# Technology – How the AI Phone Assistant Works > Source: https://hi-desk.de/en/technology/ --- Technology # The architecture for autonomous action: Precise answers, grounded in your knowledge base. HiDesk understands context in real-time by matching customer intent with centralized company knowledge. In result, no random answers, but controlled, goal-oriented conversations. [Explore Technology](https://cal.com/florestan.p/hidesk-beratung)[Book Demo](https://cal.com/florestan.p/partner) ![HiDesk Logo](https://hi-desk.de/images/brand/icon_brand_gradient.svg) Context Awareness Knowledge Retrieval Autonomous Decisions Post-Processing Workflows Multi-Agent API Call ## Knowledge from all sources and systems Calls are not isolated events. Our system is designed to bundle information from various internal sources (such as FAQs or manuals) and your active business systems (CRM, calendar, ticket systems). This seamless integration enables HiDesk to generate context-accurate answers while simultaneously executing actions in processes. Fully Structured Knowledge ∞ Relationships Recognized Real-time Updates ## Dynamic Knowledge Base and Agentic Reasoning We bundle all your company knowledge, from conversation rules to customer data to internal processes. This allows the AI to understand the context of every inquiry in real-time and independently derive the right action decisions – such as the correct priority, necessary follow-up steps, or the responsible department. minimal Hallucinations Fact-grounded Thanks to a structured knowledge base Instant Context Retrieval ## Modular Architecture: Always the best AI performance We rely on a self-developed architecture. Our Agentic Core is developed completely model-agnostic, which means we are not tied to a single AI provider. Thanks to this plug-in system architecture, we can swap out the underlying models for the best available technology at any time. For you, this means: consistently high performance, without changing the logic of your trained processes or your invested knowledge. And where a general-purpose model isn't enough, we train our own specialized models – see below. That's how [AI automation without an IT department](https://hi-desk.de/en/blog/ki-automatisierung-ohne-it/) succeeds. 10+ Model Providers Fully Audit Trail GDPR Compliant ## We train our own models Model-agnostic doesn't mean generic. For the parts of a conversation that demand precision, we develop and train our own models – so HiDesk isn't limited to what general-purpose models can do. Building the core technology ourselves lets us tune it to the way phone conversations actually work, for a better, more capable assistant. Own Models we develop ourselves Purpose-built Tuned for phone conversations More precise Technology, end to end ## How the AI phone assistant works Six steps sit between the first ring and the finished call summary. Each one runs in real time – callers never notice that several systems are working together behind the scenes. Here is the chain in detail: 1. 1 ### Call pickup on the telephony layer The call reaches HiDesk through an ordinary call forward – you keep your phone number and your provider. You set up the forward in your phone system or directly on your router. On our side, the telephony infrastructure accepts the call via SIP and opens an audio stream to the assistant. 2. 2 ### Real-time speech recognition Spoken words are converted to text continuously, not just at the end of a sentence. That lets the assistant react while someone is still talking – and it copes when callers interrupt or restructure a sentence halfway through. This works in multiple languages, not just German or English. 3. 3 ### Understanding intent and retrieving knowledge From the transcript, the Agentic Core works out what the call is actually about: a booking request, a question about opening hours, an incident report. In parallel, it matches the request against your knowledge base – the FAQs, prices and conversation rules you have stored – and pulls live data from connected systems such as your calendar or CRM where needed. If you serve several clients or locations, HiDesk recognizes from the dialed number who it is speaking for and uses the matching knowledge base. 4. 4 ### Answers from verified knowledge only The answer is generated exclusively from these stored sources. If the assistant finds no reliable information, it does not invent any: it says openly that it cannot answer the question, takes a message, or hands over to a human. That is the difference between a controlled knowledge base and a language model improvising freely. 5. 5 ### Natural speech output The finished response text is converted to speech and streamed straight into the live call. You choose the voice and the greeting yourself, so the assistant sounds like your company. 6. 6 ### Post-call processing Once the call ends, post-processing takes over: you receive a summary by email, the full transcript is available in your dashboard, and depending on your configuration HiDesk creates the record in your CRM, opens a ticket, or puts the agreed appointment in the calendar. The knowledge base itself comes together without an IT project: during onboarding you enter your website, add your call scenarios and connect your calendar or CRM – HiDesk takes care of the rest. Everything is maintained in one place, the dashboard: when opening hours, prices or responsibilities change, you update the entry and the assistant answers with the new information from the very next call. There are no copies that can drift apart – every answer rests on the same, current foundation. And what if the assistant gets stuck? For that, you define escalation rules: it can forward the call to a number you set, announce a callback, or take the request as a message – including every detail already clarified during the conversation. All of this happens on every single call, at three in the morning just as reliably as on a busy Monday at nine. And because every step is logged, you can always trace what the assistant understood, answered and triggered. ## Security and privacy Phone conversations are among the most sensitive data a company handles. The foundational decisions behind HiDesk are deliberately conservative – privacy is not an add-on here, it is part of the architecture: ### Hosted in Germany HiDesk is hosted in Germany. AI processing runs on Microsoft Azure OpenAI in data centers within the EU – your conversation data is never used to train generally available AI models. ### Encryption throughout All data is TLS-encrypted in transit and stored encrypted at rest – from the audio stream to the transcript. ### GDPR in everyday calls In practice this means: callers are told at the start of the call that they are speaking with an AI. Recordings are optional and require consent. Only what is actually needed for the request gets stored. ### DPA under Art. 28 GDPR For your use of HiDesk we provide a data processing agreement (DPA) – [just get in touch](https://hi-desk.de/en/#contact). Our own service providers, for instance for the phone connection, are contractually bound under Art. 28 GDPR as well; you can see which ones in our [privacy policy](https://hi-desk.de/en/privacy/). ### Retention periods you control You define retention and deletion periods yourself, for instance automatic deletion of recordings after a set number of days. ### Audit trail for every conversation Every interaction is documented and traceable: when the call came in, what was said, and which action the assistant triggered as a result. We also keep the EU AI Act in view. For systems like HiDesk, it primarily demands transparency: people must be able to tell they are talking to an AI. The assistant already meets that requirement today with a clear announcement at the start of the call – and we track the regulation's further obligations on an ongoing basis. Documentation and the audit trail also help you meet your own accountability duties towards customers and supervisory authorities. ## Integrations and an open platform A phone assistant is only as useful as the systems it connects to. HiDesk ships with native integrations for calendars (Google Calendar, Outlook, Calendly, Cal.com), CRMs such as HubSpot or Salesforce, and ticketing systems. Appointments, contacts and cases land exactly where your team already works – no retyping, no copy-paste between tools. For everything else there are Zapier, Make and n8n: through them, you can connect HiDesk to 250+ other tools, from industry software to internal databases. And when a system plays a special role, you can connect it directly through our open API. You'll find an overview of the most important integrations [on the homepage](https://hi-desk.de/en/#integrations). Technically, this works through actions the assistant triggers during or after the call: creating a calendar entry, updating a record, opening a ticket, sending a notification. You decide per scenario which actions are allowed – the assistant only executes what you have approved. ## Why model-agnostic? The AI world moves fast – today's best language model can be outclassed within six months. That is why HiDesk is not tied to any single provider. Our Agentic Core, the control layer we built ourselves that runs conversations, retrieves knowledge and triggers actions, treats the language model as a replaceable part. More than ten model providers are connected today – which model answers in a given case depends on the task. For you, that means we pick the model that best balances performance and cost for each task, and switch when needed – without anything changing in the workflows you have set up. Your stored knowledge, your conversation rules and your integrations stay untouched; only the engine underneath gets swapped. Model-agnostic does not mean we only buy in, though: for the parts of a conversation that demand particular precision, we train our own specialized models – more on that further up. The two together, freely selectable base models plus our own components, are why we talk about an architecture rather than a single model. ## The Result ### Natural-sounding, verifiable conversations Every interaction can be documented and is based exclusively on your central knowledge base. For the full picture, read [the complete AI phone assistant guide](https://hi-desk.de/en/blog/virtual-phone-assistant-guide/). ### Controllable dialog logic You maintain control over conversation flows, content, and rules. ### Secure data handling – fully in Germany Highest security standards and data protection according to German law. [Start with HiDesk Now](https://cal.com/florestan.p/hidesk-beratung)