Build an AI-powered booking assistant for a nail salon that lets customers book online without a phone call.
The salon tracks appointments on paper. Customers call to book, but workers are busy. There is no online booking and no way to get answers outside business hours.
The hard part is not building the tool. It is making the AI behave reliably inside a real product flow.
The Build
Someone I know needed an online presence for their business. They had no website, no online booking, no way for customers to schedule without calling. I saw an opportunity -but I did not want to build just another booking site. I was already deep into AI and large language models, working toward my master's in AI, and I wanted to tackle something real. Not a demo. Not a tutorial project. A product that could actually help a small business leverage something new.
The idea was an AI-powered booking widget. Not just a form that collects a name and a time slot. An assistant that understands services, knows worker availability, answers customer questions, and walks someone through booking without them having to search around or make a phone call.
The Problem
The salon I was building for still uses pen and paper to track appointments. Customers call to book, but the workers are busy -they are doing nails, not answering phones. Nobody has time to stop, grab the paper schedule, and check availability while a customer is waiting on the line.
There is no online booking. No way for a customer to check availability at midnight when they are planning their week. No way to get a quick answer about pricing or services without calling during business hours and hoping someone picks up.
This is not a rare situation. This is how a lot of local businesses still operate.
Assumption vs Reality
I assumed the hard part would be building the tool itself: the frontend, the chat interface, the widget embed. I figured if I wrote a good enough prompt, the AI would know exactly what to do. Collect the service, find a time, confirm the booking. Clear instructions in, correct behavior out.
Building the widget was the easy part. Getting the AI to behave reliably was not. The LLM thinks for itself. It would skip steps. It would ask for information it already had. It would try to be helpful in ways that broke the booking flow. The real work was containing the AI -defining what it can and cannot do, making it understand user intent without going off-script, and building the structure around it so it stays useful instead of unpredictable.
The Fix
I stopped thinking of this as a chatbot project and started thinking of it as a product with an AI layer. The booking flow needed structure -state management, clear rules about what questions to ask and when, guardrails around what the assistant should and should not try to do.
The prompt matters, but it is not enough on its own. The product needs to track what it knows, what it still needs, and what it should never assume.
Building something that sounds easy is not easy. An AI booking widget sounds like a weekend project until you actually try to make it work reliably for a real person at a real business. The problems you run into while building are the real curriculum. Every broken flow teaches you something about how these systems actually work -not how they work in a demo.
For a salon owner, this is about one thing: stop losing customers because nobody answered the phone. If a customer can book online at midnight, ask a question about gel nail pricing, and get a confirmed appointment without a single phone call -that is real value. For any small business, the question is the same. How much business are you losing because your booking process depends on someone being available to pick up the phone?
Builder Notes
My background before this was SAP consulting: enterprise data, client requirements, understanding what the business actually needs versus what looks good in a spec. That mindset carried over. Get the product working first. Solve the real problem. Do not chase perfect architecture before the core flow works.
Technical Note 1
The tech stack is Next.js, Supabase, and OpenAI. Nothing exotic. The complexity is not in the tools, it is in getting the AI to work inside a structured product flow.
Technical Note 2
Everyone in the AI space right now is building for tech companies and developers. Almost nobody is building AI tools for the nail salon, the barbershop, the small gym that still tracks everything on paper.
Technical Note 3
Going where the problem is real and the competition is thin is a better strategy than building the same tool as everyone else.
What I'd Do Differently
I would spend less time on the prompt early on and more time designing the booking state up front. The prompt feels like the obvious place to start, but the real leverage is in the product structure -what fields does the assistant need, what order should it collect them, and how does it know when it is done.
I would also talk to the business owner more before building. Not about features -about workflow. How do they actually handle a booking today, step by step? That understanding shapes the product more than any technical decision.