From Zero to Launch: Building an AI Consulting Business with Claude in One Session

I just built the entire digital infrastructure for my consulting firm in one working session with Claude Code. Not a mockup. Not a plan. A production-ready website, an AI voice receptionist that qualifies leads, an AI-powered contact form that drafts personalized responses, a content operations system with autonomous research agents, a Slack command center, and a marketing strategy across 10+ platforms.

This is the article about how I did it — and what it says about where AI-assisted work is headed.

What got built

Here’s the inventory of what came out of a single session:

Website (Astro 5 + Tailwind v4 + Svelte)

• 33 pages: homepage, 5 service pages, 4 industry pages, about, personal brand page, contact, insights listing, and 19 blog posts
• Full SEO infrastructure: JSON-LD schemas, Open Graph tags, auto-generated sitemap, semantic HTML
• Responsive navigation with Svelte island components
• Google Analytics (GA4) and Microsoft Clarity for traffic and engagement tracking
• Lighthouse-optimized static output with zero JavaScript on content pages

AI Voice Receptionist (Vapi + Claude)

• An AI agent that answers the business phone, qualifies inbound leads, and captures their information
• Knowledge base built from website content so it can answer questions about services accurately
• Rate limiting (3 calls per number per day) with allowlisting for known numbers
• Call summaries and lead notifications delivered via Slack
• 3-minute max duration to prevent abuse

AI-Powered Contact Form (Claude + Resend)

• Every form submission is classified by Claude: qualified lead, general question, or spam
• Qualified leads get an instant personalized response referencing their specific inquiry
• I get notified in Slack for real opportunities; spam is filtered silently
• reCAPTCHA v3 gates bot submissions before they ever hit the Claude API

Content Operations System

• 19 articles live at launch (14 converted from Word, 5 written new)
• 4 Claude Code skills: research briefing, content planner, article builder, article deployer
• 5 autonomous research agents running weekly via Claude remote triggers, each covering a different domain (AI engineering, governance, leadership, industry watch, agentic AI)
• Research delivered to a dedicated Slack workspace every Monday morning
• Content syndication pipeline across Medium, DEV.to, Hashnode, LinkedIn Articles, Substack, Hacker News, and Reddit
• 11 additional articles in the backlog with full research briefs

Slack Command Center

• Dedicated workspace with channels for research briefings, leads, call summaries, content pipeline, and spam filtering
• All inbound signals (phone, web, research) route to Slack
• Single place to monitor the business, review leads, and trigger follow-up research

Infrastructure

• Azure Container App deployment with Bicep templates and automated deploy script
• Docker containerization for the webhook server
• Scale-to-zero hosting (~$5/month when idle)
• All secrets managed through Azure’s secret store
• CORS locked to the production domain

How the session worked

The process wasn’t “tell AI what to build and walk away.” It was a collaborative design session that followed a deliberate structure:

1. Context gathering. Claude explored my existing articles (16 pieces across AI governance, healthcare, product management), my resume, and 9 active projects spanning clinical AI, multi-agent manufacturing systems, MCP servers, and nonprofit compliance. This gave it the raw material to make informed recommendations rather than generic suggestions.

2. Strategic decisions through structured questions. Rather than making assumptions, the session worked through decisions one at a time: launch strategy (MVP with marketing ready), audience targeting (tech community for reach, enterprise buyers for conversion), content voice (practitioner, not theoretical), posting cadence (batch-produce, then maintain), tech stack (Astro for content-first static output).

3. Design before implementation. A full design spec was written and reviewed before any code was written. This included site structure, content categories, article inventory, social media channel strategy, the AI receptionist call flow, and launch sequence. The spec caught positioning issues early — like the fact that “Healthcare AI” was too narrow for a service that spans healthcare, manufacturing, and defense.

4. Systematic implementation. The website was built task-by-task: project initialization, global styles, components, layouts, then pages. Each task was dispatched to a focused subagent, verified, and committed. 15 clean git commits tell the story.

5. Iterative refinement. As I previewed the site, I caught issues and fed them back in real time: “Healthcare AI should be Applied AI,” “add Manufacturing to industries,” “remove my email from public pages,” “add Financial Services — I have 15 years there.” Each change was implemented, verified, and committed within minutes.

What I learned

AI is exceptional at breadth tasks. Creating 5 service pages with consistent structure, converting 14 articles to markdown with proper frontmatter, building a contact form with reCAPTCHA and server-side validation — these are tasks where the volume of work is the challenge, not the complexity of any single piece. AI turns a week of work into an hour.

Strategic decisions still need a human. Claude proposed options and recommendations, but every positioning decision — which industries to highlight, how to describe my background, what tone to strike — required my judgment. The AI couldn’t know that Vision2’s brand is damaged, or that “Healthcare AI” undersells my breadth, or that my financial services experience matters for credibility.

The compound effect is the real story. Any single piece — one service page, one article conversion, the contact form — is straightforward. The power is in doing all of it in one session with consistent quality, connected strategy, and no context loss between components. The website copy references the same positioning as the voice receptionist’s knowledge base, which matches the article categories, which align with the social media content pillars. That coherence is hard to maintain across weeks of manual work.

Building in public is the best marketing. This article exists because the process of building the business IS the content. Every technical decision — choosing Astro over Next.js, building rate limiting for the voice agent, using Claude for contact form classification — is a potential article, video, or social media post. The meta-layer of building an AI consulting firm with AI tools is the most authentic demonstration of expertise possible.

The cost

• Astro + Vercel hosting: Free tier
• Azure Container App: ~$5/month (scale-to-zero)
• Vapi voice agent: ~$0.10/minute (low volume at launch)
• Anthropic API (contact form): ~$0.003 per submission
• Resend email: Free tier (3,000 emails/month)
• reCAPTCHA: Free
• Microsoft Clarity: Free
• Google Analytics: Free

Total infrastructure cost at launch: under $15/month.

Where AI hit its limits

Not everything went smoothly — and the friction points are instructive.

Scheduled agents and OAuth don’t mix — until they do. We set up Claude remote triggers to post weekly research briefings to Slack. The first attempt failed silently — the Slack MCP connector works in interactive sessions but the remote agent runs in an isolated cloud environment. We assumed we’d need incoming webhooks as a workaround. Then we tested a simpler trigger with just a Slack message and no web research — and it worked. The issue wasn’t OAuth; it was that the first agent was doing too much research and likely timing out before reaching the Slack step. Lesson: isolate variables when debugging distributed systems. The MCP connector works in remote triggers; the agent just needs to be scoped tightly enough to complete within its execution window.

Platform limits shape architecture. We designed 5 specialized research agents (one per domain) plus a synthesis agent. The Claude plan allows 3 scheduled triggers. Rather than compromise on depth, we had to rethink the architecture — merging domains or rotating agents weekly. The constraint forced a better design. This is a pattern I’ve seen across 28 years: artificial limits often produce better systems than unlimited resources.

Account setup should have been automated from the start. We built the entire codebase with AI, but when it came time to create Slack apps, reCAPTCHA keys, and Vapi accounts, we initially fell back to manual instructions. Then we used Claude’s Playwright browser automation to create the Slack app — navigating the Slack API site, filling forms, dismissing dialogs. It worked, but it was clunky: dialog dismissal timing, multiple overlapping modals, element references going stale. Browser automation for account setup is possible but needs a dedicated skill with retry logic and common patterns. We built that skill for next time.

Article conversion needed cleanup. Pandoc converted .docx files to markdown cleanly, but some articles had embedded images that produced broken references. The build caught them immediately, and a one-line sed command fixed all of them — a reminder that automated pipelines need validation gates.

Research agents need curation, not just automation. The first research briefing that arrived in Slack was competent but shallow — it surfaced the same headlines I already knew (Amazon Kiro, Anthropic Pentagon, MCP Linux Foundation). The prompt needed to explicitly tell the agent what I already know and to dig beneath the surface. The second version, with domain-specific agents and instructions to “find what Carl doesn’t know,” was dramatically better. Lesson: AI research agents are only as good as their briefs, just like human research assistants.

These aren’t failures — they’re the normal friction of building real systems. The difference with AI is that each issue was diagnosed, fixed, and the fix was documented in minutes, not hours. And every friction point became content for this article.

What this means for AI consulting

If I can build the entire digital infrastructure for a consulting practice in one session, what does that say about the projects I help clients with?

It says the bottleneck isn’t engineering capacity. It’s strategic clarity. Knowing what to build, why to build it, and how it fits together — that’s the hard part. The AI handles the volume. The human handles the judgment.

That’s the thesis behind AI Implementation Group: we help organizations close the gap between AI ambition and AI reality. Not by adding more engineers, but by bringing the strategic clarity that lets AI tools deliver their full potential.

The website you’re reading this on is the proof.