Building AI Employees for Hospitality: How AITropos Takes Orders Where Customers Already Are

This episode of Just Now Possible features AITropos founders Santi Marchiori (CEO) and Juan Haedo (CTO) discussing how they built an AI order-taking agent that handles the full ordering flow inside WhatsApp for hospitality venues. The conversation covers their search for a niche after exploring many startup ideas, three product iterations (hardware for waiters, a waiter app, then a customer-facing agent), and the challenge of turning non-deterministic conversation into structured POS-ready order data. It explains architectural choices such as tools over MCP or rigid pipelines, pre-fetching product context for latency, and testing with thousands of simulated conversations before live deployment. It matters to PMs building AI agents because it shows how accuracy metrics, speed constraints, and evaluation infrastructure shape product decisions in a real domain.

01Key takeaways

• Explore broadly, then commit to a specific niche where the workflow pain is concrete and measurable.
• Iterate on the form factor: a hardware or internal tool may reveal the real problem before the customer-facing product is clear.
• Define one primary accuracy KPI, such as item identification, and let it drive architecture and evaluation priorities.
• Pre-fetch likely context in parallel so the agent can respond quickly without extra round trips.
• Use large-scale simulated conversations to catch regressions before shipping to real customers.
• Build reusable domain templates so onboarding time shrinks with each new customer.

02Key sections

Finding the wedge

The founders describe spending about two years exploring many startup ideas before settling on AI-powered order taking in hospitality. They emphasize choosing a specific niche rather than a broad vision.

Three product iterations

They moved from a hardware device for waiters to a waiter-facing app, and finally to a customer-facing conversational agent in WhatsApp. Each step revealed where the real value and friction sat.

Accuracy and speed as core constraints

Order item identification accuracy became the most important KPI, while real-time response speed drove architectural decisions. Together these constraints shaped how the agent was built.

Agent architecture and latency

They chose a tools-based architecture over MCP or fixed pipelines, pre-fetch product context in parallel, and inject relevant data through a fast immediate system prompt to reduce tool calls.

Testing and onboarding

Thousands of agent-simulated customer conversations run overnight validate changes before deployment, and onboarding time for new venues dropped from three months to a few weeks.

03From the post

“Listen to this episode on: Spotify | Apple Podcasts What does it take to build an AI that can take a food order over WhatsApp — correctly, every time, fast enough that customers can't tell it's not a person? That's the core challenge Santi Marchiori and Juan Haedo set”

Summary and takeaways written by PM Atlas; quotes are short excerpts. © the original author.