Automating a business end to end
04 3 min read
I built and automated XitPlan end to end, connecting lead generation, customer communications, onboarding, and delivery through AI agents, APIs, and cloud infrastructure. I owned the architecture and operations from September 2025 to September 2026. This case study follows the business workflow and its engineering decisions.
One person, whole pipeline #
While building and operating XitPlan, every AI idea I wanted to test had to earn its place inside a real business. That constraint shaped everything. A one-person company cannot afford heroics, so each stage of the pipeline, from first contact to a live client site, had to become a repeatable system with clear points for human decisions.
Filling the top of the funnel #
Outreach runs from the CRM I built: calling, texting, and email in one place, with AI summarizing calls and coaching live during them. The marketing site does the warm half of the job, in English and French, with lead magnets that do something useful before asking for anything: a website health-check tool that audits a business's site and explains what it found, a site chat that answers questions from a curated knowledge base, and a browser demo where a visitor can talk to the voice agent directly from the page.
Pricing that lives in one place #
No price appears anywhere in the codebase. Checkout resolves prices from Stripe at runtime by lookup key, so changing an offer is a dashboard action, not a deploy, and the site can never disagree with the billing system. Subscriptions run in more than one currency through the same mechanism.
Contracts without paper #
A sale is not a handshake; it is a signed agreement. The checkout flow includes an e-signature step that generates the contract as a PDF, and payment and signature travel together. When Stripe confirms the subscription, a webhook marks the agreement paid, applies any billing credits, and notifies both sides. The legal and financial state of an engagement is machine-checkable from the moment of signature.
Delivery as a system #
Client sites are content-as-code: each one is built from a maintained template, with the business's verified facts kept in a source document and structured data files. The template enforces a zero-fabrication rule I care a lot about: nothing ships on a client's site unless it traces back to that verified source. Improving the template improves every future site, and each client's site stays isolated in its own repository, deployed to the edge.
Everything reports back #
Every live site's contact form feeds leads back into the pipeline, and the voice agent answers the phones as a usage-billed product on top. The loop closes: outreach fills the funnel, the funnel converts, delivery ships sites, and the sites themselves generate the next round of leads. Building that loop, and the discipline of making every piece of it observable and repeatable, taught me more about production AI than any single feature in it.