Automate
We build website chat, WhatsApp bots and AI assistants connected to your CRM, shop and booking systems. For each type of question we decide whether a fixed flow, a language model or a person should answer, ground AI answers in your own content, and design the handover to your team.
A chatbot is worth building when many customers ask the same questions, or when every conversation has to collect the same information. Sensaria builds website chat, WhatsApp bots and AI assistants for companies in Switzerland and abroad and connects them to the CRM, shop, booking or helpdesk systems behind them. For each type of question we decide whether a fixed flow, a language model or a person should answer, and we design the handover so no conversation ends in a dead end. The finished bot answers from your content, writes clean data into your systems and knows when to stop.
Use a rule-based flow when the path is known and the result must be exact: booking an appointment, checking an order, collecting a lead’s details, anything that writes to a system. Use a large language model (LLM) when questions are open and phrased in many ways but the answer already exists in your content: product details, policies, how-to questions. Most useful bots combine both: the model understands and routes, and fixed flows do the transactions.
| Task | Rule-based flow | LLM assistant |
|---|---|---|
| Order status | Lookup by order number and email through the shop API | Not needed |
| Appointment booking | Fixed steps against the calendar API | Turns “next Tuesday afternoon” into a slot to confirm |
| Lead qualification | Fixed questions mapped to CRM fields | Summarizes free-text needs for the sales rep |
| Product and policy questions | Works for a short FAQ | Answers from approved content, with the source |
| Complaints, legal or medical questions | Hand over | Hand over |
Give it only your approved content to answer from, and check its output before the user sees it. With each question the assistant receives the relevant passages from sources you control (help articles, product data, terms), is instructed to answer only from them, shows the source and says it doesn’t know when nothing matches. Prices, stock and order data come from live API calls, never from the model’s memory.
Around that we add the controls we use in the LLM features we have built: validation of generated text (no links outside your domains, no figures that aren’t in the source), a fixed fallback answer, a log of each answer with model and prompt version, and daily cost ceilings. Before launch, the bot is tested against a set of real customer questions with expected answers. The set is rerun after every content or model change.
When the user asks for a person, when the bot has failed to understand twice, when the topic is sensitive (complaints, cancellations, anything legal or medical), or when the customer is marked as a key account. The transcript and every field collected go with the handover into the CRM or helpdesk, so nobody asks the customer the same questions again. Outside office hours the bot says when someone will reply and opens a ticket.
The bot never pretends to be human. The Swiss data protection commissioner states that under current law users must be told when they are communicating with a machine (FDPIC on AI and data protection), and Article 50 of the EU AI Act has required the same for users in the EU since 2 August 2026.
WhatsApp bots run on Meta’s WhatsApp Business Platform, through the Cloud API or a Business Solution Provider. You may only message users who opted in. Free-form replies are possible for 24 hours after the customer’s last message; after that, only pre-approved templates. Since 15 January 2026, Meta’s Business Solution Terms exclude general-purpose AI assistants, while bots that serve a specific business purpose, such as support, bookings or orders, remain allowed.
We have built WhatsApp messaging through a provider API for an e-commerce app: template messages, chatbot triggers, cart recovery and website widgets. PingMyUsers, the communication-API directory we built, compares WhatsApp Business API providers on sourced, dated facts.
A bot that can’t reach your systems can only talk. We connect it through a small tool layer with allow-listed operations: read an order status, create a lead with its transcript in the CRM, book a slot, open a ticket. Access is read-only wherever possible: the MCP server we built for PingMyUsers, which AI agents query directly, exposes only read-only tools. For lead flows, see CRM automation.
Conversation logs are personal data under the revised FADP and the GDPR. We set a retention period for transcripts, send the LLM provider only what a question needs, and choose a provider and configuration that match your data protection requirements.
| Driver | Why it matters |
|---|---|
| Channels | Website, WhatsApp or both; WhatsApp needs a business account, a number and approved templates |
| Flows and intents | Each transaction flow is designed and tested separately |
| Content preparation | Answers are only as good as the approved sources behind them |
| Integrations | CRM, shop, calendar or helpdesk connections |
| Languages | Content and tests for each language you support |
| Running costs | LLM usage grows with conversations; WhatsApp templates are charged per message |
After a short discovery we send a fixed-price or phased proposal, with an estimate of monthly running costs.
Leads from the bot enter your CRM and can start marketing automation journeys. System access runs through API integrations. For email, SMS, WhatsApp and chat in one flow, see customer communication. Our AI and automation page covers other uses of language models in a business.
We have put language models into production features with the controls described above: validated output with a fallback, a stored record of model and prompt version per answer, and cost ceilings. We have also built WhatsApp messaging through a provider API and MCP servers for AI agents. Sensaria AG is a Swiss company, and we design and test bots in English, Italian, German and French.
What's included
A chat widget on your site in your languages, with clear bot disclosure and a visible way to reach a person.
Bots on the WhatsApp Business Platform with opt-in handling, template messages for anything outside the 24-hour window, and order or booking flows.
Answers drawn from approved sources such as help articles, product data and terms, with the source shown and a fallback when nothing matches.
Structured questions mapped to CRM fields, a summary of free-text needs for sales, and meeting booking in the same conversation.
Order status, returns, appointment changes and account questions handled through your systems' APIs.
Multi-step flows with branching, data checks and confirmations, where the language model understands the request and fixed logic carries it out.
Handover rules, the full transcript and collected fields passed to the CRM or helpdesk, and honest messages outside office hours.
A small tool layer with allow-listed operations: read order status, create a lead, book a slot, open a ticket.
A test set of real questions rerun after every change, a review of unanswered questions, and alerts on cost and error rates.
Typical scenarios
Recurring questions about delivery, returns, opening hours or product use answered around the clock, with complex cases passed to the team.
Visitors describe what they need; the bot asks the qualifying questions, creates the lead in the CRM and offers a meeting slot.
Order confirmations and delivery notices as approved templates, and answers to follow-up questions within the service window.
Booking, moving and canceling appointments against the real calendar, with a confirmation and a reminder.
Staff ask about policies, product specifications or procedures and get answers with a link to the source document.
Related projects
FAQ
It starts with the questions and tasks the bot should handle, taken from real support emails and chat logs. We then design flows and handover rules, prepare the approved content the AI may answer from, connect the CRM, shop or booking APIs, and set up the channel: website widget, WhatsApp or both. Before launch the bot is tested against real questions; after launch, unanswered questions and costs are monitored.
When questions are open and phrased in many different ways, and the answers exist in your documentation: product details, policies, troubleshooting. Rule-based flows are better when the steps are fixed and the result must be exact, such as bookings, order lookups or lead forms. Most bots in production combine both: the language model understands the request, and fixed flows carry out anything that changes data.
The bot answers only from approved sources passed to it with each question, shows them, and says it doesn't know when nothing matches. Live data such as prices or order status comes from APIs, not from the model. Generated text is validated before it is shown, a fixed fallback replaces anything that fails, and a test set of real questions is rerun after every change. Sensitive topics go to a person.
Only to customers who opted in to messages from you on WhatsApp. Within 24 hours of the customer's last message, the bot can reply freely. Outside that window, and for anything you start yourself, such as reminders or order updates, WhatsApp allows only templates approved by Meta. Promotional messages are also subject to the consent rules for advertising in the recipient's country, in Switzerland the Unfair Competition Act.
Yes. The FDPIC takes the view that under the Federal Act on Data Protection users must be told when they are communicating with a machine, and for users in the EU the AI Act has required this disclosure since 2 August 2026. A clear label in the chat window and in the bot's first message covers it. The bot should never pose as a person.
Yes. The bot asks the qualifying questions you define, maps the answers to CRM fields, and creates or updates the contact and deal with the full transcript and the page or campaign the visitor came from. Existing contacts are matched on email address or phone number. From there, the CRM's own assignment and follow-up rules take over, as for any other lead.
Running costs come from three places: language-model usage, which grows with the number and length of conversations; channel fees, such as Meta's per-message charges for WhatsApp templates; and hosting and monitoring. We set daily cost ceilings so usage can't run away, and the proposal includes an estimate of monthly running costs at your expected volume.
Technology we use
TechnologyWebsite chat, WhatsApp bots and AI assistants that answer from your own content, qualify leads into the CRM, handle routine requests through your systems and hand over to a person when they should.