How to review your agent's work before a client sees it
How to review an AI-built website before the client sees it, a core habit of AI agents for client work: check facts, click everything, test on a phone, compare.
A web designer running a one-person AI agency asks her agent to finish the contact page for a family dentist. Twenty minutes later it reports: "Contact page complete. Added opening hours, map, booking form and emergency number." It looks good. She nearly sends the link. Then she reads the hours. Saturday says 9 to 1. The practice has never opened on a Saturday. The brief didn't mention Saturday at all, so the agent filled the gap with something that looked right.
That's the job of a review: to catch what looks right and isn't, before the client does. Check the site against the brief, not against the agent's summary. Read every word, click everything, try it on a real phone, run a few basic accessibility checks, and compare it with the last version the client approved. Only then send the link.
How do I check a website my AI agent built?
Work in the same order every time. It keeps a review short when the change is small, and thorough when the site is new.
- Read the agent's report, then set it aside. Ask what it changed and what it assumed. Treat the answer as a list of places to look, not as proof.
- Check the facts against the brief. Names, address, phone numbers, hours, prices, services, staff. Anything not in the brief is a question for the client, not a fact.
- Read every word. Look for placeholder text, sample testimonials, made-up awards, and sentences that sound good but promise something the business doesn't do.
- Click everything. Menu, footer, buttons, phone numbers, email links. Send every form and check where the message lands.
- Look at it on a real phone. Not just a narrow browser window.
- Run the basic accessibility checks. Page titles, image descriptions, headings, contrast, keyboard use, form labels.
- Compare with the last approved version. What changed that you didn't ask for?
If you already use a website QA checklist before handover, this is the same list, run earlier and more often. The difference is who you're checking. At handover you check the site. Here you're also checking the agent's guesses.
What mistakes do AI agents make on websites?
Mostly confident guesses. An agent is good at producing something complete, and a complete-looking page hides gaps well. The ones worth hunting for:
| What to look for | Why it happens | How to catch it |
|---|---|---|
| Hours, prices or services not in the brief | The agent fills gaps with likely details | Compare each fact with the brief, line by line |
| Placeholder text or sample reviews | Starter content never replaced | Search the pages for "lorem", "sample", "testimonial", "Jane Doe" |
| Links to pages that don't exist | The menu was made before the pages were | Click every link |
| Forms that send nowhere | The form looks finished but isn't connected | Send a test message and find it |
| Layout that breaks on a phone | Built and checked on a wide screen | Check on a real phone |
| Images without descriptions | Easy to skip | Run the Easy Checks |
| Changes you didn't ask for | The agent "tidied" nearby sections | Compare with the last approved version |
None of these mean the agent did a bad job. They mean it did a fast job without the knowledge only you and the client have.
Why compare with the last approved version?
Because agents change more than you asked. You ask for a new testimonial section and it also adjusts the header spacing, rewrites the footer and renames a page. Each change may be fine. The problem is not knowing about them when the client asks why the footer looks different.
Keep every version the client has approved, and look at the new one side by side with it. Go page by page. For each difference, ask: did I ask for this? If not, keep it only if you'd be happy explaining it.
This also protects you when the client says "it was fine last week". You can open last week's version and see.
Can the agent check its own work?
It can help, and you should ask it to. Three prompts are worth using at the end of every task:
- "List every file and page you changed."
- "List every fact you added that wasn't in the brief."
- "List anything you couldn't finish or weren't sure about."
The second one catches most invented details. The agent will often tell you, when asked directly, that it guessed the Saturday hours.
Still, its answer isn't the review. An agent that missed a problem while building can miss it while checking. Your review is the one the client is relying on.
It also helps to stop guesses at the source. Claude Code's documentation says each session begins with a fresh context window, and that instructions you keep in a CLAUDE.md file are loaded at the start of every session. Put the brief's facts there, or in the notes your agent reads, and add a line like "never add hours, prices or services that aren't in this brief". How to give your AI agent a client brief it won't forget covers how to set that up.
What should I check on a phone?
Most of a local business's customers will visit on a phone, and Google says it uses the mobile version of a site's content for indexing and ranking. So the phone isn't a final nicety; it's the main version.
On a real phone, check:
- Can you read the text without zooming?
- Does the menu open and close, and does every item go somewhere?
- Do phone numbers start a call when tapped?
- Does the booking or contact form work with a phone keyboard?
- Are the hours and address easy to find without scrolling far?
If the site was built with Claude Code for a client website, it's worth asking the agent to describe how each page behaves on a small screen, then checking that description against the phone in your hand.
What accessibility checks should I run?
The W3C's Easy Checks are a good first pass and need no special tools. They cover page titles, text descriptions for images, headings, colour contrast, using the site with a keyboard only, and labels on form fields. Agents often get the structure right and miss the details, such as an image description that just says "image" or a button with no label.
Run them on every new page, and on any page where the agent changed the layout. The website QA checklist includes them, along with the rest of the pass.
When is it ready for the client?
When you'd be comfortable if the client found nothing wrong and you found nothing wrong later. In practice: the facts match the brief, every link and form works, it reads well on a phone, it passes the basic checks, and you know every difference from the last approved version.
Then send the client that one version, with a short note on what changed and what you'd like them to look at. Ask them to approve that version, not "the site". If they ask for changes, the next round starts from here. Keeping your approval on everything the agent produces before it reaches a client is the same idea as approving what your agent sends.
Where Volant fits
Volant has version history for every website, so you can compare versions and go back to any earlier one. You choose which one goes live. Building, reviewing and delivering client websites shows how the review and the client's approval fit together.