The channel where customers actually write in Latin America. Built on the official WhatsApp Business API, with voice note transcription and escalation to your team.
3–5 weeks
Chatbots and assistants
An assistant that knows your company and admits when it does not
Most chatbots on the market are a model wired to an FAQ document. They work in the demo and fail in month three, when questions nobody anticipated start arriving.
Technical capabilities
WhatsApp Business APIWebHuman escalationSource citationVoice notesMultilingual
2M+users served by assistants we have operated
80%of messages are usually the same dozen questions
3–6weeks to implement depending on channel and integrations
Key takeaways
- An assistant that admits when it does not know and escalates to a person preserves the team's trust; one that improvises destroys it permanently.
- That is achieved with grounding constraints in the system prompt, mandatory citation of the source passage, and explicit refusal when retrieval comes back empty.
- Quarl builds on the official WhatsApp Business API, not on unofficial automations that get the number banned by Meta.
- If the case is fewer than twenty FAQs over static information, a subscription platform solves it more cheaply — and Quarl says so before quoting.
Operating track record
FEMSA loyalty platform
Three years operating the system in production — not one delivery and an exit.
Cross-entity contamination
The most frequent failure in RAG architectures, solved with structured metadata at ingestion.
Fixed scope and date
The proposal arrives within 48 hours with fixed scope, price and date. If scope changes, it is quoted separately and approved first.
The difference
Saying "I don’t know" is a feature, not a failure
When an assistant gets a question outside its information, it has two paths: invent something that sounds right, or say it does not know and hand the conversation to a person. By default, a language model picks the first.
An assistant that recognizes its limits and escalates with context already loaded is worth ten times one that improvises, because the first preserves the team’s trust and the second destroys it permanently. Achieving that is engineering: grounding constraints, mandatory citation and explicit refusal when search comes back empty.
Channels
Where the assistant lives
Website
Widget or full page, aware of what the user is looking at, with handover to a human agent without losing the thread.
3–4 weeks
Inside your product
An assistant embedded in your application, with the user’s session, permissions and data.
5–8 weeks
Internal for your team
Over documentation, procedures and internal systems. Usually the highest-return one, and the least often requested.
4–6 weeks
Always included
What is not optional
- Answers from your real information. Catalog, pricing, availability, hours and policies — not generic text written once.
- Live lookups when needed. Order status, stock, balance: connected to your system, not to a stale copy.
- Escalation with context. When it hands off to a person, that person sees everything already discussed and does not make the customer repeat.
- A report of what it could not answer. That list is gold: it tells you what is missing from your information and what the market is looking for.
- Quality measurement. A set of real questions with correct answers, so you know whether the assistant improves or degrades with each change.
Honesty
When you should not hire us
| Your situation | Recommendation |
|---|---|
| Fewer than twenty FAQs, information that barely changes | Do not hire us. A subscription platform solves that and costs a fraction. |
| You just want a menu of options with fixed answers | Also no. That does not need a language model or the cost it implies. |
| Large catalog, many variants, information that changes | Here it starts to make sense: exactly where generic platforms confuse similar entities. |
| A wrong answer costs a customer, a fine or a legal problem | You need measurement and explicit refusal. Not optional. |
| You already have one and it does not work | Start with a diagnostic before rebuilding. It is usually cheaper. |
Related services
RAG systems
Chunking, reranking and hybrid search, evaluated with recall@k and NDCG.
View serviceAI agents
LangGraph orchestration, durable state and human approval on steps with consequences.
View serviceAI automation
Applied where there is volume and stable rules, not where there is expectation.
View serviceProject rescue
The system is already in production and answers badly. We measure it and determine what to fix.
View serviceFrequently asked questions
How is this different from using ChatGPT?
ChatGPT knows everything except your company: not your prices, your inventory, your warranty policies or an order’s status. Asked about those it answers generically or invents. What we build searches your information first, answers only with what it found, and cites the source so you can verify it.
Can we use our existing WhatsApp number?
Yes, an existing number can be migrated to the WhatsApp Business API, though the process means it stops working in the regular WhatsApp app. That is why many companies prefer dedicating a new number to the assistant and keeping the current one for personal contact. It gets decided on the first call based on how your team works.
What does Meta charge for conversations?
Meta charges per conversation initiated, with different rates depending on who starts it and variation by country. For typical mid-market volume it usually lands in the low hundreds of dollars monthly. You pay that directly to Meta, not to us, and we estimate it against your real volume before you sign.
Why the official API and not a cheaper automation?
Because unofficial methods get your number banned by Meta, and losing the number your customers write to is a hard disaster to reverse. The official API has a per-conversation cost, but it is the only way to build on WhatsApp without risking the channel.
Does it understand voice notes?
Yes. They are transcribed through a speech-to-text service and processed like a written message. It is especially relevant in Latin American markets and is usually badly handled by generic platforms. There is a small per-minute transcription cost included in the monthly estimate.
What if the customer realizes it is a bot?
They will, and that is fine: the assistant introduces itself as an assistant from the first message. Trying to pass it off as human goes badly when discovered. What people dislike is not talking to a bot, but talking to one that solves nothing and will not let them reach a person. That is why escalation is a core function, not an extra.
Does it work in multiple languages?
Yes. Current models handle Spanish and English well, and the system can reply in the language it is written to. What needs care is that the knowledge base is in the right language: if your documentation is only in Spanish, English answers will be translations and should be validated before enabling them.
What does it cost to run per month?
Model consumption plus the vector store, roughly USD 150 to 900 monthly for a mid-market company, plus Meta’s conversation cost if it is WhatsApp. We do not charge a subscription on the system: it is yours, with its code and documentation.
Book 15 minutes
Tell us what you are building, or what stopped working. You leave the call with a concrete answer: it can be fixed, it can be built, or it isn't worth it.