In January 2026, I shipped our first client website in four hours. Not a landing page. A full marketing site: homepage, about, services, portfolio, contact. Custom design. Real copy. Deployed and live.
The traditional agency I worked at before would have quoted that project at €15,000 and delivered in six weeks.
This article is a transparent look at how Kaizen works: the tools, the process, the economics, and the things that go wrong. I am sharing this because the web industry is going through a fundamental shift, and I think business owners deserve to understand what is actually happening behind the curtain.
The tools that make this possible
Two AI tools do most of the heavy lifting: Claude Code and Google's Gemini.
Claude Code is an AI coding agent that lives in my terminal. I give it instructions, and it writes, edits, tests, and deploys code. Not snippets or suggestions. Full files, full components, full pages. It understands the project structure, remembers decisions from earlier in the session, and can work across multiple files simultaneously.
I use Claude Code for all the structural work: building page layouts, writing CSS, implementing responsive design, setting up navigation, creating forms, optimizing images, and deploying to our hosting infrastructure. Work that would take a developer a day or two takes minutes.
Gemini handles content generation. Given a client brief (their industry, target audience, value proposition, and tone preferences), Gemini produces first-draft copy for every page. Headlines, body text, calls-to-action, meta descriptions, alt text. The output needs human editing, always, but it eliminates the blank page problem that makes content creation the bottleneck in most web projects.
Beyond these two, the stack is deliberately simple. Sites are static HTML, CSS, and minimal JavaScript. Hosted on Cloudflare Pages (fast, global CDN, free tier covers most projects). No WordPress, no React framework, no build pipeline complexity. The simpler the output, the faster it loads, the easier it is to maintain, and the fewer things can break.
The process, step by step
Every project follows the same sequence. I have refined this over dozens of builds, and each step exists because skipping it causes problems.
Step 1: The brief (30 minutes)
Before any AI touches anything, I need to understand the client's business. Not their preferences for colors and fonts (those come later). The fundamentals: Who are their customers? What problem do they solve? Why should someone choose them over alternatives? What action should a website visitor take?
I capture this in a structured brief document. One page. The brief is the most important document in the entire project because every downstream decision flows from it. A weak brief produces a generic site regardless of how good the tools are.
Most clients have never articulated their value proposition clearly. The brief conversation is often the first time they think about their business from their customer's perspective. This step alone is worth a significant portion of the fee.
Step 2: Content generation (45 minutes)
The brief feeds into Gemini via a structured prompt template. The template encodes everything I have learned about what makes website copy effective: lead with the customer's problem, not your solution. Use specific language, not adjectives. Every paragraph earns its place or gets cut.
Gemini produces a first draft of all page content. I then spend 30 to 45 minutes editing: cutting generic phrases, sharpening the value proposition, making claims specific ("we reduced client onboarding time by 60%" instead of "we streamline your operations"), and ensuring the tone matches the brand.
This editing pass is critical. AI-generated copy without human editing reads like AI-generated copy. It is serviceable but not persuasive. The difference between a 2% and 8% conversion rate often comes down to the specificity and credibility of the copy. I wrote more about this in why your small business website has a 4% conversion rate.
Step 3: Design and build (2 to 3 hours)
This is where Claude Code does its work. I start with a design direction (based on the client's industry, brand personality, and any existing visual assets), then instruct Claude Code to build the site section by section.
A typical sequence: "Build a hero section with a dark background. Headline: [from brief]. Supporting text: [from brief]. Primary CTA button linking to contact. Secondary text link to services page. Full-bleed layout, 100vh on desktop." Claude Code writes the HTML and CSS, I review it in the browser, request adjustments, and move to the next section.
The back-and-forth is fast. Building a complete page takes 20 to 40 minutes. An entire 5-page site takes 2 to 3 hours. The AI handles the mechanical work (responsive breakpoints, CSS grid calculations, image optimization, accessibility attributes). I handle the creative decisions (what goes where, what gets emphasized, what gets cut).
We maintain a library of section patterns that have been tested across multiple projects. This is not a template. Each site is custom designed. But the underlying structural patterns (how a testimonial section is marked up, how a pricing table handles responsive layout, how a contact form validates input) are proven and reusable. This compounds: every site we build makes the next one faster and more reliable.
Step 4: Human review (1 to 2 hours)
This is the step that separates AI-assisted work from AI-generated work, and it is the step that most people underestimate.
I review every page across three dimensions: brand coherence, conversion effectiveness, and technical quality.
Brand coherence: Does every page feel like it belongs to the same company? Is the tone consistent? Do the visual elements support the brand personality or contradict it? AI is surprisingly bad at maintaining brand consistency across pages. It tends to drift, using slightly different color tones, varying type sizes, or shifting the voice between sections.
Conversion effectiveness: Is the most important information above the fold on every page? Are CTAs visible and compelling? Is social proof positioned at decision points? Does the page flow guide the reader toward action? This is judgment work that requires understanding human psychology, not just design patterns.
Technical quality: Does the site load fast? Is it accessible? Do all links work? Does it look correct on mobile, tablet, and desktop? Are images properly sized and optimized? I run Lighthouse audits and fix any issues.
The review consistently catches problems. In a typical project, I find and fix 15 to 30 issues. Some are small (a margin that is slightly too tight on mobile). Some are significant (a CTA that is positioned below the fold when it should be above it, or copy that buries the value proposition under three paragraphs of context).
Step 5: Client presentation and revisions (varies)
The client sees a live preview site within 72 hours of the brief. Not a mockup, not a wireframe. A working website they can click through on their phone.
Most clients request one round of revisions. Common requests: "Can we change the hero image?" "Can we reword this headline?" "Can we add a team section?" These take 30 minutes to an hour. Occasionally a client wants a more significant restructuring, which might take half a day.
The speed of the revision cycle is important. In traditional agencies, revisions take days or weeks because they go through project managers, designers, and developers. Here, I can make a change and show the client the result in minutes. This builds confidence and reduces the total number of revision rounds.
Step 6: Deployment and handover (30 minutes)
Once the client approves, the site goes live. We deploy to Cloudflare Pages, which means instant global distribution via CDN, automatic HTTPS, and near-zero hosting costs. DNS configuration takes a few minutes. We set up basic analytics (privacy-respecting, consent-gated), configure any custom email addresses, and verify that everything works across browsers and devices.
The client gets a simple guide for any content updates they might need to make. For static sites, this usually means they email us with changes and we implement them quickly. For clients who want more control, we can set up a lightweight CMS or content editing workflow.
Total elapsed time from brief to live site: typically 3 to 5 business days. Not because the work takes that long, but because client review and feedback cycles take time. The actual build work happens in concentrated bursts of a few hours each.
The economics
Let me be specific about the numbers.
Total time per project: 6 to 10 hours, including the brief, content, build, review, and one revision round. Complex projects with more pages or custom functionality can take up to 15 hours.
AI tool costs: Claude Code and Gemini usage for a typical project costs €30 to €60 in API fees. This is the direct cost of the AI work.
Hosting: Cloudflare Pages is free for most projects. Annual domain registration costs €10 to €15.
Our price to the client: Starting at €2,500 for a standard go-to-market site. This is not a loss leader. The margins work because the time investment is 6 to 10 hours of skilled work, not 60 to 100 hours of team coordination.
Compare this to a traditional agency charging €15,000 for the same deliverable. Their cost structure looks different: 20 hours of design, 40 hours of development, 10 hours of project management, 10 hours of account management, plus overhead. Eighty hours of work across four or five people, stretched over six weeks, producing roughly the same output.
The difference is not that our work is lower quality. The difference is that AI eliminated the mechanical work and small teams eliminate the coordination overhead. To see how this compares to the full range of options, read how much does a website cost in 2026.
What AI gets wrong (and why humans still matter)
I want to be honest about the failures, because the hype around AI tools often obscures the real limitations.
AI does not understand your business. It can process a brief, but it cannot tell you that your value proposition is weak, that your target market is too broad, or that your pricing is misaligned with your positioning. That is strategic thinking that requires human experience and judgment.
AI generates average design. Given a prompt, AI produces the statistical average of what it has seen. This means the output is competent but never surprising. It will not invent a new interaction pattern. It will not find the unexpected visual angle that makes a brand memorable. Distinctiveness requires a human eye.
AI writes safe copy. AI-generated text avoids risk. It does not take a strong position. It does not use humor unless prompted. It defaults to corporate-friendly language that reads like every other business website. The copy that converts best is specific, opinionated, and sometimes deliberately provocative. AI needs a human editor to get there.
AI misses context. It does not know that your competitor just launched a rebrand, that your industry has a regulatory change coming, or that your customers care more about speed than price. Context is what makes a website relevant rather than just correct, and context comes from the brief conversation and from understanding the market.
AI makes confident mistakes. This is the dangerous one. AI code will sometimes implement something incorrect with complete confidence. A contact form that looks right but does not actually send emails. A responsive layout that works on common screen sizes but breaks on a specific tablet. A meta tag that is syntactically valid but semantically wrong. Every output needs to be tested by someone who knows what correct looks like.
What I would tell someone starting this today
If you are a designer, developer, or marketer thinking about building an AI-assisted web service, here is what I have learned.
Start with the process, not the tools. The AI tools change every few months. The process (brief, content, build, review, deliver) is stable. Get the process right first, then optimize each step with whatever tools are best at the time.
The brief is everything. A good brief makes every subsequent step faster and better. A bad brief wastes AI tokens and your time producing output that misses the mark. Invest your human time here.
Build a section library. Every project should make the next one faster. Extract reusable patterns. Document what works. The compounding effect is the real competitive advantage, not any individual tool.
Be transparent with clients. We tell every client that we use AI tools. Not as a selling point or an apology, but as a fact. Most clients do not care how the sausage is made. They care about the result, the timeline, and the price. The ones who ask follow-up questions are usually fascinated, not concerned.
Never skip the human review. This is where you earn your fee. The AI produces the first 80%. The human review turns it from adequate to effective. If you are just shipping AI output without review, you are not running an agency. You are running a slightly more expensive AI builder, and you will be competing on price against tools that cost €20 per month.
The web industry is changing faster than most people realize. Agencies that ignore AI will be too slow and too expensive. Businesses that rely entirely on AI builders will get generic output that does not differentiate them. The sweet spot, for now, is human judgment amplified by AI execution. If you want to see exactly what that sweet spot delivers, check our packages or read about what a new business actually needs from its website.
That sweet spot will keep shifting. But the human part (the strategy, the taste, the quality judgment) is not going away any time soon. If anything, as AI makes production cheaper, the strategic and creative layer becomes more valuable, not less. And if you are still weighing whether your business even needs a website when social media seems sufficient, here is why you almost certainly need both.