Mozt's Vita is an AI WhatsApp receptionist for clinics in Latin America, evolving into a two-sided healthcare and wellness platform. Pilot-ready as of mid-April. The architectural bet is short to describe and slow to build: keep the agentic AI loop (an AI that takes actions on its own, not just answers questions), but wrap it in process flow and guardrails dense enough to deploy in healthcare. I started building it because I watched a clinic owner in Guatemala City answer a WhatsApp message at eleven on a Sunday night. His receptionist was asleep. The patient was almost certainly going to book somewhere else by morning if he didn't pick up the phone himself. ## The version I almost shipped Most teams building AI for small businesses ship the lazy version. Drop a chat interface, wire it to an LLM with tools, push to production Friday afternoon. The architecture is "ask the model, hope for the best." On a normal product it works fine. I built that version first. It survived about ten minutes of thinking through what could go wrong inside a clinic. A patient mentions an allergy. A teenager asks about a parent's prescription. Someone asks for a price the model has never seen, and gets a confident-sounding hallucination back. The clinic books that price into the calendar. The owner finds out at intake the next day. That's a refund and a public review nightmare on a good day. On a bad day someone gets hurt. So I scrapped Friday's version and went the other way. The other way is to keep the agentic AI doing the agentic work, and surround it with layers of safety and process that turn "ask the model, hope for the best" into "the model handles what it's good at, and the system handles everything else." Inputs get screened before the agent runs. The agent operates inside a constrained workspace where it can only act on verified clinic configuration. Sensitive content classes (medical questions, complaints, anything involving a minor) route to humans before the agent gets a chance to improvise. Every decision the system makes is logged in a structured row a clinic owner can query without having to ask me for it. That last sentence is the difference between something that demos well and something a healthcare operator will actually trust with a real WhatsApp number. Easy on Friday, painful on Monday. The harder version keeps the model where it earns its weight, fences it where it doesn't, and shows the operator exactly what it did. A patient sends a WhatsApp message. The system identifies the clinic, verifies the request, and loads the clinic-specific configuration the agent will work inside. Nothing routes to the agent until the clinic context is loaded. Sensitive content classes (medical questions, complaints, minors, human-handoff requests) get filtered out before the agent runs. The agent never has to decide whether something is safe. That decision is upstream. The agent takes the request inside a bounded workspace. It can only act on the clinic's verified configuration. It does not invent prices, hours, or policies. It does not freelance outside its scope. Schedule, answer, escalate, or capture the message as a lead. Each class has its own safety profile and its own response treatment. Every decision the system made gets recorded as a structured row. Operators see real-time transcripts, can pause the bot per patient or clinic-wide, and receive an instant DM on every escalation with a subject line that says exactly what kind of escalation it is. ## What it is, what it's becoming The clinics product started from a commercial observation, not a technical one. Clinics in Latin America run patient acquisition and retention on WhatsApp, and they lose money in four specific places. Inbound latency: the receptionist can't reply at 9pm Sunday. Booking ping-pong: six to eight messages to schedule one cita. No-show drift: twenty to thirty percent of scheduled revenue walks. Zero recovery on missed appointments: a no-show is a permanent lost patient. I didn't need a smarter chatbot. I needed to take the receptionist bottleneck off the inbound and replace it with something that runs twenty-four hours a day in the clinic's own tone. What shipped first was the engine. An agentic AI bot wrapped in the safety and process layers above. Multi-tenant from day one (think apartment building: one engine running the whole place, every clinic in its own sealed unit). Every patient turn produces a structured audit record: what the system saw, what it decided, what it did, how confident it was. The audit is queryable, so a clinic owner can answer questions like "how many escalations went to a human last week" without having to message me about it. The cockpit came next. The clinic owner doesn't run the bot. They supervise it. Real-time transcript view, manual reply, pause-bot per patient or clinic-wide, manual booking, owner DM on every escalation with a subject line that differentiates medical emergency, minor consent, complaint, and human request. The lesson from earlier work was sharp: never draw a button that doesn't have a handler. Every UI element maps to a real action. That whole stack cleared the pilot-readiness bar in mid-April. Tests passing, type errors zero, audit rows queryable, escalations going to the right place. What changed after that wasn't the engineering. It was the product. The clinics work was always meant to be the beachhead, not the company. Mozt's longer thesis is that the same pattern (agentic AI inside process flow, supervised by a real operator cockpit) transplants to other small-business workflows where mistakes are expensive. So the next product, Vita, took the engine and reshaped the surface around it. Vita is two-sided. A patient-facing PWA where people in Latin America can find a fisio, a psicólogo, a dentist, or a nutricionista, book in three steps, and chat in-app. A practitioner-facing web tool where the practitioner runs their day with calendar, clinical notes, services, patient list, and team management for clinics with more than one practitioner. And an internal operations console for the team supporting the platform. The agentic AI lives inside Vita's practitioner surface, wrapped in the same safety and process flow that runs the clinics product. The repetitive workflows that today need a receptionist's attention (confirmation chains, reschedule coordination, no-show recovery, follow-up loops) get routed through that process. The agent does the work. The structure around it keeps the work inside the lines. This is also the first product I'm actually designing and battle-testing for the market. Not a side project. Not a course exercise. Real clinics, real money, real people on the other side of the messages. The lazy version is to drop an LLM (the brain behind ChatGPT or Claude) in a tool-using loop and ship. Easy on Friday, painful on Monday when an owner gets paged. The harder version keeps the agent for what only an agent can do (natural language, ambiguous patient requests, multi-turn context) and builds everything else as structure around it. Input screening, action constraints, audit, supervision, override. The intelligence is what makes the product feel modern. The structure is what makes it deployable. The bot is not unattended. Every patient interaction is visible to the clinic owner in real time. Every escalation produces an instant DM. The operator can pause the bot per patient or clinic-wide with one tap. The cockpit got built before the bot got polished, because trust in healthcare comes from visibility, not from confidence in the model. Every decision the system makes is recorded as a structured row, queryable by operators after the fact. Multi-tenancy is database-enforced, not application-enforced. If something goes wrong I can reconstruct exactly what the system saw, what it decided, and why. That recoverability is what lets one engineering team run the same engine across many clinics without losing sleep. Vita's patient-to-practitioner relationship is direct. A doctor who moves between clinics carries their patients with them. Reviews, ratings, clinical notes, and chat history attach to the practitioner, not the clinic. Workspaces (clínica, consultorio, equipo) are an optional grouping on top. The bookable unit is always a person, never a room. I'm deliberately not sharing pricing, specific clinic names, partner relationships, internal architecture, or the operator console. Those parts of the product live in pilot-grade discretion until I'm ready to talk about them publicly. ## What the easy version costs The default engineering choice when shipping AI is to lean entirely on the model. The harder choice is to keep the model where it earns its weight, surround it with structure where it doesn't, and accept that the system you actually want is six months of work, not a Friday afternoon. Treating the agentic AI as one component inside a larger process, with its own quality gate and its own failure mode, was the IE move. Operators trust systems they can supervise. Models alone are not supervisable. Models inside a process are. That's the bet. The clinics product is in pilot. Vita is next.