A customer hesitating between two liqueurs asks in the evening, rarely at eleven in the morning. At a Swiss craft liqueur maker working with Tandem, those requests landed outside shop hours and waited until the next day. The conversational agent now answers on WhatsApp and on the website, in three languages, at any hour.
The case is small, and that is exactly what makes it useful. It shows what a craft business can put in place without a technical team, and above all the decisions to make before writing a line. On the difference between an assistant, a workflow and an agent, our article on workflows, assistants and AI agents sets the vocabulary.
The problem, questions arriving when the shop is closed
A craft liqueur sells through advice. Which pairing with which dish, what strength, what format, what is in stock. In the shop, the producer answers in thirty seconds. Outside opening hours, the question goes unanswered, and some of those requests never come back.
Three constraints specific to this trade come on top. Selling alcohol requires an age check. The customer base speaks French, German and English. And the catalogue moves with each production run, so any content frozen inside a tool becomes wrong within weeks.
What the agent does, step by step
The agent answers in the language of the first message and handles four requests. An order to prepare, a tasting appointment, a food and liqueur pairing, or a callback. Every completed request triggers two emails, a confirmation to the customer and an alert to the producer, in under a minute.
Conversation history is kept per contact. The producer can see what someone asked last time, which changes the quality of the callback.
Four spreadsheet tabs, and nobody calls the agency
The most important design decision in the project is barely visible. All the content the agent uses lives in a shared spreadsheet, across four tabs, synced with its knowledge base.
Opening a WhatsApp channel, the decisions to make before coding
This is the part projects underestimate most. Connecting an agent to WhatsApp involves a series of administrative choices that have nothing to do with artificial intelligence, and they drive the schedule.
The first choice sets the free WhatsApp Business app against the WhatsApp Business Platform, which exposes an interface for programming. The app is enough for manual use. Connecting an agent that answers on its own requires the platform. The second choice is the number, and that is where most of the myths sit. Meta's documentation on business phone numbers is explicit on this point.
- No SIM card is required. The number has to be able to receive an SMS or a voice call, and that is the only real constraint.
- A mobile is recommended, but a landline, a VoIP number or a toll-free number will do, by asking for voice verification rather than SMS.
- A number already registered on WhatsApp must be deleted from the existing account before it can be used. That is the most common failure on a shop number already in use.
- Two-step verification is mandatory at registration, and the PIN is asked for again to remove the number later.
- A new portfolio starts at two numbers, rising to twenty after business verification.
What a WhatsApp agent actually costs
Meta moved to per-message billing in July 2025, replacing the 24-hour conversation model. The official pricing documentation distinguishes four categories, and the difference changes everything for a small business.
- Non-template messages are free inside the service window opened by the customer. An agent answering inbound requests therefore costs very little on Meta's side.
- Marketing templates are always charged, whenever they are sent.
- Utility and authentication templates are charged outside the service window, and utility ones are free inside it.
The practical consequence is simple. An advice agent, triggered by the customer, stays in a low-cost zone. An outbound campaign changes category and budget. On top sit the conversational platform subscription and the model usage cost, marginal at this volume.
Going step by step, and owning what you leave out
Connecting an online calendar had been identified as a useful brick during scoping. It was taken out of scope, and it is the decision Tandem is happiest with on this project.
The reasoning has three parts. The real volume of appointments was unknown, so the gain was theoretical. The brick added an integration to maintain for a business with no technical team. And it pushed the go-live date back, when the whole point was to see real conversations arrive quickly.
- Ship the journey that carries the value, here product advice and order taking, before anything else.
- Collect by hand whatever volume is unknown. Appointments are confirmed by email, until there is something to measure.
- Ship a handover guide with the agent, so the client owns the decision on what comes next.
- Reopen the subject on numbers. A brick earns its place the day volume makes the manual version painful.
Adding a brick because it is technically possible is the surest way to ship a tool nobody adopts. The sign that the scope was right came afterwards. The agent was redeployed at a second producer in the same region, on the same structure.
This step-by-step approach is the one Tandem applies across its engagements, and it is covered in our overview of AI use cases in companies.
The technical choice, the lightest build that holds
The setup first considered rested on an assembly of bricks, a messaging gateway, an automation orchestrator and a database. It was dropped before the first line was written, in favour of a single conversational platform, far lighter to maintain at this scope.
- The channel. WhatsApp Business first, then the same agent placed as a widget on the website, which cost only a setting.
- The content. Four synced spreadsheet tabs, editable by the producer.
- The automations. Email dispatch and contact tracking, orchestrated in the background.
- The email copy. A language model formats the confirmation from the collected information.
The reflex to keep is to pick the simplest assembly that covers the need. At wider scope an orchestrator becomes justified, which is the point of our n8n agency offer. For higher-volume customer support, dedicated platforms take over, compared in our article on the best AI ticketing tools.
Does a conversational agent make sense for a small business
Yes, on three conditions you can check before starting. The questions received repeat, they arrive outside opening hours, and somebody can keep the product information up to date. Without the third, the agent will be giving wrong answers within a quarter.
Company size matters less than the repetition of the requests. A cellar, a practice or a neighbourhood shop meet the same pattern. When requests arrive by phone rather than in writing, the subject shifts to a voice agent, covered in our guide to AI voice agents. To work out which applies to your case, Tandem runs an AI audit that quantifies the gain before any build, and trains the teams through its AI training offer.
