An enterprise RAG is an AI assistant that first searches your company's documents, then writes its answer from what it found, indicating its sources. For an SME, it is useful when the answer already exists in your documents and your teams lose time finding it again. By our calculation, building one ranges from about €2,400 to €60,000 excl. VAT depending on scope.
RAG stands for Retrieval-Augmented Generation. The term is intimidating, the idea is simple, and this guide is written for a manager who has to decide, without any code.
How does a RAG work, without jargon?
A RAG works like a fast archivist: it fetches the right passages, brings them to the writer, and the writer answers from those pages alone. The mechanism comes down to 4 steps:
- Prepare. Your documents are cut into passages and indexed by meaning, so that a question in everyday language finds the right paragraph even without the same words.
- Search. A colleague asks a question: the system selects the most relevant passages, among those they are allowed to see.
- Write. The AI receives the question and those passages, then writes an answer.
- Cite. The answer points back to the original documents, so that it can be checked.
The difference from using ChatGPT or Claude on their own: the assistant does not know your prices, your contracts or your procedures unless you give them to it. And the RAG does not train a new model on your data: it passes the right extracts to the model at the moment of the question.
What are the right use cases for a RAG?
The right use cases for a RAG share 3 characteristics: a large volume of documents, questions that keep coming back, and an answer that can be verified in a source. Here are 6 examples, to assess against your own volumes.
| Function | Typical question | Source queried | Who stays in control |
|---|---|---|---|
| Customer support | "What is the procedure for this error code?" | Help base, resolved tickets | The adviser reviews |
| Sales | "Have we already answered this clause?" | Previous proposals | The salesperson adapts |
| Legal | "Which contracts have a tacit renewal clause?" | Signed contracts | The lawyer checks |
| HR | "How many days for this leave under our agreement?" | Agreements and internal rules | The HR person confirms |
| Engineering consultancy | "Which rule applies to this type of structure?" | Standards, past studies | The engineer decides |
| Management | "What did we decide on this subject in June?" | Meeting minutes | The managing director |
On Jarvis, one of our development projects in landscaping and civil works, each business agent relies on a dedicated knowledge base. This principle holds for many document-heavy professions. A RAG is less suited to an exact calculation on figures: a query on your ERP will be safer. It is also less suited to a small base of 20 documents, which a standard assistant with a shared project already handles.
What is the link between RAG and the second brain?
The company second brain is the material, and the RAG is one way of searching it. The second brain gathers decisions, procedures, customers and meeting minutes in dated records, each with a status and an owner, which your teams and your agents consult.
The consequence for your budget is direct: plugging a RAG into a messy shared folder is like hiring an archivist and putting them in an attic. They will find things quickly, but sometimes it will be the 2019 version. So start with 3 clean collections (decisions, procedures, customers) and the rule "one single source per subject". If your knowledge fits in structured records, a simple connection via the MCP protocol may be enough at first, without a sophisticated search pipeline.
How much does an enterprise RAG cost?
An enterprise RAG costs anything from the price of a few seats in an existing assistant to several tens of thousands of euros if it has to be integrated with your software, with fine-grained permissions. The ranges below are a calculation based on a rate of €800 to €1,500 excl. VAT per day, taken from our pricing guide, and on a number of days that we estimate. They do not replace a quotation.
| Level | Scope | Assumption | Build (excl. VAT) |
|---|---|---|---|
| 1. Project inside an assistant | A few dozen documents loaded into a professional assistant | 1 to 2 internal days | Team time and seats |
| 2. Focused RAG | One document base, one team, a simple interface | 3 to 10 days | €2,400 to €15,000 |
| 3. Integrated RAG | Several sources, per-user permissions, automatic updates | 15 to 40 days | €12,000 to €60,000 |
For the API, let us assume 50 questions a day, 22 days a month, 8,000 tokens read and 500 written per question. With Anthropic's official prices noted on 4 November 2026, that gives $23.10 per month with Claude Sonnet 5.5 ($2 and $10 per million tokens) and $1.16 with Claude Haiku 5.5 ($0.10 and $0.50). Initial indexing, hosting and the search database come on top: their cost depends on the volume and the solution chosen, and should be requested in the quotation.
These figures show where the budget goes: into preparing the documents, access rights, testing and maintenance. If you first have to digitise 10 years of scanned PDFs, that line can exceed the AI part. Our article on the price of an AI agent details how recurring costs are calculated.
What are the limits of a RAG?
A RAG is only as good as your documents: it cannot answer correctly from a source that is wrong, out of date or missing. Here are the 6 limits to know before starting a project.
- Contradictory documents. 2 versions of a procedure give 2 answers. Every source needs a status and a date.
- The wrong passage found. The search can miss the right extract, especially on very technical terms or references.
- Tables and drawings misread. Complex PDFs need specific preparation, to be tested on your own documents.
- A plausible but wrong answer. The CNIL points out that a generative AI output can be inaccurate while looking credible.
- An internal leak. Without correct permissions, an employee can get information they were not allowed to read, such as a salary.
- A base that ages. Without upkeep, quality drops silently.
What does the CNIL say about a RAG?
The CNIL states that an organisation that connects an AI system to its own knowledge base is responsible for that processing when the base contains personal data. Its FAQ, published on 18 July 2024, cites this scenario. For personal data or sensitive documents, it generally favours an on-premises deployment, which limits the risk of extraction by a third party, and advises involving the data protection officer and, where relevant, carrying out an impact assessment.
In practice: remove from the index whatever is not necessary, apply the same access rights as in your folders, check the vendor's contract and location, and write these rules into your AI charter. A competent person must approve this framework. In a profession bound by professional secrecy, this framing comes before the choice of tool.
How do you know whether a RAG suits you?
A RAG suits you if you can write 20 real frequent questions, if their answers exist in your documents and if the lost time can be quantified. Run this test in 4 steps before any quotation:
- Have the team concerned write 20 real questions.
- For each one, note where the right answer is found. If you cannot find it yourself, the RAG will not find it either.
- Estimate the gain: minutes lost per question × number of questions × hourly cost. For example, 5 to 10 minutes × 50 questions a day × 220 days × €45 an hour gives roughly €41,000 to €82,000 a year, if the gain is real and the volumes are confirmed. Divide by 2 or 3 if you want a cautious assumption.
- Test with a small batch of documents in a professional assistant, before building.
After launch, track 3 figures: the share of answers with a correct source, the share of questions left unanswered, and the time actually saved.
It is poorly suited if your questions concern real-time figures, if your documents fit in a few files, or if nobody can maintain the base.
At Codito Ergo Sum, there are 2 paths. For a second brain and a first test with your team, the AI Partnership at €499 or €799 excl. VAT per month, no commitment, cancel at any time. For an integrated, secure RAG with an identified need, Custom Development, on quotation from €3,000 excl. VAT after scoping, the floor applying to a simple scope. To find out which suits you, let's take 30 minutes with your 20 questions.
Sources (checked on 4 November 2026)
- CNIL: questions and answers on generative AI and deploying a generative AI (both in French)
- Anthropic, API pricing: platform.claude.com/docs/en/about-claude/pricing

