Three out of four mid-market companies we spoke with in 2025 had started an AI project. Three out of four had shelved it six months later. The reason was rarely the tech — it was the missing pattern.
Pattern 1 — RAG against internal documents
The idea: a language model answers questions based on internal sources (handbooks, contracts, wikis). The model itself knows nothing about the company — it gets the facts at runtime from a vector database.
Works for: service knowledge search, onboarding help, contract analysis. Fails when: documents are messy. RAG amplifies bad data quality.
Pattern 2 — Classifying workflows
Tickets, incoming invoices, supplier requests — much of it is sorting work. A language model can do this reliably if it only assigns labels and a human decides in edge cases.
Works for: service triage, invoice recognition, standard replies. Fails when: the workflow lacks a human check step.
Pattern 3 — Prediction models
Classic machine learning on your own business data — inventory forecast, maintenance intervals, sales projections. No LLM needed, often not even a GPU.
Works for: when historical data is clean and sufficient. Fails when: the team is not a data team. Then the model produces nice numbers nobody acts on.
What we no longer recommend
AI chatbots on websites (“talk to our virtual assistant”). Conversion is poor, maintenance is high, perception is mixed. A well-maintained FAQ beats almost any chatbot.
How we set up an AI project today
Four weeks of discovery, one clearly defined pattern, a pilot with real users, then scale. Start without a pattern, and you build theatre.
Sources
- McKinsey — State of AI 2025 — www.mckinsey.com
- BSI — Secure AI Adoption — www.bsi.bund.de