-generated website copy has improved enormously, but there are still patterns that appear repeatedly.

You start seeing phrases like:

“tailored solutions”

“seamless experiences”

“elevate your business”

“cutting-edge technology”

“designed to help you thrive”

None of those phrases are necessarily wrong.

The problem is that they could belong to almost any company.

The final version of a website needs to sound like the actual business.

That usually means adding specifics, removing filler and writing around the things customers genuinely ask.

A useful homepage should make it easier to understand:

• what the company does

• who it is for

• why someone should trust it

• what makes it different

• what the next step is

AI can help draft that content, but someone still has to decide whether the words actually say anything.

Mobile usually exposes the shortcuts

AI-built interfaces often look most convincing on a large desktop screen.

That is also the easiest environment to design for.

Mobile reveals problems quickly.

Headlines wrap badly.

Two-column layouts become extremely long.

Floating elements overlap.

Buttons become awkward to tap.

Animations that looked impressive on desktop become distracting.

Navigation gets compressed rather than genuinely redesigned.

A responsive website is not just a desktop website made narrower.

Sometimes the mobile version needs different spacing, different hierarchy, different interactions or fewer decorative elements entirely.

That detailed responsive work is often part of the final 20%.

A working form is not necessarily a finished form

Forms are a good example of the difference between functionality and readiness.

An AI tool can create a contact form very quickly.

But then you need to ask:

Where does the submission go?

What happens if sending fails?

Is spam handled?

Does the user see confirmation?

Are validation messages understandable?

Does it work properly on mobile?

Are personal details stored anywhere?

Are there privacy implications?

Is analytics tracking the conversion?

That same pattern appears across almost every functional feature.

“Works” is only the first question.

SEO needs structure, not just metadata

AI tools are very good at generating meta titles and descriptions.

That does not mean the site has a good SEO structure.

Search visibility depends on broader questions:

• Are there useful pages for the services people actually search for?

• Does each page have a clear purpose?

• Are pages competing with each other for the same topic?

• Is important content reachable through internal links?

• Is the heading structure sensible?

• Are URLs stable?

• Is Google able to crawl and index the site properly?

• Are there dozens of thin pages that probably should not exist?

• Is the content genuinely useful?

Adding keywords to the homepage at the end of the build is not the same as designing a site around search intent.

The last 20% often means stepping back from individual pages and looking at the website as a whole.

AI-generated code can work and still be difficult to maintain

This becomes even more important with vibe-coded websites and web apps.

A generated feature can work perfectly during a demo while still creating problems underneath.

You may have:

• duplicated logic

• hard-coded values

• weak error handling

• unnecessary dependencies

• exposed environment variables

• inconsistent database queries

• authentication edge cases

• poor loading states

• code that becomes harder to change with every new prompt

None of that automatically means the code is bad.

Human developers create messy code too.

The difference is that AI makes it possible to create a lot of code very quickly, which means technical debt can accumulate just as quickly.

Once a project handles customer accounts, business data, payments or important workflows, it becomes worth reviewing what is actually happening behind the interface.

The admin side often gets forgotten

A website can look polished to visitors while being frustrating for the person who has to manage it.

This is especially common with AI-built projects because so much attention goes into generating the frontend.

Then the business owner realises there is no practical way to:

• edit content

• manage users

• update products

• review form submissions

• manage bookings

• upload images

• publish articles

• control SEO fields

• see what the system is doing

At that point, the missing piece is not another public-facing section.

It is the boring but important operational interface behind it.

Good admin tooling rarely gets shown in website inspiration galleries, but it can be one of the most valuable parts of the build.

Accessibility needs more than a score

Another area where AI can give a false sense of completion is accessibility.

A tool might generate semantic-looking markup or produce a respectable automated score.

That does not guarantee the interface is actually easy to use.

Accessibility also involves things like:

• keyboard navigation

• focus states

• form labels

• contrast

• readable text

• clear errors

• reduced-motion behaviour

• meaningful link text

• logical heading structure

Automated checks are useful.

They are not the same as reviewing how the interface behaves for real people.

Performance becomes more important as the site grows

A simple generated page may load quickly.

Then the site accumulates:

• animation libraries

• large images

• video

• third-party scripts

• analytics

• chat widgets

• custom fonts

• API calls

• dashboards

• component libraries

Performance usually gets worse gradually rather than failing all at once.

The final stages of a project are often where somebody needs to look at what is actually being loaded and whether all of it still deserves to be there.

Error states are part of the product

When a prototype works, you mostly see the successful path.

Production websites have unsuccessful paths too.

What happens when:

• the API is unavailable?

• payment fails?

• the user enters the wrong password?

• an image cannot load?

• the database query times out?

• a form is submitted twice?

• there are no results?

• a user does not have permission?

• a booking slot disappears while someone is checking out?

These situations are not glamorous, which is probably why they are easy to postpone.

But they are exactly the kind of details that make software feel reliable.

Analytics should answer useful questions

Another common final-stage job is making sure you can actually tell whether the website works.

Installing analytics is easy.

Deciding what matters is harder.

For a business website, useful questions might include:

• Are people reaching the enquiry page?

• Which services generate enquiries?

• Where do users leave?

• Are people reading case studies?

• Does organic traffic convert differently from paid traffic?

• Are mobile visitors struggling more than desktop users?

• Which landing pages actually contribute to leads?

Without that setup, you launch the site and then rely mostly on instinct.

Security becomes more important when the website stops being a website

For a straightforward marketing website, the security surface can be relatively small.

Once AI-assisted development introduces logins, databases, file uploads, payments, dashboards or APIs, the stakes change.

At that point it becomes important to review things such as:

• access permissions

• secrets and API keys

• authentication

• database rules

• server-side validation

• uploads

• rate limiting

• payment flows

• logging

AI can help implement these things.

The important part is not assuming that because the feature works, every security decision behind it is automatically correct.

Launching is its own piece of work

A website can be finished in the development environment and still not be ready to launch.

There may still be:

• DNS changes

• domain configuration

• redirects

• SSL

• email setup

• analytics

• Search Console

• cookie/privacy configuration

• broken links

• missing metadata

• test content

• staging URLs

• forms that still point somewhere temporary

This is the uninteresting end of the project that nobody puts in a design mockup.

It is also the part customers actually depend on.

The last 20% is mostly judgement

I do not think the lesson from any of this is that AI website builders do not work.

They clearly do.

And they are going to keep getting better.

The interesting change is that they have moved the bottleneck.

Getting something onto the screen used to be a large part of the work.

Now that can happen extremely quickly.

The harder part becomes deciding whether the result is actually good enough.

Does this section need another prompt, or should it be removed?

Is this code genuinely broken, or just untidy?

Should this website stay simple, or does the business actually need custom functionality?

Does Wix still suit the project?

Is WordPress better?

Should this remain a marketing site, or has it started becoming a web application?

Those are not really generation problems.

They are judgement problems.

You may not need a complete rebuild

This is probably the most important thing I tell people who have already built most of a site themselves.

Do not assume the professional solution is to throw everything away.

If the foundation is sound, it may make far more sense to:

• keep the working functionality

• refine the design

• rewrite weak sections

• repair mobile layouts

• improve SEO structure

• clean up fragile code

• add the missing admin tools

• review security

• finish the deployment properly

AI may have already done a lot of useful work.

The job is to work out which parts deserve to stay.

When should you bring in a developer or designer?

Probably when prompting has stopped producing meaningful progress.

If every change fixes one thing and breaks another, that is a useful signal.

If the website works but you are uncomfortable showing it to customers, that is another.

If you are about to connect real payments, customer accounts or business data, that is a particularly good time to have somebody review what is underneath.

And if you simply find yourself spending evenings trying to perfect spacing, navigation, SEO or mobile behaviour when you would rather be running the business, that counts too.

AI can get you surprisingly close

That is what makes this new way of building interesting.

Someone without a traditional development background can now create something that would have been completely unrealistic a few years ago.

I do not think the professional response to that should be to dismiss the work.

It should be to help finish it.

If your AI-built website is already 80% of the way there, the most useful question might not be:

“Should I start again?”

It might simply be:

“What does this need before I can confidently call it finished?”

If you have reached that point, take a look at my AI Website Redesign & Cleanup service. I work with existing AI-built and vibe-coded sites to improve the design, mobile experience, SEO, code and functionality without automatically throwing away what you have already built.