AI consulting
that calculates, not sells hype.

The most common misjudgement in mid-market: treating AI as a universal tool that lifts efficiency everywhere.

High-value use cases usually sit in three places:

• Recurring knowledge work with unstructured input — offer drafting, contract analysis, supplier correspondence, support triage
• Data extraction from documents, emails, PDFs — invoice capture, delivery-note processing, master-data maintenance
• Research and synthesis tasks with provable sources — RAG over internal documents, competitor monitoring, market briefings

What is rarely economical:

• AI as a replacement for well-defined ETL pipelines
• AI for tasks with high hallucination risk and hard compliance
• AI as pure chatbot gimmickry without connection to real data

Our discovery format: two workshop half-days with the operational people — not with the steering committee. We walk through real daily routines (which emails arrive, which Excel sheets are opened daily, which reports are manually compiled) and assess each candidate along four dimensions: frequency, complexity, fault tolerance and data availability.

Output is a prioritised roadmap — typically 3–5 quick wins for the next 8 weeks plus 2–3 strategic cases for 6–12 months, each with rough effort and a business-case hypothesis.

Not every task is a prompt problem. We cleanly separate four architecture classes.

Four architecture classes:

• Pure LLM tasks (rewriting, translating, summarising) — fast and cheap, little engineering
• RAG (retrieval-augmented generation) — when the model should access internal documents, product data or knowledge base without hallucinating. Stack: vector database (pgvector, Qdrant), embedding model (OpenAI, BGE, Cohere), hybrid retrieval (BM25 + semantic), re-ranking, clear source attribution
• Agent setups with tool use — when the model should perform multi-step actions (query data, write emails, trigger ERP postings). Here tool definition, eval loops and guardrails matter
• Classical ML (forecasting, classification, anomaly detection) — when the task has structured inputs and reproducible outputs, a gradient-boosted tree is often better, faster and cheaper than an LLM

We choose the architecture by task, not by tool trend. Frequent market mistake: using agentic frameworks (LangChain, LangGraph, CrewAI) for tasks solvable with two API calls and an if-condition — the result is brittle systems with high token costs. We build as little complexity as necessary.

Frontier models (US cloud) — OpenAI GPT-5, Anthropic Claude Opus 4.7, Google Gemini 2.5 — we recommend when the task needs maximum reasoning depth, the data flow is legally workable (DPA, EU region if available, consent clarity) and the token budget fits.

EU / open-source models — Mistral Large / Codestral (French hosting), Llama 3.3 / Llama 4 on own infrastructure, Qwen2.5 for multilingual workloads — we recommend when data residency is hard-required, token volume economically argues for self-hosting (typically from a few million tokens/month), or when model stability over years matters more than frontier reasoning.

Stack components by use case:

• Embeddings and re-ranking: OpenAI text-embedding-3, BGE-M3, Cohere Rerank — depending on language and domain
• Vector storage: pgvector in Postgres (sufficient for most use cases, no extra system), Qdrant (when real hybrid search and large corpora are needed), or Elasticsearch (if you operate it already)
• Agent orchestration: native tool use of the models (OpenAI function calling, Anthropic tool use) as default, LangGraph only when complexity really justifies it

Hosting of our inference containers and vector DBs runs on Hetzner Frankfurt or Fly.io Frankfurt — data processing within the EU.

We cleanly implement the technical prerequisites for GDPR-compliant and EU-AI-Act-compatible AI operation — the legally binding assessment, however, explicitly belongs to your lawyer or DPO.

What we implement technically:

• EU hosting of applications and vector DBs (Hetzner, Fly.io, Frankfurt)
• Logging discipline — which inputs/outputs are stored for eval, which are not
• Data classification before the model call — PII masking, secrets filtering
• Data-processing agreements with all model providers
• Clear separation between training and inference data — no training on customer data without explicit opt-in

On the EU AI Act: since February 2025 the bans on "unacceptable" systems are in force, from August 2025 the duties for general-purpose AI models, from August 2026 the main duties for high-risk systems.

In mid-market the most common practical question is: which of our AI applications fall into high-risk categories?

• High-risk (typical): HR screening, credit scoring, biometric identification, education
• Minimal risk (typical): internal productivity tools, marketing copy, translation

We deliver a pragmatic classification of your use cases plus the technical building blocks for the duties (risk management, logging, transparency notices, human oversight). The final compliance stamp is set by your DPO or lawyer — we deliver the technical template.

An AI application only creates leverage when connected to the real data flows.

Existing systems we integrate LLM-powered workflows with:

• ERP / accounting: SAP Business One, Microsoft Dynamics, BMD, DATEV
• CRM: Salesforce, HubSpot, Pipedrive
• E-commerce: Shopify, WooCommerce
• Productivity / collaboration: M365 (Outlook, Teams, SharePoint), Google Workspace, Atlassian (Jira, Confluence)

Connected via REST, GraphQL, webhooks or native connectors.

Typical architecture pattern: a slim integration service (Node.js on Fly.io, Python on Hetzner) as mediator between existing system and model API, with idempotency, retry logic and clean error tracking via Sentry.

For sensitive workflows: an approval step with a human in the loop before the model writes into a system — we don’t sell you the "agent does everything autonomously" fairy tale.

Over the years we have built and operated numerous own shops and digital products — dropshipping models, our own warehouse with product development, our own brands with assortment build-up and marketing.

Since 2023 we have used AI productively in our own operations:

• Product copy generation with brand-voice consistency
• Supplier email triage
• Customer-support pre-classification
• Competitive briefings via RAG over public sources
• Internal SOP search

This experience — what actually holds up in a running mid-market operation and what fails on the friction of reality — flows into every engagement.

Our team combines three disciplines:

• Software and AI engineering
• E-commerce strategy and platform know-how
• Brand and marketing practice

You talk directly with the seniors who afterwards build the system, set up the eval pipeline and write the cost reports — not through three layers of account management that "sell" AI consulting and then hand it down to junior implementers.

And we actively say no if a use case isn’t mature (too little training data, too high hallucination risk, no clear success metric) — even when it costs us an engagement.