Application example. The scenario below was put together from what ProdMan builds in real projects, without representing a specific client. The figures shown are project estimates and vary with message volume, catalog size and the rules of each business.
The scenario
A company receives hundreds of WhatsApp messages a week. Most of them are the same repeated question: price, delivery time, payment methods, address, opening hours, order status, availability in the calendar.
The team answers between one task and the next. A message arriving at 10pm sits there until 9am the following day. A customer who waits too long goes to the competitor who answered first. And nobody can say, at the end of the month, how many conversations turned into sales.
Process first, AI second
Before building anything, ProdMan reads the last few months of conversations and groups them by intent. Out of that comes the list of what the assistant answers on its own, what it answers by querying a system and what it routes to a person.
That reading also surfaces what the company has not settled yet: a discount policy with no written criteria, a lead time each salesperson quotes differently, an outdated price list. Where the company has no standard answer, neither does the AI. Sorting that out is part of the project and is often worth as much as the assistant itself.
How it works
- Official number connected through the WhatsApp Business API. The company's number stays the same, with the official account badge.
- Conversational assistant loaded with the business rules, the catalog and the tone of voice. It talks in plain language, with no "press 1" menu.
- Integrations with what already runs in the company: price list, inventory, calendar, ERP, CRM or spreadsheet.
- Handover to a human with a summary of the conversation and of what was already answered. The customer does not repeat the story.
- Operations panel with the history, what was resolved on its own, what became a human conversation and what nobody could answer.
What it does and where it stops
The limit is declared in the design, before the first line of code:
- It does: answer frequent questions, quote price and lead time straight from the official source, check order status, qualify the lead, book a time, send the catalog, payment links and documents.
- It stops: negotiations outside the price list, complaints, cancellations, sensitive topics and any question outside its knowledge base. In those cases it calls in a person and tells the customer someone will take over.
An assistant that tries to answer everything is the one that makes things up. Ours is built to admit when it does not know.
How we thought about the risks
AI can get things wrong, and a mistake here happens right in front of the customer. The design accounts for that from day one.
- Price, lead time, stock and calendar always come from the official source, queried live, never from the model's memory.
- Without enough information, the assistant routes the conversation instead of risking an answer.
- Conversations logged and auditable, with consent and personal data handling defined in the design.
- Periodic review of the conversations that went to a human. Each one becomes an improvement to the knowledge base.
- Cost per conversation monitored, with a configured ceiling and a consumption alert.
- First reply in seconds, at any hour, including weekends and holidays
- Between 50% and 70% of conversations resolved without pulling in a person, depending on catalog size and how clear the rules are
- The team recovers the hours spent today answering the same question several times a day
- A full record of what customers ask, useful for pricing, product and campaign decisions
- Figures are project estimates, calibrated during the diagnostic against each company's real volume
What the company needs to have
- A business WhatsApp number, with administrator access.
- An up-to-date catalog, price list and service policy, even if only in a spreadsheet.
- One person responsible for receiving handovers and answering whatever falls outside the standard.
How this becomes a project
It starts with the free 45-minute diagnostic, where we measure volume, map the types of conversation and size the work. The build happens as a Short Project, 2 to 4 weeks: mapping, knowledge base, integrations, testing against real conversations, go-live and training for whoever will operate it. Delivered with documentation, a 90-day warranty and a client team that knows how to run it.