I’ve been closely tracking how agentic AI reshapes frontline operations, and few case studies are as instructive as AITropos. Their north star is deceptively simple: take a food order over WhatsApp — correctly, every time, fast enough that customers can’t tell it’s not a person. That’s the challenge Santi Marchiori and Juan Haedo embraced, and it’s a masterclass in product strategy, conversation design, and systems engineering.
What they’ve built is an AI order-taking agent that handles the full flow — menu recommendations, modifiers, delivery zones, payment links, and status updates — entirely inside WhatsApp. Choosing the customer’s preferred channel wasn’t just a UX decision; it set the bar for speed, reliability, and trust. In hospitality, seconds matter. Latency becomes brand.
Their path to this solution reflects disciplined continuous discovery. They spent two years exploring hundreds of startup ideas before finding the niche of AI-powered order taking in hospitality, then iterated through three product forms — hardware for waiters, a waiter app, and finally a customer-facing WhatsApp agent — before landing on the right form factor. In my experience, this is what real product-market fit lessons look like: follow the problem, not the artifact.
Under the hood, the hardest problem is translating "non-deterministic human conversation" into structured "POS-compatible order data." To hit real-time response speed requirements, they chose a "tools-based architecture" over "MCP" or pipelines. That decision minimizes orchestration overhead and keeps the agent focused on the shortest path from intent to action — a pragmatic approach I recommend when SLAs are tight and context changes fast.
They also engineered for throughput and precision. A parallelized pipeline searches for multiple products simultaneously and pre-fetches product context before the agent even calls a tool. Complementing that, smaller, fast sub-agents assemble an "immediate system prompt" that injects relevant data into each turn without extra tool calls. Think of it as a retrieval-first pipeline designed to slash latency while preserving accuracy — a pattern every team building AI workflows should study.
Focus is evident in their KPIs. They identified order item identification accuracy as their single most important KPI. Picking one metric that truly governs customer trust is a hallmark of strong product management; it clarifies trade-offs in model selection, prompt engineering, and fallback behavior.
Quality assurance is equally rigorous. Before going live in any new venue, they test with thousands of agent-simulated customer conversations overnight. This approach de-risks deployment, surfaces edge cases early, and provides the data backbone for Agent Analytics and iteration. It’s a practical blueprint for teams operationalizing LLMs for product managers who need both scale and safety.
Operationally, the payoff shows up in onboarding. They reduced new customer onboarding from three months to a few weeks — and continue to shrink it as they build domain templates. Standardizing schemas, prompts, and flows for repeatable segments is exactly how you turn bespoke wins into a scalable go-to-market engine.
Stepping back, a few lessons stand out for product leaders building agentic AI in high-velocity environments: meet customers where they already are (WhatsApp), pick an architecture that serves your latency constraints (tools over complex workflows), pre-inject context to reduce tool calls, simulate at scale before launch, and anchor teams around one trust-defining KPI. Do these consistently, and you transform AI from a novelty into an always-on employee your customers actually prefer to use.
Inspired by this post on Product Talk.
Leave a Reply