The verdict in three sentences
A Berlin manufacturer receiving 2,500 customer requests a month, 60% of them repetitive, can hand part of that volume to an AI assistant connected to its company documentation (manuals, terms of sale, technical sheets, support history). In 2026, usage cost drops to 0.01 to 0.05 EUR per conversation, and a serious RAG project costs 15,000 to 50,000 EUR on the German market once AI Act and GDPR work is included. Reliability does not depend on the model but on document quality, guardrails and handover to a human agent: well designed, the assistant resolves 30 to 45% of requests unassisted.
What a RAG assistant does, and what it does not
RAG (retrieval augmented generation) makes the model answer only from your documents, with sources cited. The assistant knows nothing beyond what you feed it, which limits invented answers. It handles manual questions, compatibility, standard delivery times and warranty well. It must hand over as soon as a request involves a dispute, an out-of-policy refund or sensitive customer data.
| Request type | Share of volume (typical manufacturer) | Automatable with RAG |
|---|---|---|
| Manual and usage questions | 22% | yes, 70 to 85% |
| Order tracking and lead times | 18% | yes if connected to the ERP |
| Warranty and return terms | 12% | yes, 60 to 75% |
| Compatibility and spare parts | 8% | partly, 40 to 60% |
| Complaints and disputes | 15% | no, human handover |
| Sales enquiries and quotes | 25% | qualification only |
Out of 2,500 requests, 30 to 45% autonomous resolution means 750 to 1,125 requests handled without an agent. At 8 minutes of average handling time, that frees 100 to 150 hours a month, close to one full-time equivalent.
2026 costs: German market and Kolonell pricing
| Item | 2026 order of magnitude | Comment |
|---|---|---|
| API cost per conversation | 0.01 to 0.05 EUR | depends on model and conversation length |
| Monthly API cost for 2,500 conversations | 25 to 125 EUR | negligible next to salaries |
| Vector database and app hosting | 50 to 300 EUR a month | EU cloud recommended |
| RAG project, German agency, AI Act and GDPR included | 15,000 to 50,000 EUR | indexing, interface, integrations, compliance |
| Off-the-shelf SaaS chatbot | 300 to 1,500 EUR a month | less control over data |
| Kolonell, medium custom app | 4,000,000 to 7,000,000 FCFA (6,100 to 10,670 EUR) | RAG, back office, human handover |
| Kolonell maintenance | 35,000 to 120,000 FCFA a month | index updates, quality monitoring |
The most often forgotten item is document preparation: deduplicating manuals, removing outdated versions, structuring the internal FAQ. Plan 5 to 15 days of work on the client side, and this work is what separates 20% from 45% resolution.
Reliability: the guardrails that matter
A reliable assistant always shows its sources, declines to answer when confidence is low and offers handover to an agent with the conversation context. You also need a monitoring board: resolution rate, handover rate, negatively rated answers. On compliance, GDPR requires informing the user, limiting conversation retention and signing a data processing agreement with the model provider. The EU AI Act adds a transparency duty: users must know they are talking to an AI, an obligation that applies from August 2026 for this kind of system. German works councils may also need to be consulted if agent performance data is collected.
Mini case study
Illustrative example: Jonas, head of customer service at a professional kitchen equipment manufacturer in Berlin, runs a team of 6 agents for 2,500 requests a month.
Need a professional website?
Kolonell builds websites that attract clients, optimized for the Sénégalese market. Free quote in 2 minutes.
- Requests resolved by the assistant: 35% of 2,500, or 875 requests.
- Time freed: 875 x 8 minutes = 117 hours a month, at 38 EUR loaded per hour, or 4,450 EUR a month.
- Recurring costs: 75 EUR of API, 150 EUR of hosting, 115 EUR of maintenance, about 340 EUR a month.
- Investment: 25,000 EUR for a local project. Payback in about 6 months, and the team refocuses on disputes and key accounts.
FAQ
Can the assistant make up answers?
The risk exists but drops sharply with RAG, cited sources and a confidence threshold. On a clean corpus, the wrong answer rate usually falls below 3%, and every doubtful answer should trigger a human handover.
How long does deployment take?
Plan 6 to 10 weeks: 2 weeks of document preparation, 3 to 5 weeks of development and 2 weeks of testing with agents. A pilot on 20% of traffic is recommended before full launch.
Where is our data hosted?
We favour EU hosting and model providers that do not train on your data. The GDPR data processing agreement is part of the deliverables.
Should the assistant be connected to our ERP?
It is not mandatory, but order tracking often accounts for 15 to 20% of requests. A read-only connection usually adds 1 to 3 weeks to the project.
What resolution rate should we target in year one?
A realistic goal is 25 to 30% at launch, then 35 to 45% after 3 to 6 months of knowledge base enrichment. Beyond that, gains come mainly from better source documents.
Let's scope your project. Tell us your request volume, channels and the state of your documentation, and we will price the assistant, its integration and the pilot schedule. Detailed quote within 48 h. WhatsApp +221 77 596 93 33.
Mohamed Bah
Fondateur, Kolonell
Passionate about digital and entrepreneurship in Africa, Mohamed has been helping Sénégalese businesses with their digital transformation since 2020. Founder of Kolonell, he believes every SME deserves a professional and accessible online présence.
