AI in Action

Our phone agent is designed to answer when we cannot: it greets the caller, understands the request, qualifies it and books an appointment on request — invitation included. It speaks with a cloned voice (Voice Clone), and that is precisely why the first sentence is no formality: „I sound like Kai, but I am an artificial intelligence." The disclosure required by Article 50 EU AI Act is not buried in our small print — it is the greeting. As things stand it is a prototype: the pipeline runs, but we will not switch on a phone number until governance and data protection are fully settled — exactly the order we recommend to our clients.

Built as our own real-time pipeline with response times under 800 milliseconds, bilingual in German and English, designed for end-to-end data processing inside the EU. It shares its knowledge base with our website chatbot: one source, two channels. And when there is no human at the other end but another AI agent, it switches via Gibberlink to a tone-based protocol — two machines that stop reciting pleasantries and simply exchange data.

It is allowed to do more than the chatbot: make contact and book appointments. What it is not allowed to do: commit to prices or interpret the law.

Most website bots want your email address first. Ours does not. It answers your questions on AI governance, EU AI Act and ISO 42001 — and when things get specific it points to an appointment instead of collecting contact details. Data minimisation here is not an afterthought but a design decision.

The chatbot is labelled as AI, listed in our internal AI register and has a defined change process — we treat our own bot exactly as we advise our clients to treat theirs. Its knowledge base is the same one the phone agent draws on: change a statement in one place and it changes in both channels. Guardrails keep it on topic — no price commitments, no case-by-case legal advice, no invented deadlines.

A chatbot that says what it is and does not ask for what it does not need.

Screening public procurement portals by hand every day costs sales time and still leaves gaps. For a client in fire safety we built an agent that watches newly published procedures on the major platforms, matches them against service profile, region, contract volume and contracting authority — and reports only the hits that genuinely matter commercially. The smoke-detector principle: don't watch constantly, get alerted when something burns.

Every hit arrives with direct links into the tender, procurement and organisational data, so the assessment starts where the research would otherwise begin. Hours of screening become a few minutes, relevant procedures enter the evaluation in structured form — and the pitch decision rests on clean data.

Not every tender you could win is one you should pitch for. TenderFit reads tender documents in a structured way — lots, budgets, eligibility criteria, risks — and scores the strategic fit through a weighted model of capability fit, references, capacity, commercial viability and risk.

The result is a reasoned go, maybe or no-go recommendation, including an estimate of whether the pitch effort pays off. For multi-lot procedures the model also simulates consortia with lead and partner logic across different trades. Developed for an agency — transferable to any sector where public tenders tie up a lot of preparatory work.

Not every task needs a language model. To forecast NO₂ concentrations we trained a classic machine learning model on historical environmental and measurement data: data preparation and feature engineering, gradient boosting, validation and performance tuning, with a persisted model for production use.

The result is a reliable forecast of air pollutant levels as a basis for data-driven decisions in environmental and mobility contexts. We choose the method to fit the question — not to fit the headlines.