Service · Wolff Applied AI
From a validated use case to a working AI knowledge bot with sources, guardrails and labeling — as an honest proof of value.
The MVP takes your already validated use case to its first real implementation with a software result in 4–8 weeks. We sharpen the scope and measurable success criteria, connect a defined set of documents (typically via a RAG pipeline) and build a working agent, chat or knowledge bot with guardrails, robust source references and AI labeling (aligned with Art. 50 EU AI Act). We then test with real user questions and deliver an honest test and quality report as well as a recommendation for productive deployment. Deliberately set up as a proof of value — functional and meaningful, but not yet the finished production operation.
AI initiatives rarely fail because of the idea, but at the transition from pilot to production — according to S&P Global (2025), 42 % of companies abandon the majority of their AI initiatives before productive deployment, and market observers expect a substantial share of GenAI pilots to be abandoned again by the end of 2026. In most cases, feasibility and value are speculated about without a real, tested prototype. At the same time, knowledge-intensive organizations lose a lot of time searching through scattered documents; generic chatbots without a connection to their own document set do not deliver robust, source-based answers. Freely deployed chatbots create a real hallucination risk without guardrails and source references. With the staggered transparency obligations of the EU AI Act (Art. 50, planned from 02 August 2026), information systems also need labeling from the outset.
4 phases, 3 gates — you decide at every transition.
In the kickoff with the department and IT, we sharpen the target picture, user groups and typical questions, and delimit the MVP scope. We define measurable success/acceptance criteria, set the document set and target environment, and make the architecture decision (RAG as default, API integration if applicable).
We review and prepare the defined document set (formats, structure, access rights) and build the pipeline consisting of chunking, embeddings, vector index and retrieval logic. An initial retrieval quality check on sample questions fine-tunes the parameters.
We build the working bot on the RAG pipeline with a source reference for each answer. Guardrails limit the scope, provide a fallback ("no substantiated answer") instead of hallucination, and offer basic protection against prompt injection; AI labeling and transparency notices (aligned with Art. 50 EU AI Act) make the AI interaction visible.
We test in a structured way with real user questions against the success criteria and record hits, gaps and error types. The honest test and quality report (incl. limits and hallucination risks), the data protection/EU AI Act classification, as well as the recommendation and effort estimate for productive deployment are handed over.
Sharpened MVP scope with user groups, question catalog, architecture decision and measurable acceptance criteria as a shared basis for assessment.
Functional agent / chat / knowledge bot that answers questions with source support based on the defined document set — as a proof of value, not a production system.
Prepared document set and traceable pipeline (chunking, embeddings, vector index, retrieval) with documented parameters and assumptions.
Implemented guardrails, source references per answer and AI labeling/transparency notices — aligned with Art. 50 EU AI Act, not legal advice.
Honest evaluation of the test with real user questions: answer quality against the success criteria, strengths, limits, typical errors and hallucination risks with examples.
Structured technical-organizational classification (data flows, data types, hosting/provider, transparency/labeling obligations) as orientation — explicitly not legal advice.
Clear go/no-go recommendation with the necessary steps, risks and a rough effort and cost estimate for the path into productive operation.
The base package stands on its own. Optional modules dock on now or later — you decide how far you go.
Beyond the defined set, we connect a second document or knowledge source to the RAG pipeline so that the MVP delivers more broadly supported answers. Extends the data scope within the MVP frame; a full productive build-out remains separate from this.
A further structured test round with real user questions sharpens answer quality and increases the informative value of the quality report. Complements the evaluation included in the base package, but does not replace it.
The controlled path from MVP to productive operation — with success criteria, gates, change management and operational handover. Represented as a separate next step via the AI Pilot Support (offering 10), not part of the MVP.
Credit: A preceding AI Potential Analysis or an AI Use Case Lab can be credited proportionally when commissioned directly.
Note: not legal advice. Content with a legal dimension (e.g. EU AI Act, GDPR, ISO/IEC 42001) is professional orientation, not legal counsel.
You receive a clear, individual proposal before any engagement. Under § 19 UStG no VAT is charged; any ancillary costs (e.g. model/hosting/API fees, travel) are not included.
Next step
Let us clarify in a free intro call whether this offer fits your situation — honestly and without sales pressure.As a follow-up we then recommend AI Pilot Guidance.
Non-binding · online or on site · reply < 24 h
