The modules describe what Ada does for your customers. This page describes where her behaviour comes from: five steps repeated every day, and mechanisms you can inspect instead of taking on trust.
Data processed in the European Union · a human approves every data change · records without raw conversation content
Each step is a concrete mechanism in the product, not a stage on a slide. In the order they run.
Name, language, tone, voice, avatar, greeting, conversation starters, the catalogue of action buttons and the list of domains she may run on. Every customer has their own Ada, and the boundary between customers is enforced on each individual database query, not in the interface layer.
Articles with a draft, review, publish and archive cycle. Uploads of PDF, DOCX, TXT and Markdown. Import from a URL and a full site scan that returns for changes on its own. Ada sees published versions only, so a draft never answers a customer.
The coverage gate rejects an answer with no grounding in published content, before the customer sees a single word. Personal data in a question stops the turn before the model is called. An operation that changes data becomes a card for a human to click.
She answers in chat, in voice and through the avatar, keeps an e-mail thread going, books and runs demos, surfaces material from the catalogue, and hands the case to a human with the full context when the question goes beyond her scope.
The gap register collects unanswered questions and turns them into an editorial queue. Quality supervision scores answers beside the critical path. The audit log records every turn in append only form.
With an agent talking to your customers, the question that matters is why it said what it said. The answer sits in three places.
No claims we cannot back. Below are only things written in code and in documents you can read.
| Area | How it works |
|---|---|
| Separation between customers | The tenant scope is part of every database query. Somebody else's resource answers exactly like a nonexistent one, so the response never confirms that it exists. |
| Where processing happens | Database and files in Warsaw. We publish the list of parties processing data on our behalf, with locations and the basis for transfer. |
| Sensitive data | Document, account and card numbers are masked before anything reaches the log. That is pseudonymisation, and we call it that. |
| Data changes | The agent does not perform them. It issues a card that a human approves with a separate click, leaving a trace in the log. |
| Documents | The data processing agreement template, the subprocessor register and the AI systems notice are public, before any sales conversation. |
| Retention | Set for your account and enforced by a daily job, separately for each category of data. |
Technical detail and the full list of documents live on the security page, and the processing parties are listed in the subprocessor register.
In practice the cycle above begins with one afternoon of work on your side.
A website address to scan and a few files you already have: terms, pricing, frequent questions, product descriptions. The rest arrives as the gap register shows what is missing.
Name, language, tone, greeting, action buttons and the domain list. It usually takes one meeting, because the decisions are simple and the defaults work straight away.
The widget script into your page template. No site rebuild, no plugin and no changes to your code. From that moment Ada talks and books meetings.
These are the layers you ask about most often on larger rollouts.
Passing enquiries, bookings and conversation history into the system your team works in.
Notify meReading rep availability, so slots come straight from their own calendars.
Notify meGoals, guardrails and check scenarios in one place, together with a score of how well Ada meets them.
Notify meThirty minutes: we load a slice of your materials, ask the hard questions and show what remains in the log after the conversation.