Website Development by AI: What Agencies Can Ship and What Needs Review

Coding Liquids tutorial cover featuring Sagnik Bhattacharya for Website Development by AI: What Agencies Can Ship and What Needs Review.
Coding Liquids tutorial cover featuring Sagnik Bhattacharya for Website Development by AI: What Agencies Can Ship and What Needs Review.

AI can deliver a working first draft of a small client site: layout, copy, basic forms and image descriptions. It cannot deliver a finished one. A developer or designer still has to review forms, accessibility, accuracy of every claim, tracking, security and launch settings, because those are where mistakes cost the agency money or trust.

The useful way to think about it is as a division of labour. AI is quick at the first 70 per cent of a page and unreliable at the details that make it correct for this client. The agency's value moves to the review. The checklist below is that review, grouped into six areas, with the hours it takes in a worked example.

Follow me on Instagram@sagnikteaches

What AI can draft and what a person must own

Different parts of a site carry different risk. The table sorts the common jobs.

Connect on LinkedInSagnik Bhattacharya
JobAI draft useful?Who must reviewWhat goes wrong if unreviewed
Page layout and section orderYes, quicklyDesignerGeneric layout, weak hierarchy, no fit with brand
Headlines and body copyYes, as a first draftEditor plus clientFalse claims, generic wording, wrong tone
Service and product factsPoor: only from client inputClient approves in writingInvented licences, prices, guarantees
Contact and quote formsBoilerplate onlyDeveloperForms that fail silently, spam, data sent to the wrong place
Image alt textYes, as a draftEditorWrong description, decorative images announced
AccessibilityPartial, via automated checksDeveloper and manual testerKeyboard traps, poor contrast, unlabelled controls
Page titles and descriptionsYesEditorDuplicates, truncation, wrong keywords
Redirects, domain and launch settingsNoDeveloperLost traffic, broken links, site left private
Cookie notices and legal pagesTemplate onlyQualified adviserStatements that do not match what the site does

The pattern is that AI contributes where the stakes are low and a fix is cheap, and humans own anything that touches accuracy, data or the law. What ChatGPT gets right and wrong when designing a client website goes through the same split for a general-purpose tool.

Subscribe on YouTube@sagnikteaches

A worked example: a seven-page pest-control site

Illustrative case. An agency builds a seven-page site (home, rodents, insects, commercial contracts, areas, reviews, contact) for a pest-control firm. The agency's usual hours for such a site are about 45. Using AI for layouts, first-draft copy, alt text drafts and metadata, drafting drops from roughly 22 hours to 10. The review adds hours that the old process did not have as a separate step.

StageBefore AI (hours)With AI (hours)Note
Brief and structure54Same: client input is the limit
Layout and design105AI drafts, designer adjusts
Copy and images125First drafts quick; fact-check takes the time
Build and forms108Boilerplate saved, wiring still manual
Review and testing511New: the checklist below
Launch and handover33Unchanged
Total4536Saving 9 hours, not the 12 the drafting alone suggests

The point of the example is the last row. Layout and copy hours fell by 12 and the brief and build saved three more, but review hours rose by six, so the net saving is nine hours. These figures are illustrative, and your own will differ, so time a few projects before you change a quote. Agencies that cut price by the drafting saving and skip review pay for it in rework.

The Complete Excel Course with AI Integration thumbnail

The Complete Excel Course with AI Integration

From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.

By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner

Buy now

Layout review: what to check on every template

AI layouts tend to look competent and sameish: hero, three columns, testimonials, call to action. The review is about whether the layout serves this client's main action.

  1. Main action above the fold. On a phone, can a visitor see the call button or quote link without scrolling? For the pest-control firm the answer should be a visible phone number.
  2. Hierarchy. Is there one clear headline per page and a logical order of subheadings? Models sometimes stack multiple large headings, which confuses both readers and screen readers.
  3. Real content in real sizes. Placeholder text hides problems. Paste the client's actual longest service name and a long review and see what breaks.
  4. Breakpoints. Test a small phone, a tablet and a wide desktop. Check that tables, menus and buttons still work and that nothing overflows.
  5. Brand fit. Colours, fonts, spacing and imagery should match the client's brand and not drift into generic defaults.

If the layout started from a wireframe step, AI wireframes from a brief explains how to keep the draft anchored to the brief.

Copy review: every claim traced to the client

AI copy reads well and is easy to wave through. That is the risk. Every sentence that states a fact about the business needs a source in the client's own answers or documents.

  • Claims. Highlight words like "licensed", "insured", "guaranteed", "award-winning", "same day", "years of experience". Each needs the client's written confirmation. A line such as "fully insured" that the owner never supplied is exactly the sort of invention to catch.
  • Numbers. Prices, response times, areas covered, opening hours and phone numbers must match the client's information character for character.
  • Voice. Read the home page aloud. If it could be any business in the trade with the name swapped, rewrite it using the client's own phrases.
  • Duplicates. Models reuse phrases across pages. Search the site for repeated sentences, which also helps search performance.
  • Placeholders. Search for square brackets, "lorem", "your company", "[city]" and sample names before anything is shown to a client.

Illustrative miss: a draft for the rodent-control page said "we use only humane methods". The client's questionnaire never said that, and some treatments are not. The owner had to correct it before launch. A one-line claim can create a complaint. For the writing step itself, how a web agency picks an AI content platform compares the tools.

Forms: the place where AI code most often fails quietly

Forms collect names, phone numbers and sometimes addresses, and they are the main conversion point. They deserve a developer's test, not a glance.

  1. Submit it for real. Fill in every form on the live or staging site and confirm the message reaches the right inbox, in the right format, with all fields present.
  2. Test the failures. Submit with a blank required field, an invalid email and a very long message. Does the visitor see a clear error?
  3. Spam protection. Is there a challenge, a honeypot field or rate limiting? An unprotected form will fill the client's inbox within weeks.
  4. Where the data goes. Confirm which service receives the submission, what it stores and who can see it. Check that it matches the privacy notice.
  5. Confirmation message and follow-up. Does the visitor see a confirmation, and does the client have a process to reply quickly?
  6. Labels and keyboard use. Each field needs a visible label and the form must work with a keyboard alone.

A realistic failure: an AI-built form posted to a placeholder address left in the code, so the first fourteen enquiries went nowhere. No visible error appeared. A single test submission on day one would have caught it, which is why the live test is the first line of the launch checklist.

Accessibility: automated checks are the start, not the finish

The W3C's Web Accessibility Initiative says plainly that some checks cannot be automated and that tools can sometimes give inaccurate results. That matters for AI-built sites, where the output often looks tidy and still fails basic tests. Aim for WCAG 2.2 level AA as a reasonable target if the client wants a stated standard, and say in your scope what you will test and how.

  • Run an automated checker on every page template and fix what it reports, then treat the clean result as the start.
  • Use the keyboard only. Tab through each page. Focus must be visible, order logical, menus usable and nothing trapped.
  • Check headings and landmarks. One main heading, subheadings in order, regions labelled.
  • Review alt text by hand. AI descriptions are often plausible and wrong. Decorative images should have empty alt text, and informative ones a short, accurate description.
  • Check contrast and zoom. WCAG's Resize Text criterion (level AA) expects text to scale to 200 per cent without loss of content or functionality, so zoom each template and look for clipped text and overlapping buttons.
  • Read with a screen reader at least once on the main page and the form.

Whether AI can replace a manual accessibility check goes deeper into what automation can and cannot judge, and the handover version of this list is in accessibility checks before handover.

Launch checks that only a developer should run

Most launch failures are settings, not design. Run these on a staging site first, then again on the live one.

  1. Visibility. The site is public, and any "discourage search engines" or noindex setting from staging is off. The reverse applies to the staging copy: make sure it is not indexed.
  2. Redirects. Old URLs map to new pages. List the old site's top pages, and test each one.
  3. Links and images. Run a link checker across the site and fix broken links and missing images.
  4. Titles and descriptions. Each page has a unique title and description of sensible length.
  5. Analytics and consent. Tracking is installed once, consent is handled according to the client's policy, and nothing fires before consent where that is required.
  6. Security basics. HTTPS everywhere, admin accounts with strong passwords and two-step sign-in, unused plugins or test pages removed.
  7. Backups and ownership. A backup exists, the client controls the domain and hosting accounts, and logins are handed over securely.
  8. Speed. Check the heaviest page on a phone connection. AI-produced pages often load oversized images.

Where you publish many similar sites from one template, a repeatable process saves time, but the review list stays the same each time: a template that passed once can still carry a new client's wrong redirect or noindex setting.

Where human review is mandatory, in one list

If you want a short rule for your team, use this: a named person signs off, in writing, anything that falls in these categories before it goes live.

  • Any statement of fact about the client's business, licences, prices or guarantees.
  • Every form, and every place personal data is collected, stored or sent.
  • Legal pages, cookie notices, privacy wording and terms.
  • Accessibility, to the level promised in the proposal.
  • Redirects, domain, hosting, indexing and tracking settings.
  • Any code the AI wrote that runs on a server or touches the client's accounts.

Everything else, such as layout polish and first-draft copy, still gets looked at, but errors there are visible and cheap to correct.

A sign-off sheet to copy

A one-page sheet per project gives you evidence of what was checked. A filled-in sample for the pest-control site (illustrative):

Project: Pest-control site, 7 pages
Layout: reviewed by designer, phone/tablet/desktop tested, [date]
Copy: every factual claim traced to client questionnaire; 3 claims removed, 2 confirmed by owner in writing, [date]
Forms: 2 forms test-submitted on staging and live; spam protection on; data goes to [service]; [date]
Accessibility: automated check clean on 4 templates; keyboard and screen-reader pass on home and contact; alt text edited by hand; [date]
Launch: redirects 14/14 tested; noindex removed; analytics once; backup taken; client holds domain login; [date]
Signed off by: [name], [date]

Keep the sheet with the project file. If a client later disputes a claim or an accessibility complaint arrives, you can show what was checked and by whom.

Telling the client what you did, and charging for the review

Review is real work and should show up in the proposal. List it as a line: "Quality review and testing, including forms, accessibility and launch checks". Clients rarely object to testing when it is named, and they object strongly to bugs after launch. If AI tools process client content, say so in your terms and name the safeguards, such as which content is entered and what stays out.

After launch, a maintenance plan is where the review habit continues: plugin updates, link checks, form tests every month. Website maintenance packages a web agency can add sets out how to package that. If you would rather set up a review process with someone, that is what my AI implementation consultation covers, at the contact page.

A second example: the form that worked and the page that did not

Illustrative case: a four-person cleaning company's site, built mostly by AI in a single afternoon. The review found three problems, and each shows a different part of the checklist at work.

The first was in the copy. The services page said the firm offered "end-of-tenancy and oven cleaning, plus carpet shampooing". The owner only offered the first two. Tracing each claim to the questionnaire caught it in five minutes, and the sentence was edited before the client saw the draft. The second was in the form: the quote form worked but offered no confirmation message, so visitors resubmitted, and the owner received duplicates. A developer added a clear confirmation and disabled the button after the first click. The third was accessibility: the mobile menu opened with a tap but could not be reached by keyboard. The automated checker had flagged nothing, and only the keyboard-only walk-through exposed it.

None of these errors was large. Together, they would have produced a wrong claim, a messy inbox and an inaccessible menu on day one. The review took about two hours, which is cheap compared with the cost of fixing the same issues after the client's customers had met them.

How much review depends on the size of the site

Review effort does not scale evenly. A rough guide, with illustrative hours:

  • One-page site with a contact form. About two hours: one form test, one accessibility pass, a claims check and the launch list. The form and the claims are the two things that matter.
  • Five to eight pages. About eight to twelve hours, as in the pest-control example. Per-page checks repeat, but templates make later pages faster.
  • Twenty pages or more. Review by template and by sample. Test every template and every form, read all copy for claims, and sample a fifth of the remaining pages for accessibility. The risk shifts to consistency, so keep a log of fixes that must be applied to every page of that type.
  • A shop or booking system. Add payment, tax settings, stock rules and order emails. These carry real money and personal data, so they need a developer's full attention and a test purchase.

The same rule applies to the content platform. If the client will edit the site after launch, the review also has to cover what they can break: roles and permissions, editing guidance and what happens to the page layout if they paste in text from another source.

Handing the client the review, not just the site

A short handover note turns your review into a selling point. Include what was checked, what was found and fixed, what the client is responsible for (accuracy of content they add, keeping accounts secure, renewing the domain), and what to do if something breaks. Keep it to one page in plain English. Clients who understand what you did tend to value it, and the note gives them a route back to you when something needs attention, which is the start of a maintenance relationship.

Common ways the review gets skipped, and how to prevent each

  • "It looks fine." Visual polish creates false confidence. Make the checklist a required step in your project tool, so the project cannot close without it.
  • Time pressure at launch. Schedule review before the launch date, not on it. Aim to finish the review a full working day ahead.
  • Only the builder reviews. The person who prompted the AI is the least likely to see its blind spots. Ask a second person to run the copy and form checks.
  • Client changes after sign-off. A late edit by the client bypasses the checks. Route late changes through the same sheet, even if only the relevant rows.
  • Tool updates. When you change the AI tool or template, the previous review habits may not hold. Re-run a test project before using it on a client.

Questions agency owners ask about AI-built sites

Can AI build a whole client website on its own?

It can produce a working first draft of layout, copy and basic forms quickly. It cannot take responsibility for accuracy, accessibility, security or fit with the client's business. Treat the output as a draft from a fast junior, and plan review time on every page before anything is shown to the client.

Which parts of an AI-built site need a developer's eyes?

Forms and anything that handles personal data, accessibility beyond what an automated tool reports, tracking and consent, redirects and launch settings, and any code that runs on the server. Layout and first-draft copy need a designer or editor's check but not usually a developer.

Should I tell clients a site was built with AI?

Be honest if asked, and say so in your terms if AI tools touch client content. Clients care that the result is accurate, accessible and theirs to own. Say what you reviewed, who did it and what the client is responsible for.

Does AI save money on client sites?

It usually shifts hours from drafting to reviewing. Track your hours on a couple of projects before you cut prices, because review and fixing can eat much of the time saved on a rushed first draft.

Further reads

Sources: W3C Web Accessibility Initiative guidance on selecting evaluation tools; WCAG 2.2 (W3C Recommendation, December 2024), success criterion 1.4.4 Resize Text; hours and cases are illustrative.

Want a review step built into your agency's AI workflow?

On a 1:1 call we can map where AI sits in your build process, decide which checks need a named person, and turn them into a checklist your team uses on every site.

Book a 1:1 call with me