A company’s first steps into generative AI rarely fail because of technical issues. They fail because companies start with the wrong project. Those who focus on a specific pilot project with a defined metric, rather than creating a strategic presentation, will more quickly achieve results that pay off.
Three projects are ideal for getting started with generative AI because they run on the same technical foundation and deliver measurable results within a few weeks: a RAG-based document search using internal knowledge, support for quotes and requests for proposals, and a knowledge assistant for service and maintenance. The key is a Key figure before the start.
Why Generative AI Projects Get Stuck
According to McKinsey, two-thirds of companies have not yet begun scaling AI applications, and only 39 percent are achieving measurable added value (McKinsey, State of AI 2026). The figures from Germany are even more striking.
of German companies have systematically integrated AI into their processes, even though 97 percent consider it relevant. Between ambition and implementation there is usually a first project that is too large.
KPMG, Study: Generative AI in the German Economy, July 2026
Anyone who wants to automate their entire sales management process right away will need months to set up data access, obtain approvals, and run tests before users even see anything. Another problem is the lack of performance metrics. 35 percent of companies do not track the impact of their AI projects at all (FPT and Forrester, 07/2026). Without metrics, a pilot project can neither be extended nor halted—it simply fizzles out.
This applies to industry and retail in particular. Production planning, purchasing, and service rarely have staff to spare for months of polishing a strategy before the first benefit becomes visible.
| Project | Benefits | Effort | Requirement |
|---|---|---|---|
| RAG Document Search | Search time is reduced by 60 to 70 percent | 60,000 to 120,000 euros, payback period of 4 to 6 months | searchable documents |
| Quotation Support | Less tedious work for standard solicitations | Same basis, different sources | offers with similar structures |
| Knowledge Assistance Service | Responses Directly at the Scene | Expansion of the existing base | well-maintained incident logs |
Project 1: RAG Document Search Using Internal Knowledge
A retrieval-augmented generation search combines a language model with a company’s own documents, manuals, and files. Employees ask a question in natural language and receive an answer with a source reference, rather than having to click through twenty results in a full-text search. A prototype can be built in two to four weeks; the production rollout takes another four to eight weeks (Pexon Consulting, 03/2026). The investment ranges from 60,000 to 120,000 euros for installation, plus 500 to 2,000 euros in monthly operating costs. With a 60 to 70 percent reduction in search time, the project pays for itself in four to six months.
We describe in detail how such a search works from a technical standpoint and what its limitations are in the article RAG explained.
Project 2: Support for Proposals and Bids
Sales and project teams spend a large portion of their time searching through old proposals, compiling reference texts, and reviewing bid documents. A language model with access to past proposals, price lists, and service descriptions extracts relevant text blocks, flags contradictory wording, and suggests an initial outline. Humans remain responsible for the calculation and final approval, while the machine handles the legwork.
The benefits are most evident in standard requests for proposals that include recurring sets of questions. For highly customized large-scale projects, the amount of manual work remains high; in such cases, no one should expect full automation. A realistic starting point is a single product line or department whose proposals are structurally similar.
Project 3: Knowledge Assistance for Service and Maintenance
Service technicians work on the road, often without time for lengthy searches in spare parts catalogs or old fault logs. A knowledge assistant that consolidates machine manuals, maintenance histories, and internal fault reports answers questions right on the job site. Those who have already set up a RAG document search system for administrative purposes can extend that same infrastructure to the field service team, rather than starting from scratch.
What an entry-level project needs to ensure it doesn't remain in the pilot phase
36 percent of German companies now use AI in some form (Bitkom, 02/2026). The difference between a pilot project that fizzles out after three months and one that continues to operate productively rarely comes down to the choice of model. It comes down to clear responsibilities for approvals, a key metric established before launch, and a realistic assessment of which tasks a language model can reliably handle and which it cannot.
Anyone who wants to use generative AI productively clarifies these questions before the first line of code and saves themselves the second kickoff round. That includes the question of who in the organization signs off a new AI system and who spot-checks the results. These roles cannot be managed on the side; they need a named person. Which roles and approvals belong to this is set out in the article on AI Governance, and how to calculate the benefits, in the article on ROI of AI Projects.
Which of the three projects best fits your situation?Tell us what your biggest time-consuming task is. We'll let you know whether a pilot program is worth it—and, if so, based on which metric.
Frequently Asked Questions
How much does an initial pilot project for generative AI cost?
For an RAG document search, the investment ranges from 60,000 to 120,000 euros, plus 500 to 2,000 euros in monthly operating costs (Pexon Consulting, 03/2026). Quotation support and knowledge assistance fall within a similar range when they use the same technical foundation. Smaller prototypes for individual departments are also feasible with a fraction of this budget, but they then deliver correspondingly more limited results.
Does a starter project need its own servers?
No, that’s not necessary. Many companies start with a cloud connection because it allows them to test a prototype without having to make an upfront investment in hardware. When data protection requirements are high or the volume of requests is large, operating their own system often becomes more cost-effective later on; read more about this in the article on On-premises models.
What does the EU AI Act require of these projects?
A RAG document search generally falls into a limited risk category with a transparency requirement; therefore, it must make it clear that users are interacting with an AI system. High-risk requirements apply only to applications involving personnel decisions or safety-critical functions. The exact deadlines and obligations are outlined in the article on EU AI Act.
The Next Step
torck implements entry projects for generative AI itself, with its own development teams in Maxhütte-Haidhof, Vienna, and Rabat. After the rollout we stay responsible for operation and maintenance. For industry and retail, we clarify in the initial consultation which of the three projects fits your own starting position. Schedule an Initial Consultation.
This article refers to laws and regulations to put technical decisions in context. It is not legal advice. Whether and how a rule applies to your company is a question for your legal department or a law firm.
German 
English 
French