Start With the Process: How RAG Moves EBIT
In short: Enterprise AI pays off when it is tied to one process with a measurable result. In this example, a machine builder's sales engineers spend hours finding answers to technical customer questions. Retrieval-augmented generation (RAG) finds the right datasheet passage in seconds and answers with the source. The effect shows up in revenue, cost and quality, which is to say in EBIT.
Process first, then technology
The starting point is not a model but a bottleneck: technical pre-sales questions that sit in a sales engineer's inbox while someone searches datasheets, old offers and test reports. The video uses an illustrative case of 2,000 requests a year at six hours each, which adds up to 12,000 hours of senior engineering time spent searching. Your own numbers will differ; the point is to calculate them before building anything.
Three levers come out of the same process:
- Revenue: faster answers win more deals.
- Cost: less time searching.
- Quality: fewer wrong answers, fewer claims.
How RAG works, in three steps
- Chunk. Every document is cut into small passages of a few paragraphs.
- Find by meaning. An embedding model turns each passage into a vector that represents its meaning, stored in a vector database. A question about "hot oil" finds the passage on "thermal fluid up to 150 °C", although the words differ.
- Answer with a source. The model receives the question and the best few passages, with one rule: answer only from this text and cite it.
When a datasheet changes, nothing is retrained. The passage is updated and the next answer is correct. The pump, values and datasheet in the video are a fictional example.
Questions people ask
Why not simply ask ChatGPT?
A public model has never seen your datasheets. They are private, and many were written after the model was trained. Without your documents it can only guess.
Why not put all documents into the prompt?
Thousands of datasheets do not fit into the model's context window. Even where they would, you would pay for all that text with every question, and answer quality drops when the relevant detail is buried.
Does RAG require retraining the model?
No. The model stays as it is. Only the document index is updated, which is why answers stay current and every answer can point to its source.
Transcript
Read the full transcript
Most companies start AI with the technology. We start with the process.
Take a typical machine builder. Every week, customers send technical questions. "Can your pump handle hot oil at 12 bar?" "Which seal do I need for this chemical?" Today, a sales engineer searches through datasheets, old offers and test reports. It takes hours, sometimes days.
Imagine 2,000 requests a year, six hours each. That's 12,000 hours of your best engineers spent searching, not selling. And while you search, a competitor sends the offer first.
So the business case is clear. Faster answers mean more deals won. Less searching means lower cost. Fewer mistakes mean fewer claims. Revenue, cost and quality, all in one process. That's EBIT.
So why not just ask ChatGPT? Because it has never seen your datasheets. They are private, and many were written after the model was trained.
The simple idea is to put all your documents into the prompt. But thousands of datasheets are far too much text. They don't fit into what the model can read at once, its context window. And even if they did, you would pay for all of it with every single question, and the model gets lost in the details. So we only give the model what matters.
Step one: we cut every document into small pieces, a few paragraphs each. Thousands of documents become many small cards.
Step two: we find the right cards. A keyword search for "hot oil" might fail, because the datasheet says "thermal fluid up to 150 degrees". Different words, same meaning. So we use an embedding model. It turns every card into a long list of numbers, a vector, that captures its meaning. Think of a map: cards about temperature limits sit close together, cards about delivery terms sit far away. All these vectors go into a vector database. When a customer question comes in, it becomes a vector too. We drop it on the same map and take the closest cards. The right datasheet section shows up, even with completely different words.
Step three: we give the model the question, the best few cards and one rule: answer only from this text, and name the source. The answer: "Yes, pump P-200 handles thermal oil up to 150 degrees at 16 bar. Datasheet P-200, section 3.1."
When product management updates a datasheet, there is no retraining. We update the cards, and the very next answer is already correct.
That's RAG: retrieve first, then generate. But the result is not a chatbot. It's a faster sales process. Answers in minutes instead of days. Engineers back with customers. Offers out before the competition. That's where AI meets EBIT, and where strategy meets execution.
Narrated by Christian Rose. Transcript lightly edited for readability.
Christian Rose (罗仕, Luó Shì) is the Founder and Managing Director of Q-Bridges GmbH, a Berlin-based strategy, AI and technology advisory firm working between Europe and China. He has 25 years of consulting experience and has worked as a Managing Director and Partner in Germany and China. Q-Bridges is vendor-neutral: it sells no licences and takes no commissions.
→ www.q-bridges.com · LinkedIn
We start with one process and its business case, then build the knowledge base, retrieval and answer layer around it, with sources, access rights and a measurable pilot.
See Corporate LLM & Company BrainAlso watch: A Chatbot Is Just the Tip: What Enterprise AI Really Needs