Plain English, no buzzword bingo. Whenever an AI vendor throws a term at you and your gut says they are dressing up nothing, look it up here.
Optimising your content so it gets cited inside ChatGPT, Claude and Gemini answers.
AEO is the next layer above classical SEO. Instead of fighting for clicks on a search-result list, the goal is to be cited as a source inside a fully-formed AI answer. If you are not in the AI answer, you are not in a German mid-market buyer's research path. Prerequisites: clear definitions, concrete numbers, named sources, clean structured markup.
A method where the AI only answers from your own documents and ships the source with each answer.
With RAG, the system first retrieves the matching passages from your knowledge base (FAQs, contracts, manuals) for every query and ships them to the model alongside the question. The model formulates an answer strictly from those sources and returns the citation. Benefits: no hallucinations, no fine-tuning when documents change, clear accountability.
When the AI invents something plausible but false because it has no source and still wants to answer.
Generative models without source grounding tend to produce confident-sounding wrong answers. In a mid-market business this is a compliance and reputation risk. Counter-measures: source grounding via RAG, conservative escalation thresholds, sample reviews by a responsible human.
A contract under Article 28 GDPR that governs how an external service handles your personal data.
Once personal data flows through an external service, you need a DPA. Any AI vendor should hand it to you with the offer, not after the contract is signed. The DPA defines rights, retention periods, sub-processors and access controls.
EU regulation that categorises AI applications into four risk tiers and attaches obligations to each.
The AI Act has been in force since 2024 with phased applicability. For most mid-market use cases the risk tier is low, with transparency and documentation duties. High-risk applications (hiring decisions, credit scoring, medical diagnosis) require richer documentation and conformity assessment.
A structured risk analysis for processing operations with high risk to data subjects.
A DPIA is not mandatory for every AI use case but is good practice for any application that handles customer data. It documents which data is processed, which risks exist, and how they are mitigated. White Fox ships a DPIA template with every setup.
A large language model that understands and generates text based on its training data.
Common LLMs include GPT (OpenAI), Claude (Anthropic), Gemini (Google) and Mistral (France). For GDPR-clean setups in the mid-market we lean on EU vendors with DPAs, mainly Mistral from France or Anthropic with EU configuration.
Smallest billing unit of an LLM, roughly half a word or a word fragment.
LLMs are billed by tokens. A 20-word German sentence is around 30 to 40 tokens. To understand cost behaviour you should look at input and output tokens separately, since output is usually more expensive.
A systematic process to filter buyer-ready prospects out of incoming inquiries.
Sales typically spends a large share of time sorting unfit inquiries. AI can clarify need, budget and timeline through dialogue. What remains is a clean list with a clear lead score that sales can pick up.
The structured collection of your documents, manuals and FAQs that the AI uses as ground truth.
The knowledge base is the foundation of any Service Machine. The cleaner and fresher your documents, the better the AI answers. Before every setup we do a source audit and prioritise what should be reworked first.
Mathematical representation of your content that captures semantic similarity between texts.
Embeddings are the mechanism a RAG system uses to find the right document passages for a question. In our mid-market setups, embeddings live in an EU-hosted vector database, typically Qdrant or pgvector on a server we run for you.
The craft of describing a task to an AI clearly enough that it delivers reliably.
During setup we write system prompts that define how the AI should respond, its tone, when it escalates, and which statements it must never make. A prompt is more contract than magic.
An operating model where humans sample-check AI output and steer the system.
We use the term carefully because it is often misunderstood. At 800 service tickets per month no one can manually review every answer. The reality is sample reviews, clear escalation rules, and a responsible human who steers tone and quality.
An AI system that plans and executes multiple steps instead of producing a single answer.
An agent can take a lead from your form, research the company, compute a lead score and write it to your CRM. In the mid-market we deploy agents for clearly defined workflows, not open-ended tasks.
A contract model where you know the cost before signing, with no hourly billing.
The alternative is consultant days at hourly rates, which can quickly drift open. With fixed price the setup is one defined number, the monthly retainer another. You get planning certainty, we get the right incentive to optimise.
How often your company or URL gets cited in AI-generated answers.
From our live projects, realistic targets are eight to fourteen AI mentions per month plus two to four new top-ten Google rankings. The condition is substance: concrete numbers, real experience, clear definitions.
We reply within one business day. No sales pitch, no drip campaign. Florian or a teammate calls, listens, frames it.
A personal callback
You write here, someone from our Hamburg team calls you. No sales team, no call centre, no email funnel. We listen, frame the problem, and tell you honestly whether our solution fits your situation.
The White Fox team · Hamburg