04 Every message enters a single circuit and leaves either answered or in your team's hands.
We install your company's first line of customer service in your own environment, on the channels and the software you already use. Its behaviour is specified in writing: which intents it answers, from which sources, above which confidence level and to whom it hands over the rest.
Four channels, one circuit
Each channel has its own connector, and all of them deliver the message in a common format: sender, channel, time, text and attachments. From there on, the handling is identical.
-
01 WhatsApp
Through the WhatsApp Business Platform, always with your company's account and number and as an auxiliary function of your team, as Meta's terms require. The connector respects the 24-hour customer service window, falls back on approved templates outside it and stays within the platform's messaging limits. Voice notes are transcribed and follow the same path as a text.
-
02 Email
Connection to the company mailbox over IMAP and SMTP. Each email is handled within its thread and the proposed reply is left as a draft in the mailbox itself. Attachments that are orders, invoices or delivery notes go to the document circuit.
-
03 Web
Your website's chat and your contact forms reach the same connector by webhook.
-
04 Telephone
On the lines where your company takes inbound calls, a real-time voice conversation answers and transfers the call to a person when needed. Voicemail messages are transcribed and come in as a written message.
The path of every message
The models act in bounded steps: filtering, classifying, drafting and checking. Storing, querying, applying the threshold and recording are done by deterministic code, which is verified step by step.
-
01 Stored exactly as it arrives
The message is stored raw in your environment before any model reads it. If a provider fails mid-process, it is reprocessed from that original and no message is lost.
-
02 Input filter
Llama Prompt Guard, an open classifier that runs inside your environment, detects instructions hidden in a message or an attachment before any other model reads it.
-
03 Intent
Each query is classified against a catalogue of intents specific to your company: availability, order status, terms, incident or complaint, for example. A message that mixes several queries is given one intent for each.
-
04 A grounded reply
It is drafted solely from your knowledge base and from what the queries to your software return, as an object validated against a JSON Schema with two mandatory fields: the confidence level and whether it should go to a person. Handovers caused by missing information flag what needs adding to the base during monthly maintenance.
-
05 Output filter and decision
Granite Guardian checks that every statement is supported by the sources and that the text complies with your rules. If the query is informative, the check passes and the confidence exceeds the threshold, the reply is sent; in any other case, it goes to your team.
What the system looks up and what your team decides
The boundary lies in the architecture, not in an instruction to the model: the model requests a query, our layer runs it, and each conversation has a cap on steps and on spend.
Read-only queries
Free slots in the calendar, including when there are several schedules by person, room or site, and the status of an order, a delivery or an incident in the ERP or the CRM. If the software is legacy, the query goes through an MCP server that exposes read access only.
Your team's queue
Confirming an appointment, accepting a change to an order, handling a complaint or setting a special price arrives in a queue with the conversation, the intent, the data looked up, the proposed reply and the reason for the handover. Whoever picks it up replies without having to ask again.
Confirmations under an explicit rule
If you want a type of operation to be confirmed without waiting for anyone, such as an appointment in a specific time slot, it is defined as a bounded rule approved in writing. Deterministic code executes it; the model only proposes it.
The customer's file in your CRM
Each conversation is added to the customer's file as a new entry, written by deterministic code with a schema common to all channels: channel, transcript, intent, reply, who handled it and how it ended.
Transparency
The system introduces itself as automatic at the start of every conversation, in writing or by voice, and whoever writes or calls can ask for a person at any time.
Models, voice and languages
In customer service, each step has its own model and its own latency budget: a WhatsApp reply can take a few seconds; a phone conversation cannot. Before any change of model or instructions, this circuit's test suite runs, with cases taken from your real conversations.
- Classification and drafting: today, a model from the Mistral family served in the European Union for the daily volume and Claude for complex queries. If conversations must stay in your environment, the same circuit runs on open models installed on your server, such as Granite, ALIA/Salamandra or EuroLLM.
- Voice: Whisper, self-hosted, transcribes calls, voicemail messages and voice notes.
- Languages: the system replies in the language of each message, within those your company works in. ALIA/Salamandra, trained with particular attention to Spanish, Catalan, Basque and Galician, is the self-hosted option for those languages.
Results and how they are measured
Each indicator is measured with the same method as the diagnostic's baseline and appears in the monthly report.
- Time to first response: from the time of the message to the time of the reply, by channel and intent.
- Queries resolved at first line: the proportion of informative queries closed with the company's information.
- Handovers and their reason: how many conversations go to your team and why.
- Your team's time per handover: how long a person takes to close what they receive, with the context already prepared.
- Queue status: its size and the age of the oldest item.
- Complete record: conversations received versus conversations recorded in the CRM, by channel.
- Cost per resolved conversation: with an alert when it deviates from the forecast.
Shall we bring your company up to date?
We start with a 30-minute call to understand your case and explain what the diagnostic consists of and what it costs.
Request the diagnostic